Findings of the Association for Computational Linguistics: EACL 2026
Unveiling the Deficiencies of Pre-trained Text-and-Layout Models in Real-world Visually-rich Document Information Extraction (2026.findings-eacl)
Copied to clipboard
Chong Zhang, Yixi Zhao, Yulu Xie, Chenshu Yuan, Yi Tu, Ya Guo, Mingxu Chai, Ziyu Shen, Yue Zhang, Qi Zhang
| Challenge: | PTLMs have shown remarkable success in multiple information extraction tasks . however, their performance in real-world scenarios falls short of expectations . |
| Approach: | They propose to use an entity-centric dataset to evaluate PTLMs' performance . they find that inadequate annotations in benchmark datasets lead to spurious correlations . |
| Outcome: | The proposed dataset disentangles the falsely-coupled segment and entity annotations that arises from the block-level annotation of FUNSD. |
Entity-aware Cross-lingual Claim Detection for Automated Fact-checking (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing work on verifiable claims detection is focused on monolingual solutions . identifying and validating claims related to global concerns requires a fact-checking pipeline capable of processing claims written in multiple languages. |
| Approach: | They propose an entity-aware cross-lingual claim detection model that generalizes well to handle multilingual claims. |
| Outcome: | The proposed model shows consistent performance gains across 27 languages and robust knowledge transfer between languages seen and unseen during training. |
WorkForceAgent-R1: Incentivizing Reasoning Capability in LLM-based Web Agents via Reinforcement Learning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing web agents relying on supervised fine-tuning struggle with generalization and robustness due to insufficient reasoning capabilities when handling the inherently dynamic nature of web interactions. |
| Approach: | They propose a large language model-empowered web agent that trains using a rule-based reinforcement learning framework to enhance single-step reasoning and planning for business-oriented web navigation tasks. |
| Outcome: | The proposed agent outperforms baseline LLM-based agents on the WorkArena benchmark by 10.26–16.59%. |
Being Kind Isn’t Always Being Safe: Diagnosing Affective Hallucination in LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly engaged in emotionally vulnerable conversations that extend beyond information seeking to moments of personal distress. |
| Approach: | They propose AHaBench, a benchmark of 500 mental-health-related prompts with expert-informed reference responses, evaluated along three dimensions: Emotional Enmeshment, Illusion of Presence, and Fostering Overdependence. |
| Outcome: | The proposed model is based on 500 mental-health-related prompts with expert-informed reference responses and a 5K-instance preference dataset enabling direct preference optimization (DPO) for alignment with emotionally responsible behavior. |
Joint Multimodal Preference Optimization for Fine-Grained Visual-Textual Alignment (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent research has focused on addressing multimodal hallucinations in Large Vision-Language Models (LVLMs) however, these methods lack fine-grained visual contrast mechanisms and rely on single-margin optimization. |
| Approach: | They propose a framework that integrates text-conditioned preference loss with visual ranking-based objective. |
| Outcome: | The proposed framework improves cross-modal alignment and fine-grained visual grounding. |
Let’s Put Ourselves in Sally’s Shoes: Shoes-of-Others Prefilling Improves Theory of Mind in Large Language Models (2026.findings-eacl)
Copied to clipboard
Kazutoshi Shinoda, Nobukatsu Hojo, Kyosuke Nishida, Yoshihiro Yamazaki, Keita Suzuki, Hiroaki Sugiyama, Kuniko Saito
| Challenge: | Existing methods for Theory of Mind (ToM) are specialized for inferring beliefs from contexts involving changes in the world state. |
| Approach: | They propose a method which makes fewer assumptions about contexts and is applicable to broader scenarios. |
| Outcome: | The proposed method makes fewer assumptions about contexts and is applicable to broader scenarios. |
Plane Geometry Problem Solving with Multi-modal Reasoning: A Survey (2026.findings-eacl)
Copied to clipboard
| Challenge: | Plane geometry problem solving has gained significant attention as a benchmark to assess the multi-modal reasoning capabilities of large vision-language models. |
| Approach: | They present a systematic review of existing work in PGPS and summarize their results. |
| Outcome: | The proposed frameworks are compared with existing frameworks and analyze them according to their architectural designs. |
Examining the Utility of Self-disclosure Types for Modeling Annotators of Social Norms (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent work has explored the use of personal information in the form of persona sentences to improve modeling of individual characteristics and prediction of annotator labels for subjective tasks. |
| Approach: | They categorize self-disclosures and use them to build annotator models for predicting judgments of social norms by analyzing comments from original post. |
| Outcome: | The proposed model improves the model and its ability to predict annotator labels. |
Position Paper: How Should We Responsibly Adopt LLMs in the Peer Review Process? (2026.findings-eacl)
Copied to clipboard
| Challenge: | a recent paper criticizes the current use of Large Language Models (LLMs) for simple review text generation. |
| Approach: | They propose to use Large Language Models to support key aspects of the review process . they argue that this approach overlooks more meaningful applications of LLMs . authors argue that the increased reviewing burden per reviewer is a factor . |
| Outcome: | The proposed approach would support reproducibility, correctness and relevance of citations and ethics review flagging. |
Rad-Flamingo: A Multimodal Prompt driven Radiology Report Generation Framework with Patient-Centric Explanations (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing reports are labor-intensive and expert-intensive, resulting in inconsistencies and a lack of patient-centered insight. |
| Approach: | They propose a multimodal prompt-driven report generation framework that integrates diverse data modalities to produce comprehensive and context-aware radiology reports. |
| Outcome: | The proposed framework improves report quality, improves understandability and could foster better patient-doctor communication. |
I-MCTS: Enhancing Agentic AutoML via Introspective Monte Carlo Tree Search (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing LLM-based agents struggle with low diversity and suboptimal code generation. |
| Approach: | They propose an approach that iteratively expands tree nodes through an introspective process that meticulously analyzes solutions and results from parent and sibling nodes. |
| Outcome: | The proposed approach shows a 4% improvement in performance compared to the strong open-source AutoML agents. |
ThinkNote: Enhancing Knowledge Integration and Utilization of Large Language Models via Constructivist Cognition Modeling (2026.findings-eacl)
Copied to clipboard
Zhipeng Xu, Zhenghao Liu, Yukun Yan, Shuo Wang, Shi Yu, Zheni Zeng, Chaojun Xiao, Zhiyuan Liu, Ge Yu, Chenyan Xiong
| Challenge: | Large Language Models (LLMs) exhibit suboptimal behaviors and inconsistencies when exposed to unfamiliar external information, underscoring their limitations in effectively leveraging such knowledge. |
| Approach: | They propose a framework that enhances the external knowledge utilization of Large Language Models through a two-stage constructivist cognitive modeling process. |
| Outcome: | The proposed framework achieves a 10% improvement over baseline methods on various question-answering benchmarks. |
Mitigating Copy Bias in In-Context Learning through Neuron Pruning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models have shown impressive few-shot in-context learning abilities, but they are prone to a ‘copying bias’, where they copy answers from provided examples instead of learning the underlying patterns. |
| Approach: | They propose a method to prune neurons that prioritize copying over generalization and adopt a task-recognition perspective on ICL and examine task vectors induced by the model. |
| Outcome: | The proposed method improves performance across a diverse set of ICL tasks while maintaining or improving the model’s general capabilities. |
How to Make LMs Strong Node Classifiers? (2026.findings-eacl)
Copied to clipboard
Zhe Xu, Kaveh Hassani, Si Zhang, Hanqing Zeng, Michihiro Yasunaga, Limei Wang, Dongqi Fu, Ning Yao, Bo Long, Hanghang Tong
| Challenge: | Language Models (LMs) are increasingly challenging the dominance of domain-specific models, such as Graph Neural Networks (GNNs) and Graph Transformers (GTs). |
| Approach: | They propose a novel approach that empowers off-the-shelf LMs to achieve performance comparable to state-of-the art (SOTA) GNNs on node classification tasks without requiring any architectural modifications. |
| Outcome: | The proposed approach outperforms existing GNNs on node classification tasks and is open-source upon publication. |
Rethinking Data Mixture for Large Language Models: A Comprehensive Survey and New Perspectives (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for data mixture improve the generalization capability of large language models (LLMs) on downstream tasks. |
| Approach: | They propose a fine-grained categorization of existing methods and propose three subtypes of offline and online methods. |
| Outcome: | The proposed methods extend beyond offline and online classifications and highlight key challenges in the field of data mixture. |
The Mediomatix Corpus: Parallel Data for Romansh Language Varieties via Comparable Schoolbooks (2026.findings-eacl)
Copied to clipboard
| Challenge: | a small-scale human evaluation confirms that the segments are highly parallel, making the dataset suitable for NLP applications. |
| Approach: | They present a first parallel corpus of Romansh idioms from 291 schoolbooks . they use automatic alignment methods to extract 207k multi-parallel segments from the books . |
| Outcome: | The proposed corpus is based on 291 schoolbook volumes, which are comparable in content for the five idioms. |
Unleashing the Unseen: Harnessing Benign Datasets for Jailbreaking Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Despite significant efforts in safety alignment, large language models (LLMs) such as GPT-4 and LLaMA 3 remain vulnerable to jailbreak attacks that can induce harmful behaviors. |
| Approach: | They propose a feature extraction method to extract sample-agnostic features from benign datasets in the form of adversarial suffixes and propose 'suffix maybe features' they show that adversarials generated from jailbreak attacks may contain meaningful features, i.e. appending the same suffix to different prompts results in responses exhibiting specific characteristics. |
| Outcome: | The proposed method extracts sample-agnostic features from benign datasets and shows that they may contain meaningful features. |
JEEM: Vision-Language Understanding in Four Arabic Dialects (2026.findings-eacl)
Copied to clipboard
Karima Kadaoui, Hanin Atwany, Hamdan Al-Ali, Abdelrahman Mohamed, Ali Mekky, Sergei Tilga, Natalia Fedorova, Ekaterina Artemova, Hanan Aldarmaki, Yova Kementchedjhieva
| Challenge: | Existing evaluation datasets feature Western-centric images and English text, while their non-English counterparts are often derived from the latter. |
| Approach: | They propose to evaluate Vision-Language Models (VLMs) on visual understanding across four Arabic-speaking countries: Jordan, The Emirates, Egypt, and Morocco. |
| Outcome: | The proposed model underperforms in visual understanding and dialect-specific generation across four Arabic-speaking countries. |
Detecting Primary Progressive Aphasia (PPA) from Text: A Benchmarking Study (2026.findings-eacl)
Copied to clipboard
Ghofrane Merhbene, Fabian Lecron, Philippe Fortemps, Bradford C. Dickerson, Mascha Kurpicz-Briki, Neguine Rezaii
| Challenge: | Primary progressive aphasia (PPA) is a neurodegenerative disorder characterized by progressive language deficits as the primary symptom. |
| Approach: | They benchmarked the performance of traditional machine learning models with various feature extraction techniques, transformer-based models, and large language models (LLMs) they found that transformer-Based models exceeded chance-level performance in terms of balanced accuracy, while MLP using MentalBert’s embeddings achieved the highest accuracy. |
| Outcome: | The proposed models outperform chance-level models in terms of balanced accuracy while using MentalBert’s embeddings achieve the highest accuracy. |
Geometric Interpretation of Layer Normalization and a Comparative Analysis with RMSNorm (2026.findings-eacl)
Copied to clipboard
| Challenge: | a novel geometric interpretation of LayerNorm is presented . layer normalization is a crucial yet often overlooked component of the transformer architecture . |
| Approach: | They propose a geometric interpretation of LayerNorm and explore how LayerNorm influences the norm and orientation of hidden vectors in the representation space. |
| Outcome: | The proposed interpretation of LayerNorm shows that it is redundant to remove a component along the uniform vector during training and inference. |
Continual Pretraining on Encrypted Synthetic Data for Privacy-Preserving LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to protect PII from training on small corpora are difficult to implement in real-world applications. |
| Approach: | They propose an entity-based framework that synthesizes encrypted training data to protect PII. |
| Outcome: | The proposed framework outperforms base models and ensures PII security on limited-scale datasets while exhibiting a modest performance gap compared to models trained on unencrypted synthetic data. |
Do Diacritics Matter? Evaluating the Impact of Arabic Diacritics on Tokenization and LLM Benchmarks (2026.findings-eacl)
Copied to clipboard
| Challenge: | Diacritics can significantly influence language processing tasks in Arabic . their presence can increase subword fragmentation during tokenization, reducing performance . |
| Approach: | They analyze the impact of diacritics on tokenization and benchmark task performance across major Large Language Models. |
| Outcome: | The proposed model is robust to diacritics, but full diacritization leads to token fragmentation and degraded performance. |
Pelican Soup Framework: A Theoretical Framework for Language Model Capabilities (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing theoretical frameworks for large language models (LLMs) do not explain how pretraining leads to in-context learning. |
| Approach: | They propose a theoretical framework that allows LLMs to generalize to unseen instructions and perform in-context learning even when verbalizers are irrelevant to the task. |
| Outcome: | The proposed framework can be used to analyze LLMs' ability to perform in-context learning . it can be applied to linguistic, psychology, and philosophy tasks . |
I Know, but I Don’t Know! How Persona Conflict Undermines Instruction Adherence in Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies on persona-grounded dialogue assume idealized scenarios where persona and user utterances are fully aligned. |
| Approach: | They propose a taxonomy that categorizes model behaviors into three response types . they propose sycophantic, adherent, and wavering responses as response types. |
| Outcome: | The proposed framework categorizes model behaviors into three response types and develops a measurement schema grounded in this taxonomy. |
Persona-driven Simulation of Voting Behavior in the European Parliament with Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models exhibit a progressive left-leaning bias, but can also produce behavior that aligns with socioeconomic groups. |
| Approach: | They analyze whether persona prompting can accurately predict individual voting decisions . they find that they can simulate the voting behavior of European Parliament members reasonably well . |
| Outcome: | The proposed model can predict the voting behavior of European Parliament members reasonably well, with a weighted F1 score of approximately 0.793. |
Exploring Iterative Controllable Summarization with Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) excel at abstractive summarization tasks, but their ability to precisely control summary attributes remains underexplored. |
| Approach: | They propose a guide-to-explain framework for controllable summarization that enables the model to identify misaligned attributes in the initial draft and guides it to self-explan errors in the previous output. |
| Outcome: | The proposed framework generates well-adjusted summaries that satisfy the desired attributes with robust effectiveness while requiring surprisingly fewer iterations than other iterative approaches. |
The Price of Thought: A Multilingual Analysis of Reasoning, Performance, and Cost of Negotiation in Large Language Models (2026.findings-eacl)
Copied to clipboard
Sherzod Hakimov, Roland Bernard, Tim Leiber, Karl Osswald, Kristina Richert, Ruilin Yang, Raffaella Bernardi, David Schlangen
| Challenge: | Negotiation is a fundamental challenge for AI agents as it requires an ability to reason strategically, model opponents, and balance cooperation with competition. |
| Approach: | They propose to use a self-play setup to compare commercial and open-weight large language models to their vanilla counterparts in three different languages to examine trade-offs between performance and cost. |
| Outcome: | The proposed model improves GPT-5's performance by 31.4 % while increasing its cost by nearly 400 %. |
ART: Adaptive Reasoning Trees for Explainable Claim Verification (2026.findings-eacl)
Copied to clipboard
Sahil Wadhwa, Himanshu Kumar, Guanqun Yang, Abbaas Alif Mohamed Nishar, Pranab Mohanty, Swapnil Shinde, Yue Wu
| Challenge: | Large Language Models (LLMs) are powerful candidates for complex decision-making, leveraging vast encoded knowledge and remarkable zero-shot abilities. |
| Approach: | They propose a hierarchical method for claim verification that uses a root claim and a pairwise tournament of its children to determine an argument's strength. |
| Outcome: | The proposed method outperforms baseline methods on multiple datasets and shows that it is more reliable and clearer than existing methods. |
VortexPIA: Indirect Prompt Injection Attack against LLMs for Efficient Extraction of User Privacy (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have been widely deployed in Conversational AIs . however, the methods proposed in the study rely on a white-box setting . |
| Approach: | They propose an indirect prompt injection attack that induces privacy extraction in LLMs . they use token-efficient data containing false memories to inject LLM data . |
| Outcome: | The proposed method outperforms baselines and achieves state-of-the-art performance. |
VisDoT : Enhancing Visual Reasoning through Human-Like Interpretation Grounding and Decomposition of Thought (2026.findings-eacl)
Copied to clipboard
| Challenge: | Lack of perceptual grounding limits vision-language models' ability to interpret visual data . prior work on visualized data understanding focused on adapting VLMs to instruction tuning and chain-of-thought supervision . |
| Approach: | They propose a framework that enhances visual reasoning through human-like interpretation grounding. |
| Outcome: | The proposed framework improves on ChartQA and ChartQAPro benchmarks by +11.2%. |
KNN-SSD: Enabling Dynamic Self-Speculative Decoding via Nearest Neighbor Layer Set Optimization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have proven highly capable in handling downstream tasks, but the token-by-token generation in autoregressive decoding results in quadratic computational complexity. |
| Approach: | They propose a method that proposes skipping certain layers to construct a draft model, which eliminates the need for additional parameters or training. |
| Outcome: | The proposed method achieves 1.31.6 speedup in LLM inference while being sensitive to domain shifts. |
Towards Robust Evaluation of Visual Activity Recognition: Resolving Verb Ambiguity with Sense Clustering (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing evaluation methods for visual activity recognition systems fail to capture ambiguities in verb semantics and image interpretation. |
| Approach: | They propose a framework that constructs verb sense clusters to evaluate visual activity recognition systems. |
| Outcome: | The proposed framework provides a more robust evaluation of visual activity recognition systems. |
HiKE: Hierarchical Evaluation Framework for Korean-English Code-Switching Speech Recognition (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent advances in automatic speech recognition (ASR) have pushed error rates below 5% on standard monolingual benchmarks. |
| Approach: | They propose a framework for the evaluation of multilingual ASR models using loanword labels and a hierarchical CS-level labeling scheme that allows for fine-tuning with synthetic CS data. |
| Outcome: | The proposed framework provides a means for the precise evaluation of multilingual ASR models and fosters research in the field. |
Complexity-aware fine-tuning (2026.findings-eacl)
Copied to clipboard
| Challenge: | General-purpose Large Language Models (LLMs) are often fine-tuned through supervised fine- tuning (SFT) to enhance performance in specific domains. |
| Approach: | They propose a novel approach that uses reasoning only for complex data identified by entropy to refine large language models. |
| Outcome: | The proposed model outperforms the standard SFT approach while using 81% less data. |
Multilingual Retrieval-Augmented Generation for Knowledge-Intensive Question Answering Task (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies focus on English as the data language for RAG, resulting in limited coverage of multilingual RAG. |
| Approach: | They propose a method that translates retrieved documents into a common language before generating the response. |
| Outcome: | The proposed approach improves efficiency on knowledge-intensive tasks but introduces inconsistencies due to cross-lingual variations in the retrieved content. |
SpatialMath: Spatial Comprehension-Infused Symbolic Reasoning for Mathematical Problem-Solving (2026.findings-eacl)
Copied to clipboard
| Challenge: | Current models struggle to accurately decompose intricate visual inputs and connect perception with structured reasoning, leading to suboptimal performance. |
| Approach: | They propose a Spatial Comprehension-Infused Symbolic Reasoning Framework to integrate spatial representations into structured symbolic reasoning chains. |
| Outcome: | The proposed framework outperforms existing models in vision-intensive mathematical problems. |
Skill Discovery for Software Scripting Automation via Offline Simulations with LLMs (2026.findings-eacl)
Copied to clipboard
Paiheng Xu, Gang Wu, Xiang Chen, Tong Yu, Chang Xiao, Franck Dernoncourt, Tianyi Zhou, Wei Ai, Viswanathan Swaminathan
| Challenge: | Large Language Models (LLMs) can generate code from natural language queries, but runtime code generation is limited due to unverified code, security risks, longer response times, and higher computational costs. |
| Approach: | They propose an offline simulation framework to curate a software-specific skillset by exploiting large language models and publicly available scripting guides. |
| Outcome: | The proposed framework significantly improves automation success rates, reduces response time, and saves runtime token costs compared to traditional runtime code generation. |
How Important is ‘Perfect’ English for Machine Translation Prompts? (2026.findings-eacl)
Copied to clipboard
Patrícia Schmidtová, Niyati Bafna, Seth Aycock, Gianluca Vico, Wiktor Kamzela, Kathy Hämmerl, Vilém Zouhar
| Challenge: | Large language models (LLMs) are largely trained on and respond best to English prompts, but are also sensitive to errors in user prompts. |
| Approach: | They propose to model a range of error types exhibited by second language English speakers and quantify their impact on LLM performance. |
| Outcome: | The proposed model is brittle to natural spelling errors but not to errors at the phrasal level, but the variance in quality caused by these errors is lower than the variance over the initial prompt choice. |
KETCHUP: K-Step Return Estimation for Sequential Knowledge Distillation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Empirical evaluation shows that our approach yields superior performance in both standard task metrics and large language model (LLM)-based evaluation. |
| Approach: | They propose a K-step return estimation method for reinforcement learning (RL)-based knowledge distillation in text generation tasks using the Bellman Optimality Equation. |
| Outcome: | The proposed method performs better on standard task metrics and large language model evaluations on three text generation tasks. |
Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for steering concept vectors suffer from noisy features in diverse datasets that undermine steering robustness. |
| Approach: | They propose a Sparse Autoencoder-Denoised Concept Vector (SDCV) which selectively keeps the most discriminative SAE latents while reconstructing hidden representations. |
| Outcome: | The proposed method improves steering success rates by 4-16% across six challenging concepts while maintaining topic relevance. |
Shifting Perspectives: Steering Vectors for Robust Bias Mitigation in LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Despite efforts to mitigate social bias in large language models, representational harms such as stereotyping continue to exist in both open and closed-source models. |
| Approach: | They propose a method to modify model activations in forward passes by applying steering vectors to a BBQ dataset and comparing their results to bias mitigation methods. |
| Outcome: | The proposed method outperforms 3 other bias mitigation methods on the BBQ dataset and shows the lowest impact on MMLU scores. |
MAPS: A Multilingual Benchmark for Agent Performance and Security (2026.findings-eacl)
Copied to clipboard
Omer Hofman, Jonathan Brokman, Oren Rachmil, Shamik Bose, Vikas Pahuja, Toshiya Shimizu, Trisha Starostina, Kelly Marchisio, Seraphina Goldfarb-Tarrant, Roman Vainshtein
| Challenge: | Existing benchmarks do not provide a comprehensive, multi-domain, security-aware evaluation of multilingual agentic AI systems. |
| Approach: | They propose a multilingual benchmark suite to evaluate agentic AI systems across languages and tasks. |
| Outcome: | The proposed framework evaluates agentic AI systems across languages and tasks. |
Linking Knowledge to Care: Knowledge Graph-Augmented Medical Follow-Up Question Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing large language models (LLMs) fail to identify information gaps across diverse symptoms. |
| Approach: | They propose a Knowledge Graph-augmented LLM with active in-context learning to generate relevant and important follow-up questions. |
| Outcome: | The proposed framework outperforms state-of-the-art methods by 5% - 8% on relevant benchmarks. |
DebateQA: Evaluating Question Answering on Debatable Knowledge (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing QA benchmarks that provide fixed answers to debatable questions are inadequate for evaluating their performance. |
| Approach: | They propose to use a dataset of 2,941 debatable questions to assess their ability to provide comprehensive answers to inherently debatably asked questions. |
| Outcome: | The proposed model performs well on 2,941 debatable questions accompanied by human-annotated partial answers that capture a variety of perspectives. |
Personal Information Parroting in Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Modern language models memorize millions of PI instances, increasing privacy risks. |
| Approach: | They develop a model that parrots 13.6% of PI verbatim on a manually curated set of 483 instances . they recommend that pretraining datasets be aggressively filtered and anonymized to minimize PI parroting. |
| Outcome: | The proposed model outperforms the best regex-based PI detectors on a manually curated set of 483 instances of PI. |
Harmful Factuality: LLMs Correcting What They Shouldn’t (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are trained for factual accuracy, but can conflict with the critical demand for source fidelity. |
| Approach: | They propose a reproducible framework to elicit and measure HFH using controlled entity-level perturbations and strategic entity selection. |
| Outcome: | The proposed framework reduces HFH rates by 50% across summarization, rephrasing, and QA tasks. |
Toward Beginner-Friendly LLMs for Language Learning: Controlling Difficulty in Conversation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Practicing conversations with large language models is a promising alternative to traditional in-person language learning. |
| Approach: | They propose a new token-level evaluation metric, Token Miss Rate, that measures the proportion of incomprehensible tokens per utterance and correlates strongly with human judgments. |
| Outcome: | The proposed methods improve comprehensibility for beginner speakers from 39.4% to 83.3%, compared with prompting alone and a token-level evaluation metric, Token Miss Rate (TMR). |
CodeGuard: Improving LLM Guardrails in CS Education (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly embedded in Computer Science classrooms to automate code generation, feedback, and assessment. |
| Approach: | They propose a guardrail framework for educational AI systems that can handle unsafe and irrelevant prompts. |
| Outcome: | The proposed framework reduces potentially harmful or policy-violating code completions by 30-65% without degrading performance on legitimate educational tasks. |
ATOM: AdapTive and OptiMized dynamic temporal knowledge graph construction using LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Unstructured data is expanding at an unprecedented rate, and static knowledge graphs are often overlooked due to their dynamic nature and lack of time-sensitive features. |
| Approach: | They propose a few-shot approach that builds and continuously updates Temporal Knowledge Graphs (TKGs) from unstructured texts. |
| Outcome: | Empirical results show that ATOM achieves 18% higher exhaustivity, 33% better stability, and over 90% latency reduction compared to baseline methods. |
On the Interplay between Human Label Variation and Model Fairness (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies on the impact of human label variation on model fairness have not explored the interaction between HLV and performance. |
| Approach: | They compare human label variation (HLV) training methods with four other methods . they find that HLV methods improve performance without harming fairness . |
| Outcome: | The proposed methods improve fairness without explicit debiasing under certain configurations. |
Where do LLMs currently stand on biomedical NER in both clean and noisy settings ? (2026.findings-eacl)
Copied to clipboard
| Challenge: | despite advances in medicine, many diseases remain without effective treatments . clinical meta-analysis is essential for drug discovery and clinical research . |
| Approach: | They investigate the performance of large language models (LLMs) on biomedical NER tasks . findings suggest LLMs exhibit a notable degree of robustness to noise . |
| Outcome: | The proposed models are closing the performance gap with BERT-based models and demonstrate particular strengths in low-data settings. |
Scaling Data-Constrained Language Models with Synthetic Data (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) improve with more training data, but practical limitations on data collection constrain further scaling. |
| Approach: | They compare three strategies to generate Japanese text, repeat the limited Japanese Web text, and use English Web text to fill the data shortfall. |
| Outcome: | The proposed model outperforms baselines and achieves the performance achieved when the entire token budget is filled with additional organic Japanese Web text. |
The Unintended Trade-off of AI Alignment: Balancing Hallucination Mitigation and Safety in LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Hallucination in large language models has been studied, but a side effect remains unrecognized . a new study examines the trade-off between truthfulness and safety alignment . |
| Approach: | They propose a method that disentangles hallucination from hallucinian features using sparse autoencoders. |
| Outcome: | The proposed method preserves refusal behavior and task utility while maintaining safety alignment. |
The Model’s Language Matters: A Comparative Privacy Analysis of LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models are increasingly deployed in multilingual settings that process sensitive data . prior privacy evaluations focused on English, but new research shows that language matters for privacy leakage . |
| Approach: | They quantify six corpus-level linguistic indicators and evaluate vulnerability under three attack families. |
| Outcome: | The results show that language matters for privacy leakage in large language models . Italian exhibits the strongest exposure, while English and French are more resilient . |
Towards the First NLP Benchmark for Ladin - an Extremely Low-Resource Language (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are limited in low-resource languages due to lack of labeled training data. |
| Approach: | They propose to use Ladin as a model for sentiment analysis and question answering by incorporating Italian data into machine translation training. |
| Outcome: | The proposed method improves on existing Italian–Ladin translation baselines. |
DRAGON: Domain-specific Robust Automatic Data Generation for RAG Optimization (2026.findings-eacl)
Copied to clipboard
Haiyang Shen, Hang Yan, Zhongshi Xing, Mugeng Liu, Yue Li, Zhiyang Chen, Yuxiang Wang, Jiuzheng Wang, Yun Ma
| Challenge: | Existing retrieval-augmented generation paradigms rely heavily on public knowledge . Existing RAGs reliant on public information and often falter when faced with domain-specific queries. |
| Approach: | They propose a framework that combines a data-construction modeling approach with a scalable synthetic data-generation pipeline to optimize domain-specific retrieval performance. |
| Outcome: | The proposed framework optimizes domain-specific retrieval performance and bolsters retriever robustness. |
Causal Activation Steering via Sparse Mediation (2026.findings-eacl)
Copied to clipboard
| Challenge: | a sparse mediation steering approach to control language-model behavior is feasible, says a new study . existing methods that learn dense steering vectors modify thousands of activation dimensions simultaneously . |
| Approach: | They propose a sparse mediation steering approach that learns targeted behavioral interventions via regularized training. |
| Outcome: | The proposed method achieves 97-100% of dense baseline effectiveness across four tasks while using only 10-30% of activation dimensions. |
Causal Direct Preference Optimization for Language Model Alignment (2026.findings-eacl)
Copied to clipboard
| Challenge: | Empirical evaluations show that CDPO surpasses DPO-based baselines by achieving unbiased fine-tuning through causal reasoning. |
| Approach: | They propose a framework that incorporates causal inference principles to mitigate the influence of confounders and sharpen the signal of genuine human preferences. |
| Outcome: | The proposed framework preserves the tractability of direct optimization while enhancing robustness to spurious correlations and annotation biases. |
LLMs Faithfully and Iteratively Compute Answers During CoT: A Systematic Analysis With Multi-step Arithmetics (2026.findings-eacl)
Copied to clipboard
Keito Kudo, Yoichi Aoki, Tatsuki Kuribayashi, Shusaku Sone, Masaya Taniguchi, Ana Brassard, Keisuke Sakaguchi, Kentaro Inui
| Challenge: | Specifically, we examine when the LLMs’ answer is (pre)determined, especially before the CoT begins or after, and how strongly the information from CoT specifically has a causal effect on the final answer. |
| Approach: | They examine when the LLMs’ answer is (pre)determined, especially before the CoT begins or after, and how strongly the information from CoT specifically has a causal effect on the final answer. |
| Outcome: | The proposed model can generate reasoning chains while generating the reasoning chain on the fly. |
VaseVQA: Multimodal Agent and Benchmark for Ancient Greek Pottery (2026.findings-eacl)
Copied to clipboard
Jinchao Ge, Tengfei Cheng, Biao Wu, Zeyu Zhang, Shiya Huang, Judith Bishop, Gillian Shepherd, Meng Fang, Ling Chen, Yang Zhao
| Challenge: | MLLMs that use domain-specific data are limited in understanding cultural heritage artifacts such as ancient Greek pottery . supervised fine-tuning improves adaptation to domain knowledge, but it struggles with deeper reasoning tasks. |
| Approach: | They propose a visual question-answer tool that augments SFT with reinforcement learning using verifiable rewards. |
| Outcome: | The proposed model outperforms baseline models on reasoning-intensive questions on ancient Greek pottery. |
PromptPrism: A Linguistically-Inspired Taxonomy for Prompts (2026.findings-eacl)
Copied to clipboard
| Challenge: | PromptPrism is a linguistically-inspired taxonomy that enables prompt analysis across three hierarchical levels. |
| Approach: | They propose a linguistically-inspired taxonomy that enables prompt analysis across three hierarchical levels: functional structure, semantic component, and syntactic pattern. |
| Outcome: | The proposed taxonomy bridges traditional language understanding with modern LLM research . it improves prompt quality and improves model performance across tasks . |
HiGraAgent: Dual-Agent Adaptive Reasoning over Hierarchical Knowledge Graph for Open Domain Multi-hop Question Answering (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to multi-hop question answering lack a robust and flexible approach to QA . prior work showed compositionality gap persists even for Large Language Models . |
| Approach: | They propose a framework that unifies graph-based retrieval with adaptive reasoning . HiGraAgent uses a hierarchical knowledge Graph with entity alignment . |
| Outcome: | The proposed framework outperforms the strongest graph-based method on hotpotQA, 2WikiMultihopQA, and MuSiQue. |
Hire Your Anthropologist! Rethinking Culture Benchmarks Through an Anthropological Lens (2026.findings-eacl)
Copied to clipboard
| Challenge: | anthropological accounts of culture often focus on static facts or homogeneous values . large language models are being implemented in translation systems, educational tools and search engines . |
| Approach: | They propose to categorize how benchmarks frame culture such as knowledge, preference, performance, or bias. |
| Outcome: | The proposed framework categorizes how benchmarks frame culture, such as knowledge, preference, performance, or bias. |
Suppressing Final Layer Hidden State Jumps in Transformer Pretraining (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing models exhibit only slight changes in the angular distance between the input and output hidden state vectors in the middle layers . |
| Approach: | They propose a jump-suppressing regularizer which penalizes large hidden state displacements near the final layer during pre-training. |
| Outcome: | The proposed method significantly reduces hidden state jumps in the final layer and increases model capacity. |
Intention-Adaptive LLM Fine-Tuning for Text Revision Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing work on large language models (LLMs) has demonstrated impressive capabilities in context-based text generation tasks, such as summarization and reasoning. |
| Approach: | They propose an intention-adaptive layer-wise LLM fine-tuning framework that dynamically selects a subset of LLM layers to learn intentions and transfers them to revision generation. |
| Outcome: | The proposed framework outperforms PEFT baselines on small revision corpora while maintaining fast convergence and accuracy. |
ReAttn: Improving Attention-based Re-ranking via Attention Re-weighting (2026.findings-eacl)
Copied to clipboard
| Challenge: | Attention-based re-ranking methods are highly concentrated a small subset of tokens within a few documents, making others indistinguishable. |
| Approach: | They propose a post-hoc re-weighting strategy that uses attention weights to reduce lexical bias and emphasize distinctive terms. |
| Outcome: | The proposed method reduces lexical bias and emphasizes distinctive terms across documents, while maintaining a balanced distribution across informative tokens. |
MapAgent: A Hierarchical Agent for Geospatial Reasoning with Dynamic Map Tool Integration (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing frameworks for large language models are tailored to domains such as mathematics, coding, or web automation. |
| Approach: | They propose a hierarchical multi-agent plug-and-play framework with customized toolsets and agentic scaffolds for map-integrated geospatial reasoning. |
| Outcome: | The proposed framework decouples planning from execution and reduces cognitive load on users. |
Comprehensive Study of Bilingual and Multi-category Instruction Pre-training (2026.findings-eacl)
Copied to clipboard
| Challenge: | Instruction pre-training (IPT) has recently gained attention as an intermediate stage between pre- and post-training for large language models. |
| Approach: | They study the optimal balance between raw and instruction-response data, languages, and task categories in an LLM instruction-respondence dataset. |
| Outcome: | The proposed model improves on English-centric and bilingual models using bilingual instruction-response datasets. |
Reflect, Rewrite, Repeat: How Simple Arithmetic Enables Advanced Reasoning in Small Language Models (2026.findings-eacl)
Copied to clipboard
Mengdie Flora Wang, Haochen Xie, Mun Young Kim, Baishali Chaudhury, Meghana Ashok, Suren Gunturu, Sungmin Hong, Jae Oh Woo
| Challenge: | Recent advances in language model reasoning require computationally intensive reinforcement learning and massive datasets. |
| Approach: | They propose a framework that combines Direct Preference Optimization and Supervised Fine-Tuning with selective guidance from larger models and iteratively refining solutions through a "reflect, rewrite, repeat" cycle. |
| Outcome: | The proposed framework shows significant performance improvements across arithmetic, symbolic and cognitive reasoning benchmarks. |
Don’t Judge Code by Its Cover: Exploring Biases in LLM Judges for Code Evaluation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly used as evaluators for code evaluation tasks . however, whether they can handle superficial variations remains unclear . |
| Approach: | They define six types of potential biases in code evaluation and reveal their impact on LLM judges. |
| Outcome: | The proposed method can be used to evaluate semantically equivalent code with superficial variations without reference implementations. |
COMMUNITYNOTES: A Dataset for Exploring the Helpfulness of Fact-Checking Explanations (2026.findings-eacl)
Copied to clipboard
| Challenge: | X, Meta, and TikTok are experimenting with community-based factchecking . community-driven verification is a way to provide explanatory notes that clarify why a post might be misleading . |
| Approach: | They propose a framework that optimizes the helpfulness of explanatory notes and the reason for this by automatically optimizing the prompt definitions. |
| Outcome: | The proposed framework improves helpfulness and reason prediction on 104k posts with user-provided notes and helpfulness labels. |
Persona Jailbreaking in Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies focus on narrative or role-playing tasks and overlook how adversarial conversational history alone can reshape induced personas. |
| Approach: | They propose a framework that embeds semantically loaded cues into user queries to gradually induce reverse personas. |
| Outcome: | The proposed framework predictably shifts personas, triggers collateral changes in correlated traits, and exhibits stronger effects in multi-turn settings. |
ParsTranslit: Truly Versatile Tajik-Farsi Transliteration (2026.findings-eacl)
Copied to clipboard
| Challenge: | Despite significant similarities between the two written standards, script differences hinder simple one-to-one mapping, hindering written communication and interaction between Tajikistan and its Persian-speaking “siblings”. |
| Approach: | They propose to use a sequence-to-sequence model to convert between two scripts in a Persian-speaking country using two datasets. |
| Outcome: | The proposed model achieves chrF++ and Normalized CER scores of 87.91 and 0.05 from Farsi to Tajik and 92.28 and 0.04 from Tajikistan to Farsis. |
One Sentence, Two Embeddings: Contrastive Learning of Explicit and Implicit Semantic Representations (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing sentence embedding methods lack the ability to capture the implicit semantics of sentences. |
| Approach: | They propose a sentence embedding method that assigns two embeddables to each sentence . one represents the explicit semantics and the other represents the implicit semantics . results show DualCSE can effectively encode both explicit and implicit meanings - they argue . |
| Outcome: | The proposed method can effectively encode both explicit and implicit meanings and improve the performance of the downstream task. |
ETOM: A Five-Level Benchmark for Evaluating Tool Orchestration within the MCP Ecosystem (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing benchmarks assess tools in isolation, overlooking challenges such as functional overlap and cross-server orchestration, which can lead to overly optimistic evaluations. |
| Approach: | They propose a five-level benchmark for evaluating multi-hop, end-to-end tool orchestration by LLM agents within a hierarchical Model-Context Protocol (MCP) ecosystem. |
| Outcome: | The proposed framework evaluates end-to-end tool orchestration by agents in hierarchical Model-Context Protocol (MCP) environments. |
SymCode: A Neurosymbolic Approach to Mathematical Reasoning via Verifiable Code Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Current approaches to mathematical reasoning are inference-time prompting and model fine-tuning. |
| Approach: | They propose a neurosymbolic framework that reframes mathematical problem-solving as a task of verifiable code generation using the SymPy library. |
| Outcome: | The proposed framework improves accuracy on MATH-500 and OlympiadBench benchmarks. |
Unsupervised Detection of LLM-Generated Text in Korean Using Syntactic and Semantic Cues (2026.findings-eacl)
Copied to clipboard
| Challenge: | Prior work focused on English, leaving low-resource languages such as Korean underexplored. |
| Approach: | They propose an unsupervised framework that integrates syntactic token cohesiveness and semantic regeneration similarity to detect Korean text. |
| Outcome: | The proposed framework outperforms baselines in Korean and other low-resource languages without training. |
NLP Privacy Risk Identification in Social Media (NLP-PRISM): A Survey (2026.findings-eacl)
Copied to clipboard
| Challenge: | Social media platforms such as X (formerly Twitter), Facebook, and Reddit generate user-generated content. |
| Approach: | They propose a framework to assess privacy risks in social media by evaluating vulnerabilities across six dimensions: data collection, preprocessing, visibility, fairness, computational risk, and regulatory compliance. |
| Outcome: | The proposed framework assesses privacy risks across six dimensions . it achieves F1-scores of 0.58–0.84, but incurs 1% - 23% drop under fine-tuning . |
CrowdSelect: SyntheticInstruction Data Selection with Multi-LLM Wisdom (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for capturing instruction-following complexity rely on single-dimensional signals, but they fail to capture complexity across diverse fields. |
| Approach: | They propose three foundational metrics that leverage Multi-LLMs wisdom to capture instruction-response pair characteristics and propose CrowdSelect, an integrated metric incorporating a clustering-based approach to maintain response diversity. |
| Outcome: | The proposed metrics outperform existing models on MT-bench and Arena-hard and show improvements of 4.81% on full and LoRA fine-tuning. |
Bias in the Ear of the Listener: Assessing Sensitivity in Audio Language Models Across Linguistic, Demographic, and Positional Variations (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent advances extend language understanding beyond text to speech, enabling unified reasoning across modalities. |
| Approach: | They construct and release a speech-augmented benchmark based on Global MMLU Lite and a data set spanning English, Chinese, and Korean. |
| Outcome: | The proposed model is robust to demographic factors but sensitive to language and option order, suggesting that speech can amplify structural biases. |
Pushing the Frontiers of Scientific Fact-Checking: The SCINLP Dataset (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly being used to understand how scientific research evolves, drawing growing interest from the research community. |
| Approach: | They propose a scientific fact-checking dataset, SCINLP, tailored to the NLP domain that verifies the veracity of scientific research questions across varying rationale contexts. |
| Outcome: | The proposed framework examines scientific claims and research focus from a curated collection of influential and reputable NLP papers published between 2000 and 2024. |
SCAN: Semantic Document Layout Analysis for Textual and Visual Retrieval-Augmented Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent research indicates that using VLMs yields better RAG performance, but processing rich documents remains a challenge. |
| Approach: | They propose a VLM-friendly approach that enhances both textual and visual RAG systems. |
| Outcome: | The proposed approach outperforms conventional methods and commercial document processing solutions. |
Unified Multimodal Interleaved Document Representation for Retrieval (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods focus on textual content, ignoring the fact that documents can contain multiple modalities. |
| Approach: | They propose a method that holistically embeds documents interleaved with multiple modalities . they use vision-language models that combine text, images, and tables into a unified format . |
| Outcome: | The proposed method outperforms baselines on textual and multimodal queries. |
TELLME: Test-Enhanced Learning for Language Model Enrichment (2026.findings-eacl)
Copied to clipboard
Minjun Kim, Inho Won, HyeonSeok Lim, MinKyu Kim, Junghun Yuk, Wooyoung Go, Jongyoul Park, Jungyeul Park, KyungTae Lim
| Challenge: | Continual pre-training (CPT) has been widely adopted as a method for domain expansion in large language models, but has faced challenges such as acquiring large-scale domain-specific datasets and high computational costs. |
| Approach: | They propose a method that integrates the Test-Enhanced Learning principle with CPT to promote efficient domain-specific knowledge acquisition and long-term memory retention. |
| Outcome: | The proposed method outperforms existing methods by 23.6% in the financial domain and achieves 9.8% improvement in long-term memory retention. |
Beyond Accuracy: Alignment and Error Detection across Languages in the Bi-GSM8K Math-Teaching Benchmark (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent advances in LLMs have significantly improved mathematical problem-solving, with models like GPT-4 achieving human-level performance. |
| Approach: | They propose a bilingual English-Korean dataset enriched with teacher solutions, student solutions, and annotations marking students’ initial errors. |
| Outcome: | The proposed model achieves high agreement with human judgments and lower latency and resource usage than commercial APIs, demonstrating strong computational efficiency. |
VN-MTEB: Vietnamese Massive Text Embedding Benchmark (2026.findings-eacl)
Copied to clipboard
| Challenge: | a lack of large-scale test datasets makes it difficult to evaluate AI models before deploying them in real-world projects. |
| Approach: | They propose a Vietnamese benchmark for embedding models that leverages large language models and embeddable models to translate and filter samples from the Massive Multilingual Text Embedding Benchmark. |
| Outcome: | The proposed benchmark outperforms existing models in Vietnamese and English tasks with 41 datasets. |
See More, Store Less: Memory-Efficient Resolution for Video Moment Retrieval (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing video moment retrieval methods rely on sparse frame sampling, risking information loss. |
| Approach: | a new video-based framework enhances memory efficiency while maintaining high information resolution . SMORE uses query-guided captions to encode semantics aligned with user intent . |
| Outcome: | a new framework improves memory efficiency while maintaining high information resolution . it achieves state-of-the-art performance on QVHighlights, Charades-STA, and ActivityNet-Captions benchmarks . |
RB-LoRA: Rank-Balanced Aggregation for Low-Rank Adaptation with Federated Fine-Tuning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Low-rank adaptation (LoRA) improves fine-tuning of foundation models by updating only compact adapter matrices . varying client device capabilities lead to different adapter ranks, causing rank heterogeneity that undermines aggregation. |
| Approach: | They propose a rank-balanced aggregation framework that decomposes each update into rank-wise components and aligns them using analytically derived weights. |
| Outcome: | Experiments on language and vision models show that RB-LoRA improves under one and three rounds of communication in federated learning environments. |
Garbage In, Reasoning Out? Why Benchmark Scores are Unreliable and What to Do About It (2026.findings-eacl)
Copied to clipboard
| Challenge: | Using social reasoning benchmarks, we uncover pervasive flaws in both benchmark items and evaluation methodology. |
| Approach: | They audit three widely used social reasoning benchmarks and identify flaws in their design and evaluation methodology. |
| Outcome: | The results challenge the validity of current benchmark-based claims about social reasoning in large language models. |
Confidence-Driven Multi-Scale Model Selection for Cost-Efficient Inference (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have revolutionized inference across diverse natural language tasks, with larger models performing better but at higher computational costs. |
| Approach: | They propose a confidence-driven strategy that dynamically selects the most suitable model based on confidence estimates. |
| Outcome: | The proposed approach reduces token usage by approximately 60% and improves cost efficiency on the Massive Multitask Language Understanding (MMLU) benchmark. |
Quantifying the Impact of Structured Output Format on Large Language Models through Causal Inference (2026.findings-eacl)
Copied to clipboard
| Challenge: | Prior studies have examined the impact of structured output on LLMs’ generation quality, often presenting one-way findings. |
| Approach: | They propose to derive five potential causal structures characterizing the influence of structured output on LLMs’ generation using one assumed and two guaranteed constraints. |
| Outcome: | The proposed pipeline can be extended to other modules and is not limited to structured output but can be used in industrial applications. |
Breaking the Illusion of Reasoning in Polish LLMs: Quality over Quantity of Thought (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent advances in large language models have introduced explicit reasoning capabilities . however, the precise role of reasoning in improving model performance remains unclear . |
| Approach: | They disentangle effects of reasoning quality and sequence length by fine-tuning 8B models on Polish variants of the Mixture-of-Thoughts dataset. |
| Outcome: | The proposed model trained on high-quality reasoning traces achieved better average performance than other models. |
RV-Syn: Rational and Verifiable Mathematical Reasoning Data Synthesis based on Structured Function Library (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for generating high-quality reasoning data are limited in quality and availability. |
| Approach: | They propose a method that constructs mathematical operations and generates verifiable graphs that are back-translated into complex problems. |
| Outcome: | The proposed method achieves a 6.3% performance gain over existing methods on LLaMA-3-8B and outperforms others with only half the training data (50k vs. 100k). |
WebNovelBench: Placing LLM Novelists on the Web Novel Distribution (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing benchmarks for long-form novel generation lack scale, diversity, or objective measures. |
| Approach: | They propose a framework that assesses long-form novel generation using an LLM-as-Judge approach. |
| Outcome: | The proposed framework differentiates between human-written masterpieces, popular web novels, and LLM-generated content. |
From Semantics to Style: A Cross-Dataset Comparative Framework for Sentence Similarity Predictions (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing frameworks for analyzing text embedding models are limited. |
| Approach: | They propose a framework that uses lightweight poolers to analyze STS, PI, and Triplet datasets. |
| Outcome: | The proposed framework shows that the model captures semantic differences between sentences and is consistent across datasets. |
Feature Drift: How Fine-Tuning Repurposes Representations in LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Sparse autoencoders (SAEs) are a powerful tool for interpreting neural networks by extracting concepts (features) represented in their activations. |
| Approach: | They propose to use Sparse Autoencoders to extract concepts from their activations to explain how fine-tuning changes model capabilities. |
| Outcome: | The proposed model recombines existing concepts rather than learning new ones, and shows that it is a better explanation for how fine-tuning changes model capabilities. |
Detecting Winning Arguments with Large Language Models and Persuasion Strategies (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent studies have focused on predicting winning arguments, i.e., those that effectively convince a reader to adopt a certain opinion. |
| Approach: | They propose to use large language models with a chain-of-thought framework to guide reasoning over six persuasion strategies to determine persuasiveness. |
| Outcome: | The proposed approach leverages large language models with a chain-of-thought framework that guides reasoning over six persuasion strategies. |
The Devil is in the Distributions: Explicit Modeling of Scene Content is Key in Zero-Shot Video Captioning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for zero-shot video captioning focus on one key aspect of the scene and ignore the rest of the visual input. |
| Approach: | They propose a novel textual prompting strategy for zero-shot video captioning that uses a category-aware retrieval mechanism to promote prompt diversity while ensuring visual relevance. |
| Outcome: | The proposed method outperforms existing methods on in-domain and cross-domain settings. |
Best-of-L: Cross-Lingual Reward Modeling for Mathematical Reasoning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent studies have focused on improving reasoning ability in English models, with multilingual models receiving comparatively little attention. |
| Approach: | They propose a framework that ranks candidate reasoning traces across languages rather than within a single language. |
| Outcome: | The proposed framework improves accuracy by up to 10 points in English compared to using reward modeling within a single language. |
Nuanced Toxicity Detection in Spanish: A New Corpus and Benchmark Study (2026.findings-eacl)
Copied to clipboard
Alba María Mármol-Romero, Robiert Sepúlveda-Torres, Estela Saquete, María-Teresa Martín-Valdivia, L. Alfonso Ureña
| Challenge: | Existing corpora for Spanish are under-resourced for toxic content detection . sarcasm, indirect aggression, irony, and other toxicity are not detected in English . |
| Approach: | They propose to extend the NECOS-TOX corpus to include 4,011 Spanish comments . each comment is annotated across three levels of toxicity, with substantial inter-annotator agreement . |
| Outcome: | The proposed model performs on par with larger models and is released publicly . the proposed model is based on a human-in-the-loop active learning strategy . |
Persona Switch: Mixing Distinct Perspectives in Decoding Time (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies show that role-play prompting improves zero-shot reasoning, but these improvements are inconsistent across tasks and instances. |
| Approach: | They propose a method that dynamically combines the benefits of both prompting strategies. |
| Outcome: | The proposed method outperforms baselines and shows that output confidence is an important measure for selecting the more reliable output. |
Revealing the Truth with ConLLM for Detecting Multi-Modal Deepfakes (2026.findings-eacl)
Copied to clipboard
Gautam Siddharth Kashyap, Harsh Joshi, Niharika Jain, Ebad Shabbir, Jiechao Gao, Nipun Joshi, Usman Naseem
| Challenge: | Existing methods for deepfake detection suffer from two limitations: modality fragmentation and shallow inter-modal reasoning. |
| Approach: | They propose a framework for multimodal deepfake detection that uses contrastive learning and large language models to mitigate modality fragmentation and refine embeddings to address shallow inter-modal reasoning. |
| Outcome: | ConLLM reduces audio deepfake EER by 50%, improves video accuracy by 8%, and achieves approximately 9% accuracy gains in audio-visual tasks. |
Detection of Adversarial Prompts with Model Predictive Entropy (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly deployed in high-impact scenarios raising concerns about their safety and security. |
| Approach: | They propose an attack-agnostic pipeline for detecting adversarial inputs without prior knowledge of attack specifications. |
| Outcome: | The proposed pipeline outperforms traditional defenses in terms of adaptability and resource efficiency. |
Actors, Frames and Arguments: A Multi-Decade Computational Analysis of Climate Discourse in Financial News using Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | a new study examines how financial news media portrays climate change . financial news is the nervous system of the global economy . |
| Approach: | They propose a three-stage Actor–Frame–Argument pipeline that uses large language models to extract actors, stances, frames, and argumentative structures from a 980,061-article corpus. |
| Outcome: | The proposed pipeline extracts actors, stances, frames, and argumentative structures from a 980,061-article corpus of climate-related financial news from the Dow Jones Newswire (2000–2023) it is based on a human-annotated gold standard and a Decompositional Verification Framework (DVF) that decomposes evaluation into completeness, faithfulness, coherence, and relevance, with multi-judge scoring calibrated against human ratings. |
RECAP: REwriting Conversations for Intent Understanding in Agentic Planning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Understanding user intent is essential for effective conversational assistants . however, real-world dialogues are often ambiguous, underspecified, or dynamic . |
| Approach: | They propose a benchmark to evaluate intent rewriting in user-agent dialogues . they propose rewriters that reframe user-goal dialogues into concise representations of user goals . |
| Outcome: | The proposed benchmark outperforms baselines in terms of plan preference and fine-tuning two DPO-based rewriters yields additional utility gains. |
Modeling Turn-Taking with Semantically Informed Gestures (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing computational models of turn-taking relied on verbal cues and prosody. |
| Approach: | They propose a framework that integrates text, audio, and gestures to model multimodal turn-taking using semantic annotations. |
| Outcome: | The proposed framework shows that incorporating semantically guided gestures yields consistent performance gains over baselines. |
Do Large Language Models Reflect Demographic Pluralism in Safety? (2026.findings-eacl)
Copied to clipboard
Usman Naseem, Gautam Siddharth Kashyap, Sushant Kumar Ray, Rafiq Ali, Ebad Shabbir, Abdullah Mohammad
| Challenge: | Existing datasets that focus on demographics and safety are narrow in their annotator pools. |
| Approach: | They propose to decouple value framing from responses by modeling pluralism directly at the prompt level. |
| Outcome: | Demo-SafetyBench decouples value framing from responses to model pluralism at the prompt level. |
Adversarial Decoding: Generating Readable Documents for Adversarial Objectives (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for generating adversarial documents produce gibberish that is easy to detect and filter out. |
| Approach: | They propose a generic text generation technique that produces readable adversarial documents . they demonstrate that adversarials can be used for different objectives . |
| Outcome: | The proposed technique outperforms existing methods while producing readable documents for adversarial objectives. |
MEDAL: A Framework for Benchmarking LLMs as Multilingual Open-Domain Dialogue Evaluators (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing meta-evaluation benchmarks are static, outdated, and lacking in multilingual coverage. |
| Approach: | They propose a framework for curating more representative open-domain dialogue evaluation benchmarks . they leverage several LLMs to generate user-chatbot multilingual dialogues conditioned on varied seed contexts based on a state-of-the-art LLM . |
| Outcome: | The proposed framework exploits state-of-the-art LLMs to perform multilingual evaluations of open-domain chatbots. |
Which Works Best for Vietnamese? A Practical Study of Information Retrieval Methods across Domains (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies on Large Language Models (LLMs) are limited to single domains or curated datasets. |
| Approach: | They propose a domain-normalized, multi-domain benchmark for Vietnamese IR . they evaluate lexical, neural-sparse, late-interaction, dense, and hybrid paradigms . |
| Outcome: | The proposed benchmarks cover six domains and ten datasets across education, legal, healthcare, customer support, lifestyle reviews, and open-domain knowledge. |
MemeWeaver: Inter-Meme Graph Reasoning for Sexism and Misogyny Detection (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to detect hate speech on social media are limited by heuristic graph construction, shallow modality fusion, and instance-level reasoning. |
| Approach: | They propose a multimodal framework for detecting sexism and misogyny using a graph reasoning mechanism that can be used to train multiple visual-textual fusion strategies. |
| Outcome: | The proposed framework outperforms state-of-the-art methods on MAMI and EXIST benchmarks while achieving faster training convergence. |
SEAM: Bridging the Temporal-Semantic Granularity Gap for LLM-based Speech Recognition (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing duration-based methods generate embeddings at fixed rates, creating distributional mismatch with LLM pre-training. |
| Approach: | They propose an encoder-decoder architecture that generates embeddings at variable rates through cross-attention between speech features and text embeddables. |
| Outcome: | The proposed architecture achieves competitive performance on LibriSpeech (2.6%/5.2% WER) and 4.7% WER on TED-LIUM-v2 with a multi-stage training strategy and First Token Guidance. |
Foundations of LLM Knowledge Materialization: Termination, Reproducibility, Robustness (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models encode substantial factual knowledge, yet measuring and systematizing it remains challenging. |
| Approach: | They systematically analyze LLM knowledge materialization using miniGPTKBs . they find high termination rates, though model-dependent, and mixed reproducibility . |
| Outcome: | The proposed model can reliably surface core knowledge, but it has limitations. |
Investigating Gender Stereotypes in Large Language Models via Social Determinants of Health (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing benchmarks evaluate biases related to individual social determinants of health (SDoH) but they overlook interactions between these factors and lack context-specific assessments. |
| Approach: | They investigated the relationship between gender and other SDoH in french patient records to determine whether LLMs rely on embedded stereotypes to make gendered decisions. |
| Outcome: | The proposed models can probe stereotypes and make gendered decisions based on the data. |
FOL-Traces: Verified First-Order Logic Reasoning Traces at Scale (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to evaluate language models fail to provide structural clarity and verifiable inference. |
| Approach: | They propose to use a large-scale dataset of programmatically verified reasoning traces to evaluate structured logical inference. |
| Outcome: | The proposed model achieves 45.7% accuracy on masked operation prediction and 27% on two-step completion. |
Uncertainty Quantification for Evaluating Gender Bias in Machine Translation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing models can reproduce existing social inequalities but cannot be reduced. |
| Approach: | They propose that models should maintain uncertainty when input is ambiguous to avoid reinforcing biases. |
| Outcome: | The proposed model can detect gender bias when translated to ambiguous and unambiguous sources and shows that it does not correlate with high translation accuracy and debiases the two cases differently. |
PIRA: Preference-Oriented Instruction-Tuned Reward Models with Dual Aggregation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to align large language models with human preferences are limited by their large-scale annotation and prone to reward overoptimization. |
| Approach: | They propose a training paradigm that integrates three complementary strategies to address these challenges by reformulating question–answer pairs into preference-task instructions, averaging the rewards aggregated from diverse preference- task instructions for each sample, and a balancing outputs from the value head under different dropout rates. |
| Outcome: | Experiments on public datasets show that PIRA improves performance considerably, enhances generalization, and effectively mitigates reward overoptimization. |
The Hidden Bias: A Study on Explicit and Implicit Political Stereotypes in Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly integral to information dissemination and decision-making processes. |
| Approach: | They investigate political bias and stereotype propagation across eight prominent LLMs using the two-dimensional Political Compass Test. |
| Outcome: | The political bias and stereotype propagation of large language models is investigated using the two-dimensional Political Compass Test (PCT) key findings reveal a left-leaning political alignment across all investigated models. |
TIPA: Typologically Informed Parameter Aggregation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Massively multilingual language models enable cross-lingual generalization but underperform on low-resource and unseen languages. |
| Approach: | They propose a typologically informed framework that constructs proxy language adapters by aggregating existing ones, weighted by typological similarity. |
| Outcome: | The proposed framework outperforms baselines on five NLP tasks and over 230 languages. |
Can Calibration of Positional Encodings Enhance Long Context Utilization? (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models suffer from positional biases that reduce effective utilization of long contexts. |
| Approach: | They propose a training-free framework for calibrating Positional Encodings at inference time. |
| Outcome: | The proposed framework improves on needle-in-a-haystack and cross-chunk reasoning benchmarks and provides a lightweight method for improving long-context utilization. |
FiNERweb: Datasets and Artifacts for Scalable Multilingual Named Entity Recognition (2026.findings-eacl)
Copied to clipboard
| Challenge: | Named entity recognition (NER) is the task of identifying tokens that belong to a predefined set of classes such as "person" or "location" |
| Approach: | They propose a dataset-creation pipeline that scales the teacher-student paradigm to 91 languages and 25 scripts. |
| Outcome: | The proposed model achieves comparable or improved performance in English, Thai, and Swahili despite being trained on 19x less data than strong baselines. |
Bias in the East, Bias in the West: A Bilingual Analysis of LLM Political Bias on U.S.- and China-Related Issues (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) can exhibit political biases, which creates a risk of undue influence on LLM users and public opinion. |
| Approach: | They use a dataset of 36k real-time test prompts to measure LLM political bias on U.S. and Chinese issues. |
| Outcome: | The proposed model origin and prompt language influence bias on 60 political issues. |
Ask Me Again Differently: GRAS for Measuring Bias in Vision Language Models on Gender, Race, Age, and Skin Tone (2026.findings-eacl)
Copied to clipboard
| Challenge: | Using vision language models, we examine demographic biases in VLMs across gender, race, age, and skin tone. |
| Approach: | They propose a benchmark for uncovering demographic biases in Vision Language Models . they propose 'Gras Bias Score' to quantify bias in VLMs based on gender, race, age and skin tone . |
| Outcome: | The proposed model achieves 98, far from the unbiased ideal of 0. |
A Simple and Efficient Learning-Style Prompting for LLM Jailbreaking (2026.findings-eacl)
Copied to clipboard
| Challenge: | Learning-style queries can reliably elicit harmful responses, highlighting a critical safety blind spot in modern LLMs. |
| Approach: | They propose a new reframing paradigm that hides intention by learning from LLMs and uses 4 conceptual components to construct learning-style queries. |
| Outcome: | The proposed framework achieves top attack success rates on most models and across malicious categories while maintaining high efficiency with concise prompts. |
Aggregating Crowd of LLMs for Cost-Effective Data Annotation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent advances in Large Language Models (LLMs) have shown promise for automated data annotation, yet reliance on expensive commercial models like GPT-4 limits accessibility. |
| Approach: | They propose to build a crowd of LLMs which aggregates annotations from multiple sLLMs using label aggregation algorithms. |
| Outcome: | The proposed approach outperforms individual sLLMs and human crowd labels yields superior results compared to either method alone. |
Representation Collapse in Machine Translation Through the Lens of Angular Dispersion (2026.findings-eacl)
Copied to clipboard
| Challenge: | Effective training of Transformer models for sequential language tasks is difficult due to various forms of collapse of the internal representations learned. |
| Approach: | They propose to use angular dispersion to analyze representation collapse at different levels of discrete and continuous transformers throughout training. |
| Outcome: | The proposed method mitigates collapse and improves translation quality. |
Can LLMs Reason Like Doctors? Exploring the Limits of Large Language Models in Complex Medical Reasoning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have shown remarkable progress in reasoning across multiple domains, but it remains unclear whether their abilities reflect genuine reasoning or sophisticated pattern matching. |
| Approach: | They conduct one of the largest evaluations to date, assessing 77 LLMs . they select three medical question answering (QA) benchmarks targeting reasoning processes . |
| Outcome: | The results highlight the need to improve specific reasoning strategies to better reflect medical decision-making. |
Testing Low-Resource Language Support in LLMs Using Language Proficiency Exams: the Case of Luxembourgish (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are used in research and society at large, but are mostly developed with English-speaking users in mind. |
| Approach: | They investigate the viability of language proficiency exams as evaluation tools for Luxembourgish . large models such as Claude and DeepSeek-R1 typically achieve high scores . |
| Outcome: | The proposed models can predict performance in Luxembourgish language tests. |
Unveiling Decision-Making in LLMs for Text Classification : Extraction of influential and interpretable concepts with Sparse Autoencoders (2026.findings-eacl)
Copied to clipboard
| Challenge: | Concept-based explanations for large language models are not well understood in text classification. |
| Approach: | They propose a model with a specialized classifier head and activation rate sparsity loss for sentence classification . they compare it to existing models with HI-Concept and ConceptShap . |
| Outcome: | The proposed model improves both the causality and interpretability of the extracted features. |
TextMineX: Data, Evaluation Framework and Ontology-guided LLM Pipeline for Humanitarian Mine Action (2026.findings-eacl)
Copied to clipboard
| Challenge: | Humanitarian Mine Action (HMA) authorities publish large amount of life-saving operational knowledge, but much remains locked away in unstructured reports. |
| Approach: | They propose a dataset, evaluation framework and ontology-guided large language model pipeline for knowledge extraction from text in the HMA domain. |
| Outcome: | The proposed framework improves extraction accuracy by 44.2% and reduces hallucinations by 22.5% . the proposed framework can be used to analyze human-annotated triples and an LLM-as-Judge protocol . |
MathMist: A Parallel Multilingual Benchmark Dataset for Mathematical Problem Solving and Reasoning (2026.findings-eacl)
Copied to clipboard
Mahbub E Sobhani, Md. Faiyaz Abdullah Sayeedi, Tasnim Mohiuddin, Md Mofijul Islam, Swakkhar Shatabda
| Challenge: | Existing benchmarks primarily focus on English or a narrow subset of high-resource languages, leaving significant gaps in assessing multilingual and cross-lingual mathematical reasoning. |
| Approach: | They propose a parallel multilingual benchmark for mathematical problem solving and reasoning that encompasses 2,890 parallel Bangla-English gold standard artifacts. |
| Outcome: | The proposed model encompasses 2,890 parallel Bangla-English gold standard artifacts, totaling 30K aligned question–answer pairs across thirteen languages, representing high-, medium-, and low-resource linguistic settings. |
Enhancing Reliability in Community Question Answering with an Expert-Oriented RAG System (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing LLMs require substantial computational resources and are prone to generating hallucinated or unreliable content. |
| Approach: | They propose an expert-oriented Retrieval-Augmented Generation framework which leverages user modeling to identify archived questions with answers that fully or partially address the user’s new query. |
| Outcome: | The proposed framework synthesizes expert-written answers from similar questions to generate unified answers. |
Unsupervised Text Style Transfer for Controllable Intensity (2026.findings-eacl)
Copied to clipboard
| Challenge: | Unsupervised Text Style Transfer (UTST) aims to transfer the stylistic properties of a given text without parallel text pairs. |
| Approach: | They propose a SFT-then-PPO paradigm to fine-tune an LLM with parallel data and reward functions for distinguishing stylistic intensity in hierarchical levels. |
| Outcome: | The proposed system can transfer stylistic properties without parallel text pairs even for adjacent levels of intensity. |
SchemaGraphSQL: Efficient Schema Linking with Pathfinding Graph Algorithms for Text-to-SQL on Large-Scale Databases (2026.findings-eacl)
Copied to clipboard
AmirHossein Safdarian, Milad Mohammadi, Ehsan Jahanbakhsh Bashirloo, Mona Shahamat Naderi, Heshaam Faili
| Challenge: | Text-to-SQL systems translate natural language questions into executable SQL queries. |
| Approach: | They propose a schema linking approach that first constructs a graph based on foreign key relations and then uses a single prompt to a lightweight LLM to extract source and destination tables from the user query. |
| Outcome: | The proposed method outperforms specialized, fine-tuned, and complex multi-step approaches on BIRD and Spider 2.0 benchmarks. |
Binary Token-Level Classification with DeBERTa for All-Type MWE Identification: A Lightweight Approach with Linguistic Enhancement (2026.findings-eacl)
Copied to clipboard
| Challenge: | Current approaches focus on specific MWE types, such as transformer-based models that incorporate linguistic features like dependency parsing for verbal discontinuous patterns. |
| Approach: | They propose a binary token-level classification approach that integrates linguistic feature integration and data augmentation to improve multiword expression (MWE) identification. |
| Outcome: | The proposed model outperforms the Qwen-72B model on the CoAM dataset by 12 points while using 165 times fewer parameters. |
The Correlation Between Emotion in Text and Speech Segments is Limited: A Cross-Modal Study (2026.findings-eacl)
Copied to clipboard
| Challenge: | a recent study has shown that text-to-speech systems can capture human-like emotion, but they lack the ability to predict emotion in speech. |
| Approach: | They propose to use 8 large language models for identifying emotion in text and 2 audio models for emotion in speech to investigate the correlation between emotion and speech. |
| Outcome: | The proposed models perform well on emotion recognition from situational text and audiobooks, but show weak correlation for Valence only. |
Seeing All Sides: Multi-Perspective In-Context Learning for Subjective NLP (2026.findings-eacl)
Copied to clipboard
| Challenge: | Modern language models excel at factual reasoning but struggle with value diversity, authors say . task-sensitive tasks such as hate speech expose this limitation . human disagreement captures the diversity of plausible human perspectives, authors argue . |
| Approach: | They evaluate four large language models with human disagreements on five datasets . they find multi-perspective in-context learning outperforms standard prompting . |
| Outcome: | The proposed approach outperforms standard prompting on English labels while disaggregated soft predictions better align with human judgments in Arabic and Italian datasets. |
Beyond Divergent Creativity: A Human-Based Evaluation of Creativity in Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models are increasingly used in verbal creative tasks. |
| Approach: | They propose a divergent association task that focuses on novelty, ignoring appropriateness, a core component of creativity. |
| Outcome: | The proposed model scores are lower than baselines with no creative abilities, undermining its validity for model evaluation. |
Are Multimodal LLMs Movie Buffs? (2026.findings-eacl)
Copied to clipboard
| Challenge: | Multimodal large language models are increasingly used for movie understanding . however, their performance on movies lags behind other video understanding tasks . |
| Approach: | They analyze movie knowledge, cinematographic knowledge, and critical analysis to identify where MLLMs fail . ML models are increasingly used for movie understanding . |
| Outcome: | The results show that MLLMs outperform existing methods in small-scale settings involving factual knowledge but fail when cinematographic and critical analysis is required. |
Process Evaluation for Agentic Systems (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent adoption of LLM-based assistants has led to premature assumptions about their reliability and general capability. |
| Approach: | They propose to assess the feasibility of automatic process evaluation for critical applications such as medicine, finance, law and infrastructure. |
| Outcome: | The proposed evaluations are based on a small-scale study to assess the feasibility of automated process evaluation, present a compliance score, analyse use cases of bad and good behaviours, and offer recommendations for more holistic evaluation. |
MIMIC: Multi-party Dialogue Augmentation via Speaker Stylistic Transfer (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing discourse annotations are limited and annotated data scarcity has hindered progress in discourse parsing. |
| Approach: | They propose a framework for augmenting discourse-annotated corpora via speaker stylistic transfer using Large Language Models (LLMs). |
| Outcome: | The proposed framework outperforms parsers trained on STAC and Molweni corpora on a multi-party dialogue with consistent gains for underrepresented discourse patterns and in low-resource scenarios. |
TechING: Towards Real World Technical Image Understanding via VLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Modern day vision language models struggle when it comes to understanding technical diagrams . a large synthetically generated corpus is needed to train and evaluate VLMs on hand-drawn images . |
| Approach: | They propose a large synthetically generated corpus for training VLMs and evaluate them on hand-drawn images. |
| Outcome: | The proposed model improves ROUGE-L performance of Llama 3.2 11B-instruct by 2.14x on synthetic images on real-world images. |
Multilingual-To-Multimodal (M2M): Unlocking New Languages with Monolingual Text (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing multimodal models rely on machine translation, but performance drops for other languages due to limited multilingual multimodal resources. |
| Approach: | They propose a lightweight alignment method that learns only a few linear layers using English text alone to map multilingual text embeddings into multimodal space. |
| Outcome: | M2M achieves strong zero-shot transfer on XTD Text-to-Image retrieval in English and spanish . it learns only a few linear layers to map multilingual text embeddings into multimodal space . |
Do GUI Grounders Truly Understand UI Elements? (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing grounding models and benchmarks are skewed toward web and mobile environments, neglecting desktop interfaces (especially windows). |
| Approach: | They propose a GUI Grounding Sensitivity Benchmark to assess UI grounding sensitivity to multiple descriptions of the same UI element. |
| Outcome: | The proposed model generates multiple valid instructions per UI element and develops nuanced validation methods to validate them. |
Ensemble Privacy Defense for Knowledge-Intensive LLMs against Membership Inference Attacks (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are the foundation of modern natural language processing, powering applications across diverse domains. |
| Approach: | They propose a model-agnostic defense framework which aggregates and evaluates the outputs of a knowledge-injected LLM, a base LLM and a dedicated judge model to enhance resistance against membership inference attacks. |
| Outcome: | The proposed framework reduces MIA success by up to 27.8% for SFT and 526.3% for RAG compared to inference-time baseline while maintaining answer quality. |
SafeSearch: Do Not Trade Safety for Utility in LLM Search Agents (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language model (LLM) based search agents are more likely to produce harmful outputs than base models. |
| Approach: | They propose a query-level shaping term that rewards safe queries and penalizes unsafe ones. |
| Outcome: | The proposed approach reduces harmfulness by over 70% across three red-teaming datasets while producing safe, helpful responses. |
SAGE : A Top-Down Bottom-Up Knowledge-Grounded User Simulator for Multi-turn Agent Evaluation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing evaluation methods rely on static benchmarks or narrow task-specific datasets that fail to capture the open-ended nature of real-world interactions. |
| Approach: | They propose a user Simulation framework for multi-turn AGent Evaluation that integrates top-down knowledge from business contexts and bottom-up knowledge from agent infrastructure. |
| Outcome: | The proposed framework produces interactions that are more realistic and diverse while identifying up to 33% more agent errors. |
Better as Generators Than Classifiers: Leveraging LLMs and Synthetic Data for Low-Resource Multilingual Classification (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable multilingual capabilities, making them promising tools in both high- and low-resource languages. |
| Approach: | They use a multilingual LLM to generate synthetic datasets covering 11 languages and 4 classification tasks and use them to train smaller models. |
| Outcome: | The proposed model outperforms the large generator in low-resource languages and tasks. |
Dialogue is Better Than Monologue: Instructing Meidcal LLMs via Strategic Conversations (2026.findings-eacl)
Copied to clipboard
Zijie Liu, Xinyu Zhao, Jie Peng, Jinhao Duan, Zhuangdi Zhu, Qingyu Chen, Kaidi Xu, Xia Hu, Tianlong Chen
| Challenge: | Existing tuning methods for medical AI models are monologue-based . existing benchmarks are based on licensing exams or research articles . |
| Approach: | They propose a benchmark to expose limitations of monologue-based tuning for medical AI models . they use a large dialogue dataset to capture stepwise diagnostic reasoning . |
| Outcome: | The proposed model outperforms monologue-tuned models on a medical question answering task and improves accuracy on standard medical QA benchmarks. |
DF-RAG: Query-Aware Diversity for Retrieval-Augmented Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Retrieval-augmented generation (RAG) is a common technique for grounding language models in domain-specific information. |
| Approach: | They propose a new retrieval technique that incorporates diversity into the retrieval step to improve performance on reasoning-intensive QA benchmarks. |
| Outcome: | The proposed method outperforms baselines on reasoning-intensive QA benchmarks by 4–10%. |
Hearing Between the Lines: Unlocking the Reasoning Power of LLMs for Speech Evaluation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Model (LLM) judges are limited to textual content, resulting in expensive and opaque evaluation methods. |
| Approach: | They propose a framework that enables large language model judges to reason over audio cues . they introduce a human chain-of-thought annotation protocol to improve judge diagnostic capability . |
| Outcome: | The proposed framework achieves higher agreement with human raters than ALMs and transcript-only LLM judges while being significantly more cost-effective. |
Improving Chain-of-Thought for Logical Reasoning via Attention-Aware Intervention (2026.findings-eacl)
Copied to clipboard
| Challenge: | Modern logical reasoning with LLMs relies on employing complex interactive frameworks that decompose the reasoning process into subtasks solved through carefully designed prompts or requiring external components, which limit their scalability. |
| Approach: | They propose a non-interactive, end-to-end framework for reasoning tasks that enables reasoning to emerge within the model itself. |
| Outcome: | The proposed framework improves generalization while preserving analyzability without external resources. |
Thinking Long, but Short: Stable Sequential Test-Time Scaling for Large Reasoning Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Inducing models to think for longer can increase accuracy, but as the length of reasoning is further extended, it has also been shown to result in accuracy degradation and model instability. |
| Approach: | They propose a sequential test-time scaling method which induces models to think for longer, but which also generates an increasingly long output. |
| Outcome: | The proposed method improves model accuracy significantly over a wide range of induced thoughts, stabilizing the accuracy of sequential scaling, and eliminating the need for reasoning length fine-tuning. |
Defeating Cerberus: Privacy-Leakage Mitigation in Vision Language Models (2026.findings-eacl)
Copied to clipboard
Boyang Zhang, Istemi Ekin Akkus, Ruichuan Chen, Alice Dethise, Klaus Satzke, Ivica Rimac, Yang Zhang
| Challenge: | Existing models that process multiple modalities of data have been used for multimodal tasks, but their advanced capabilities raise privacy concerns. |
| Approach: | They propose a method to modify the model’s internal states associated with PII-related content and to reduce the risk of PI I leakage by modifying the model's internal state. |
| Outcome: | The proposed method achieves on average 93.3% refusal rate for various PII-related tasks with minimal impact on unrelated model performances. |
TruthTrap: A Bilingual Benchmark for Evaluating Factually Correct Yet Misleading Information in Question Answering (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly used to answer factual, information-seeking questions (ISQs). |
| Approach: | They propose to use a dataset to evaluate large language models to generate human-like text on ISQs in two languages, English and Farsi, and then use it to evaluate nine LLMs. |
| Outcome: | The proposed dataset shows that accuracy drops by 25% when models encounter misleading yet factual hints. |
FLAT-LLM: Fine-grained Low-rank Activation Space Transformation for Large Language Model Compression (2026.findings-eacl)
Copied to clipboard
| Challenge: | Low-rank decomposition methods suffer from accuracy degradation and expensive calibration procedures. |
| Approach: | They propose a fast and accurate, training-free structural compression method based on fine-grained low-rank transformations in the activation space. |
| Outcome: | The proposed method outperforms pruning baselines in generalization and downstream performance while delivering inference speedups. |
Negative Sampling Techniques in Dense Retrieval: A Survey (2026.findings-eacl)
Copied to clipboard
| Challenge: | Information Retrieval (IR) is fundamental to many modern NLP applications. |
| Approach: | They propose a taxonomy that categorizes negative sampling techniques in dense IR . they analyze them with respect to trade-offs between effectiveness, computational cost, implementation difficulty . |
| Outcome: | The proposed taxonomy categorizes techniques using random, static/dynamically mined, and synthetic datasets. |
Multi-Agent Procedural Graph Extraction with Structural and Logical Refinement (2026.findings-eacl)
Copied to clipboard
Wangyang Ying, Yanchi Liu, Xujiang Zhao, Wei Cheng, Zhengzhang Chen, Wenchao Yu, Yanjie Fu, Haifeng Chen
| Challenge: | Recent advances in large language models (LLMs) show potential for graph extraction, but often yield ill-formed structures or misinterpret logical constructs such as gateways. |
| Approach: | They propose a framework that treats procedural graph extraction as a multi-round reasoning process with structural and logical refinement agents. |
| Outcome: | The proposed framework achieves significant improvements in structural correctness and logical consistency over strong baselines. |
MADIAVE: Multi-Agent Debate for Implicit Attribute Value Extraction (2026.findings-eacl)
Copied to clipboard
| Challenge: | Implicit Attribute Value Extraction (AVE) is essential for accurately representing products in e-commerce due to the complexity of multidimensional data and gaps in vision-text understanding. |
| Approach: | They propose a multi-agent framework that employs multiple MLLM agents to iteratively refine inferences through debate rounds. |
| Outcome: | The proposed framework improves inference performance and robustness through debate rounds. |
DeepSieve: Information Sieving via LLM-as-a-Knowledge-Router (2026.findings-eacl)
Copied to clipboard
Minghao Guo, Qingcheng Zeng, Xujiang Zhao, Yanchi Liu, Wenchao Yu, Mengnan Du, Haifeng Chen, Wei Cheng
| Challenge: | Existing RAG methods lack fine-grained control over query and source sides, resulting in noisy retrieval and shallow reasoning. |
| Approach: | They propose an agentic RAG framework that integrates information sieving via LLM-as-a-knowledge-router. |
| Outcome: | Experiments on multi-hop QA tasks across heterogeneous sources demonstrate improved reasoning depth, retrieval precision, and interpretability over conventional approaches. |
Analyzing LLM Instruction Optimization for Tabular Fact Verification (2026.findings-eacl)
Copied to clipboard
Xiaotang Du, Giwon Hong, Wai-Chung Kwan, Rohit Saxena, Ivan Titov, Pasquale Minervini, Emily Allaway
| Challenge: | evaluating instruction optimization for tabular fact verification is a key challenge for reliable NLP systems. |
| Approach: | They compare instruction optimization for tabular fact verification with a framework based on DSPy . they find that instruction optimization consistently improves verification accuracy . |
| Outcome: | The proposed method improves verification accuracy across four benchmarks and three model families. |
XMAD-Bench: Cross-Domain Multilingual Audio Deepfake Benchmark (2026.findings-eacl)
Copied to clipboard
Ioan-Paul Ciobanu, Andrei-Iulian Hîji, Nicolae Catalin Ristea, Paul Irofti, Cristian Rusu, Radu Tudor Ionescu
| Challenge: | Recent advances in audio generation led to an increasing number of deepfakes . however, these methods are typically tested in an in-domain setup . |
| Approach: | They propose a large-scale cross-domain audio deepfake benchmark comprising 668.8 hours of real and deepfak speech. |
| Outcome: | The proposed benchmark compares audio deepfake detectors with existing methods in the wild . the results show that the proposed methods perform better in different languages than existing methods . |
CLEAR-3K: Assessing Causal Explanatory Capabilities in Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing natural language understanding benchmarks inadequately address the ability to evaluate causal relationships. |
| Approach: | They propose to use CLEAR-3K to evaluate whether language models can determine if one statement causally explains another. |
| Outcome: | The proposed questions show that language models often confuse semantic similarity with causality, relying on lexical and semantic overlap instead of inferring actual causal explanatory relationships. |
Imbalanced Gradients in RL Post-Training of Multi-Task LLMs (2026.findings-eacl)
Copied to clipboard
Runzhe Wu, Ankur Samanta, Ayush Jain, Scott Fujimoto, Jeongyeol Kwon, Ben Kretzu, Youliang Yu, Kaveh Hassani, Boris Vidolov, Yonathan Efroni
| Challenge: | Large-gradient tasks can achieve similar or even much lower learning gains than small-grading ones. |
| Approach: | They show that large-gradient tasks can achieve lower learning gains than small-grading ones . large-grade tasks can accomplish similar or even lower learning gain than small grade ones if they are large . |
| Outcome: | The proposed approach fails when certain tasks produce larger gradients . Large-gradient tasks can achieve lower learning gains than small-gradent ones . |
BayesFlow: A Probability Inference Framework for Meta-Agent Assisted Workflow Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable generality, often solving tasks with a single carefully engineered prompt. |
| Approach: | They propose to cast automatic workflow generation as Bayesian inference over a posterior distribution on workflows and instantiate BayesFlow as Bayer-based workflow generation framework. |
| Outcome: | The proposed framework improves accuracy by 9 percentage points over baselines and 65 percentage points on pool-wide benchmarks. |
HapticLLaMA: A Multimodal Sensory Language Model for Haptic Captioning (2026.findings-eacl)
Copied to clipboard
| Challenge: | haptic captioning is the task of generating natural language descriptions from haptics, such as vibrations, for use in virtual reality and rehabilitation applications. |
| Approach: | They propose a multimodal sensory language model that interprets vibration signals into descriptions in a given sensory, emotional, or associative category. |
| Outcome: | The proposed model interprets vibration signals into descriptions in a given sensory, emotional, or associative category. |
Token-Level Precise Attack on RAG: Searching for the Best Alternatives to Mislead Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to attack large language models rely heavily on retrieval and generation stages, limiting their effectiveness in black-box scenarios. |
| Approach: | They propose a retrieval-augmented generation framework that leverages a white-box LLM as an attacker to generate and iteratively optimize malicious passages at the token level. |
| Outcome: | The proposed framework outperforms existing approaches in retrieval-stage and end-to-end attacks on black-box RAG systems. |
PagedEviction: Structured Block-wise KV Cache Pruning for Efficient Large Language Model Inference (2026.findings-eacl)
Copied to clipboard
Krishna Teja Chitty-Venkata, Jie Ye, Siddhisanket Raskar, Anthony Kougkas, Xian Sun, Murali Emani, Venkatram Vishwanath, Bogdan Nicolae
| Challenge: | Large Language Models (LLMs) are exploding to large sizes, including GPT, LLaMA, and DeepSeek. |
| Approach: | They propose a fine-grained, structured KV cache pruning strategy that enhances the memory efficiency of vLLM’s PagedAttention. |
| Outcome: | The proposed method integrates seamlessly with PagedAttention without any modifications to its CUDA attention kernels. |
SMART-Editor: A Multi-Agent Framework for Human-Like Design Editing with Structural Integrity (2026.findings-eacl)
Copied to clipboard
| Challenge: | SMART-EDITOR is a framework for compositional layout and content editing for structured visual domains. |
| Approach: | They introduce a framework for compositional editing for structured images like posters or websites . SMART-EDITOR maintains global coherence through two complementary strategies . |
| Outcome: | The proposed framework maintains global coherence through two complementary strategies. |
ConVerse: Benchmarking Contextual Safety in Agent-to-Agent Conversations (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are rapidly transitioning from passive text generators to autonomous agents that act and communicate on behalf of users. |
| Approach: | a new benchmark evaluates privacy and security risks in agent–agent interactions . a converse model enables attackers to embed malicious requests within plausible discourse . the model is based on a three-tier taxonomy assessing abstraction quality . |
| Outcome: | ConVerse tests privacy and security risks in agent–agent interactions with 12 user personas and over 864 contextually grounded attacks. |
SIRAJ: Diverse and Efficient Red-Teaming for LLM Agents via Distilled Structured Reasoning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing red-teaming frameworks do not cover all the risks associated with arbitrary black-box LLMs. |
| Approach: | They propose a generic red-teaming framework for arbitrary black-box LLM agents that iteratively constructs and refines model-based adversarial attacks based on the execution trajectories of former attempts. |
| Outcome: | The proposed model improves attack success rate by 100%, surpassing the 671B Deepseek-R1 model. |
Who You Are, What You Say: Intra- and Inter- Context Personality for Emotion Recognition in Conversation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to Emotion Recognition in conversation (ERC) focus on modeling speaker dynamics within dialogues. |
| Approach: | They propose a personality-aware ERC framework that segregates conversational context into intra- and inter-speaker components and models static or dynamic personality traits to represent stable and evolving speaker dispositions. |
| Outcome: | The proposed framework improves weighted F1 by 2.74% over non-LLM methods and 0.98% over recent LLM-based methods. |
DRIVINGVQA: A Dataset for Interleaved Visual Chain-of-Thought in Real-World Driving Scenarios (2026.findings-eacl)
Copied to clipboard
| Challenge: | Chain-of-thought (CoT) prompting is a prompting strategy that improves reasoning in large language models, but its effectiveness in vision-language models remains limited due to over-reliance on textual cues and memorized knowledge. |
| Approach: | They propose a visual question-answering dataset derived from driving theory exams that incorporates textual explanations with visual tokens extracted from entities relevant to the reasoning process. |
| Outcome: | The proposed approach outperforms chain-of-thought prompting in large language models and vision-language models in real-world scenarios. |
SAGE: Steerable Agentic Data Generation for Deep Search with Execution Feedback (2026.findings-eacl)
Copied to clipboard
Fangyuan Xu, Rujun Han, Yanfei Chen, Zifeng Wang, I-Hung Hsu, Jun Yan, Vishy Tirumalashetty, Eunsol Choi, Tomas Pfister, Chen-Yu Lee
| Challenge: | High-quality, complex question-answer pairs are pivotal for training and evaluating capable deep search agents. |
| Approach: | They propose a pipeline that generates high-quality, difficulty-controlled deep search question-answer pairs for a given corpus and a target difficulty level. |
| Outcome: | The proposed pipeline generates high-quality, difficulty-controlled deep search question-answer pairs for a given corpus and a target difficulty level. |
Negative-Aware Diffusion Process for Temporal Knowledge Graph Extrapolation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Temporal Knowledge Graphs (TKGs) are dynamic structures representing entities and their evolving relationships through time. |
| Approach: | They propose a non-parametric model that encodes subject-centric histories into sequential embeddings. |
| Outcome: | The proposed model encodes subject-centric histories of entities, relations and temporal intervals into sequential embeddings. |
DS2-Instruct: Domain-Specific Data Synthesis for Large Language Models Instruction Tuning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing data synthesis methods focus on general-purpose tasks and fail to capture domain-specific terminology and reasoning patterns. |
| Approach: | They propose a framework that generates domain-specific instruction datasets without human supervision by pairing task-informed keywords with different cognitive levels from Bloom’s Taxonomy. |
| Outcome: | The proposed framework generates domain-specific instruction datasets without human supervision and achieves significant improvements over existing methods. |
DashboardQA: Benchmarking Multimodal Agents for Question Answering on Interactive Dashboards (2026.findings-eacl)
Copied to clipboard
Aaryaman Kartha, Ahmed Masry, Mohammed Saidul Islam, Thinh Lang, Shadikur Rahman, Ridwan Mahbub, Mizanur Rahman, Mahir Ahmed, Md Rizwan Parvez, Enamul Hoque, Shafiq Joty
| Challenge: | Existing question-answering benchmarks for data visualizations focus on static charts instead of interactive dashboards. |
| Approach: | They propose a benchmark to assess how vision-language GUI agents comprehend and interact with real-world dashboards. |
| Outcome: | The first benchmark explicitly designed to assess how vision-language GUI agents comprehend and interact with real-world dashboards. |
Reasoning about Uncertainty: Do Reasoning Models Know When They Don’t Know? (2026.findings-eacl)
Copied to clipboard
| Challenge: | Reasoning models are prone to generating confident, plausible responses that are incorrect (hallucinations). |
| Approach: | They introduce introspective uncertainty quantification to examine whether reasoning models are well-calibrated and does deeper reasoning improve their calibration? |
| Outcome: | The proposed model calibrations show that models are overconfident, overconfent and overconfust with deeper reasoning. |
AfriMMT-EA: Multi-domain Machine Translation for Low-Resource East African Languages (2026.findings-eacl)
Copied to clipboard
Naome A Etori, Kelechi Ezema, Nathaniel Romney Robinson, Davis David, Alfred Malengo Kondoro, Elisha Ondieki Makori, Michael Samwel Mollel, Maria Gini
| Challenge: | Recent advances in open-source large language models have demonstrated strong multilingual capabilities through data-efficient adaptation strategies. |
| Approach: | They propose to use AfriMMT-EA to refine two multilingual versions of Gemma-3 to better understand the region's linguistic and cultural diversity. |
| Outcome: | The proposed datasets comprise 54 local languages across five East African countries. |
Diffusion Language Model Inference with Monte Carlo Tree Search (2026.findings-eacl)
Copied to clipboard
Zheng Huang, Kiran Ramnath, Yueyan Chen, Aosong Feng, Sangmin Woo, Balasubramaniam Srinivasan, Zhichao Xu, Kang Zhou, Shuai Wang, Haibo Ding, Lin Lee Cheong
| Challenge: | Existing methods for inference use heuristics to determine which positions to unmask and which tokens to commit . MEDAL is an inference-time scaling framework that integrates Monte Carlo Tree SEarch initialization for Diffusion Language Model inference. |
| Approach: | They propose a framework that integrates Monte Carlo Tree SEarch initialization for Diffusion Language Model inference. |
| Outcome: | The proposed framework achieves 22.0% improvement over existing inference strategies across multiple benchmarks. |
DWA-KD: Dual-Space Weighting and Time-Warped Alignment for Cross-Tokenizer Knowledge Distillation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing cross-tokenizer distillation methods are limited by suboptimal alignment across sequence and vocabulary levels. |
| Approach: | They propose a cross-tokenizer distillation framework that enhances token-wise distillation . they use dual-space entropy-based weighting to achieve precise sequence-level alignment . |
| Outcome: | The proposed framework outperforms state-of-the-art methods in large language models but has high computational and memory costs. |
Harnessing Consistency for Robust Test-Time LLM Ensemble (2026.findings-eacl)
Copied to clipboard
Zhichen Zeng, Qi Yu, Xiao Lin, Ruizhong Qiu, Xuying Ning, Tianxin Wei, Yuchen Yan, Jingrui He, Hanghang Tong
| Challenge: | Existing efforts to improve LLM ensemble quality have focused on model consistency, but failures are often due to heterogeneous tokenization schemes and varying model expertise. |
| Approach: | They propose a plug-and-play technique that harnesses model consistency for robust LLM ensemble. |
| Outcome: | The proposed technique improves ensemble performance and robustness against erroneous signals. |
AutoAnoEval: Semantic-Aware Model Selection via Tree-Guided LLM Reasoning for Tabular Anomaly Detection (2026.findings-eacl)
Copied to clipboard
Suhee Yoon, Sanghyu Yoon, Ye Seul Sim, Seungdong Yoa, Dongmin Kim, Soonyoung Lee, Hankook Lee, Woohyung Lim
| Challenge: | Existing approaches to tabular anomaly detection fail to reflect domain specific nature of real-world anomalies. |
| Approach: | They propose a framework that constructs pseudo-evaluation sets with semantically grounded synthetic anomalies. |
| Outcome: | The proposed framework generates pseudo-evaluation sets with semantically grounded synthetic anomalies. |
Conformal Feedback Alignment: Quantifying Answer-Level Reliability for Robust LLM Alignment (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing uncertainty-aware approaches weight preferences, but ignore reliability of the answers being compared. |
| Approach: | They propose a framework that grounds preference weighting in Conformal Predictions to address this problem. |
| Outcome: | The proposed framework improves alignment robustness and data efficiency across different datasets. |
ThinkPilot: Steering Reasoning Models via Automated Think-prefixes Optimization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Reasoning Models (LRMs) are powerful but still suffer from inefficient and off-target reasoning. |
| Approach: | They propose a training-free framework that automatically optimizes Large Reasoning Models' reasoning by generating think-prefixes that evolve driven by a taxonomy of reasoning behaviors. |
| Outcome: | The proposed framework significantly improves accuracy-length trade-off for efficient reasoning, drastically improves safety and improves instruction following. |
LitE-SQL: A Lightweight and Efficient Text-to-SQL Framework with Vector-based Schema Linking and Execution-Guided Self-Correction (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods rely on proprietary models to generate SQL queries. |
| Approach: | They propose a lightweight framework that translates natural language questions into SQL queries. |
| Outcome: | The proposed framework achieves 72.10% execution accuracy on BIRD and 88.45% on Spider 1.0 . it offers a practical solution for privacy-sensitive and resource-constrained settings. |
Beyond Coherence: Improving Temporal Consistency and Interpretability in Dynamic Topic Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing topic models capture bag-of-words statistics but lack semantic priors . interpretability remains shallow, relying on noisy top-word lists that obscure thematic clarity. |
| Approach: | They propose a variational framework to capture more faithful temporal trajectories . they propose to use entropy-regularized optimal transport to align entire topic constellations . |
| Outcome: | The proposed framework captures more faithful temporal trajectories and improves interpretability. |
Interpretable Graph-Language Modeling for Detecting Youth Illicit Drug Use (2026.findings-eacl)
Copied to clipboard
Yiyang Li, Zehong Wang, Zhengqing Yuan, Zheyuan Zhang, Keerthiram Murugesan, Chuxu Zhang, Yanfang Ye
| Challenge: | Illicit drug use among teens and young adults remains a public health concern . existing models ignore latent and interconnected structures among survey variables . |
| Approach: | They propose a joint graph-language modeling framework to detect illicit drug use among TYAs . they use large-scale surveys such as the Youth Risk Behavior Survey and the National Survey on Drug Use and Health to analyze data . |
| Outcome: | The proposed framework outperforms baseline models on YRBS and NSDUH datasets in predictive accuracy. |
Tailoring Memory Granularity for Multi-Hop Reasoning over Long Contexts (2026.findings-eacl)
Copied to clipboard
| Challenge: | Extensive experiments on long-context multi-hop question answering benchmarks show TAG achieves state-of-the-art performance. |
| Approach: | They propose a framework that prestructures memory into diverse granularities and employs a reward-guided navigator to adaptively compose hybrid memory tailored to each query. |
| Outcome: | Experiments on long-context multi-hop question answering show that the framework achieves state-of-the-art performance. |
Unlocking Large Audio-Language Models for Interactive Language Learning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Computer-Assisted Pronunciation Training (CAPT) systems provide unintuitive feedback that lacks actionable guidance. |
| Approach: | They propose to use audio-language models to provide more user-friendly feedback for pronunciation training. |
| Outcome: | The proposed model outperforms baselines on mispronunciation detection and suggestion generation. |
Blind Spot Navigation in Large Language Model Reasoning with Thought Space Explorer (2026.findings-eacl)
Copied to clipboard
Jinghan Zhang, Fengran Mo, Tharindu Cyril Weerasooriya, Xinyue Ye, Dongjie Wang, Yanjie Fu, Kunpeng Liu
| Challenge: | Existing studies show that large language models have strong reasoning capabilities through chain-structured methods. |
| Approach: | They propose a framework for navigating and expanding thought structures to overcome blind spots in LLM reasoning. |
| Outcome: | The proposed framework overcomes blind spots in large language models by expanding thought structures . the proposed framework improves accuracy of the final answer and intermediate reasoning steps . |
StrucSum: Graph-Structured Reasoning for Long Document Extractive Summarization with LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have shown strong performance in zero-shot summarization, but struggle to model document structure and identify salient information in long texts. |
| Approach: | They propose a training-free prompting framework that injects structural signals into prompts via sentence-level graph structures. |
| Outcome: | The proposed framework improves summary quality and factual consistency over baselines and vanilla prompting. |
Logits-Based Block Pruning with Affine Transformations for Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for pruning models rely on calibration data and neglect cumulative effects of pruning on subsequent blocks. |
| Approach: | They propose to use the Logit Disruption Score (LDS) to measure the impact of pruning by comparing the cosine similarity between the logits of the original and pruned models. |
| Outcome: | Experiments on multiple datasets show that the proposed pruning technique reduces reliance on calibration data and improves generalization, achieving competitive results with existing methods. |
MultiCW: A Large-Scale Balanced Benchmark Dataset for Training Robust Check-Worthiness Detection Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are beginning to reshape how media professionals verify information, but support for detecting check-worthy claims remains limited. |
| Approach: | They propose a multilingual benchmark for check-worthy claim detection spanning 16 languages, six topical domains, and two writing styles. |
| Outcome: | The proposed model outperforms zero-shot LLMs on claim classification and strong generalization across languages, domains, and styles. |
What Really Matters for Table LLMs? A Meta-Evaluation of Model and Data Effects (2026.findings-eacl)
Copied to clipboard
Naihao Deng, Sheng Zhang, Henghui Zhu, Shuaichen Chang, Jiani Zhang, Alexander Hanbo Li, Chung-Wei Hang, Hideo Kobayashi, Yiqun Hu, Patrick Ng
| Challenge: | a series of paradigm shifts have come with distinct characteristics and challenges associated with table modeling. |
| Approach: | They propose to replicate four table LLMs by instruction-tuning three foundation models on four existing datasets. |
| Outcome: | The results show that base model choice plays a more dominant role than training data itself. |
Evaluating Morphological Plausibility of Subword Tokenization via Statistical Alignment with Morpho-Syntactic Features (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing metric for subword tokenization evaluation for morphological plausibility requires unavailable or inconsistent gold segmentation data. |
| Approach: | They propose a morpho-syntactic feature-based metric for subword tokenization evaluation. |
| Outcome: | The proposed metric correlates well with traditional morpheme boundary recall while being more broadly applicable across languages with different morphological systems. |
MOSAIC: Masked Objective with Selective Adaptation for In-domain Contrastive Learning (2026.findings-eacl)
Copied to clipboard
| Challenge: | a new framework for domain adaptation of text embedding models addresses the challenges of adapting general-domain text embeds to specialized domains. |
| Approach: | They propose a framework for domain adaptation of text embedding models that integrates masked supervision and mangled objectives within a unified training pipeline. |
| Outcome: | The proposed framework improves on high-resource and low-resourced domains while preserving the robustness of the original model. |
BitBypass: A New Direction in Jailbreaking Aligned Large Language Models with Bitstream Camouflage (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have inherent risk of generating harmful and unsafe content. |
| Approach: | They develop a black-box jailbreak attack that leverages hyphen-separated bitstream camouflage to bypass aligned Large Language Models' safety alignment. |
| Outcome: | The proposed attack outperforms state-of-the-art jailbreak attacks in stealthiness and attack success. |
The Problem of Ambiguity in Table Question Answering (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to question answering on tabular data have limited capabilities due to ambiguousness inherent to tabular datasets. |
| Approach: | They propose to use large language models to answer questions on tabular data by analyzing tabular tables and detecting ambiguity. |
| Outcome: | The proposed model can detect ambiguity in tabular data and provide an initial ground for a deeper discussion on how to approach it in the age of LLMs. |
Beyond Multiple Choice: Evaluating Steering Vectors for Summarization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent methods for controlling language models can often be classified into three main strategies: prompt engineering, trainable decoding mechanisms, fine-tuning according to specific objectives. |
| Approach: | They evaluate steering vectors for controlling topical focus, sentiment, toxicity, and readability in abstractive summaries across the SAMSum, NEWTS, and arXiv datasets. |
| Outcome: | The proposed method is effective in free-form generation, but high steering strengths induce degenerate repetition and factual hallucinations. |
Similar Region Search using LLMs on Spatial Feature Space (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing similarity search methods fail to capture contextual richness of spatial data . existing methods fail in capturing regional characteristics, authors say . |
| Approach: | They propose a similar region search framework that ranks candidate regions based on their similarity to a query region using large language models. |
| Outcome: | The proposed similar region search framework outperforms state-of-the-art methods on real-world city datasets. |
Learning to Ask: Multi-Decoder Fine-Tuning for Multi-Hop Visual Question Generation with External Knowledge (2026.findings-eacl)
Copied to clipboard
| Challenge: | Traditional supervised QG methods rely on tokenlevel alignment with fixed gold labels struggle to capture diverse valid question formulations. |
| Approach: | They propose a model-agnostic framework that integrates multimodal inputs with a multi-decoder architecture to optimize for multiple labels per sample. |
| Outcome: | The proposed framework improves fluency, reasoning depth, and relevance of visual questions. |
SLANG-GraphRAG: Multi-Layered Retrieval with Domain-Specific Knowledge for Low Resource Social Media Conversations (2026.findings-eacl)
Copied to clipboard
| Challenge: | Standard NLP benchmarks often miss subtle, culturally-specific cues in social media . incorporating structured cultural knowledge into the retrieval process improves accuracy by up to 31% . |
| Approach: | They propose a retrieval-augmented framework that integrates a culture-specific slang knowledge graph into large language models via one-shot prompting. |
| Outcome: | The proposed framework outperforms traditional and unstructured retrieval methods in slang-based models by 31% and 28%. |
Science Across Languages: Assessing LLM Multilingual Translation of Scientific Papers (2026.findings-eacl)
Copied to clipboard
| Challenge: | a large number of scientific journals are published exclusively in English . this creates barriers for non-native English speakers to access scientific knowledge . |
| Approach: | They propose a way to translate scientific articles while preserving native JATS XML formatting. |
| Outcome: | The proposed approach shows that the key scientific details are accurately conveyed. |
TABED: Test-Time Adaptive Ensemble Drafting for Robust Speculative Decoding in LVLMs (2026.findings-eacl)
Copied to clipboard
Minjae Lee, Wonjun Kang, Byeongkeun Ahn, Christian Classen, Kevin Galim, Seunghyuk Oh, Minghao Yan, Hyung Il Koo, Kangwook Lee
| Challenge: | Large Vision Language Models (LVLMs) are advanced models that process multiple modalities, such as images, audio, and video, alongside text. |
| Approach: | They propose to use a method to generate and verify draft tokens in parallel . they compare existing methods with small draft models and observe performance fluctuations . |
| Outcome: | The proposed method achieves an average walltime speedup of 1.74 over autoregressive decoding and a 5% improvement over single drafting methods. |
KGHaluBench: A Knowledge Graph-Based Hallucination Benchmark for Evaluating the Breadth and Depth of LLM Knowledge (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing benchmarks for large language models are limited by static and narrow questions, leading to limited coverage and misleading evaluations. |
| Approach: | They propose a Knowledge Graph-based hallucination benchmark that assesses Large Language Models across the breadth and depth of their knowledge and provides a fairer and more comprehensive insight into LLM truthfulness. |
| Outcome: | The proposed framework assesses LLMs across breadth and depth of their knowledge, and provides a fairer and more comprehensive insight into LLM truthfulness. |
Marking Code Without Breaking It: Code Watermarking for Detecting LLM-Generated Code (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for identifying LLM-generated code are limited by syntax-critical tokens, which can introduce syntax errors. |
| Approach: | They propose a syntax-aware watermarking method that embeds watermarks only in non-syntactic tokens and preserves code integrity. |
| Outcome: | The proposed method outperforms baseline methods on Python, C++, and Java. |
VIGiA: Instructional Video Guidance via Dialogue Reasoning and Retrieval (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing models focus on text-only guidance or treat vision and language in isolation. |
| Approach: | They propose a multimodal dialogue model that supports grounded, plan-aware dialogue . they use a dataset with rich video dialogues aligned with cooking and DIY plans . |
| Outcome: | The proposed model outperforms existing models on all tasks in a conversational plan guidance setting, reaching over 90% accuracy on plan-aware VQA. |
Attribute-Controlled Translation with Preference Optimization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Attribute-controlled translation (ACT) is a natural language processing task that produces translations that satisfy specific constraints on linguistic and stylistic attributes. |
| Approach: | They propose to leverage the contrastive nature of ACT tasks with preference optimization . they also propose to exploit knowledge distillation with synthetically-generated training samples . |
| Outcome: | The proposed approach improves attribute matching and translation quality in small-medium size models. |
ReciFine: Finely Annotated Recipe Dataset for Controllable Recipe Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing resources, such as RecipeNLG, extract food items only from ingredient lists, overlooking entities expressed in instructions, such tools, chef actions, food and tool states, and durations. |
| Approach: | They extend RecipeNLG to extract 97 million entities from 2.2 million recipes. |
| Outcome: | The proposed model outperforms existing models trained on ingredient-list data on both automatic and human evaluations. |
ReBPE: Iteratively Improving the Internal Structure of a Structured Tokeniser by Mining its Internal Structure (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent work explores pruning merges from BPE subword tokenisers using corpus data as a signal for which merges to prune. |
| Approach: | They propose a pruning algorithm that inspects the effects left by pruning . they propose reification of the tokenisers and a new pruning algorithm . |
| Outcome: | The proposed algorithm outperforms the original BPE-knockout algorithm on alignment in all 14 languages tested by over 11% F1 on average. |
Emotion Recognition in Multi-Speaker Conversations through Speaker Identification, Knowledge Distillation, and Hierarchical Fusion (2026.findings-eacl)
Copied to clipboard
| Challenge: | Emotion recognition in multi-speaker conversations faces significant challenges due to speaker ambiguity and severe class imbalance. |
| Approach: | They propose a speaker identification module that leverages audio-visual synchronization to accurately identify the active speaker and hierarchical attention fusion with composite loss functions to handle class imbalance. |
| Outcome: | The proposed framework achieves 67.75% and 72.44% weighted F1 scores on MELD and IEMOCAP datasets, with notable improvements on minority emotion classes. |
Demystifying Mixed Outcomes of Self-Training: Pre-training Analyses on Non-Toy LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent studies on self-training report seemingly contradictory outcomes. |
| Approach: | They use OLMo-2 models as non-toy LLMs and perform multiple rounds of continual pre-training using self-generated text with different prompting strategies and data filtering. |
| Outcome: | The proposed model collapse is inherent to the training procedure itself, while self-improvement is likely owes its success to human-designed, strategic synthetic pipelines that inject external intelligence. |
Revealing Redundant Syntax in Large Language Models through Multi-Hop Dependency Paths (2026.findings-eacl)
Copied to clipboard
| Challenge: | Prior work on attention–syntax alignment has focused on single-hop Universal Dependency edges (DPs). |
| Approach: | They extract 2–3 hop MDPs from UD-parsed English and quantify head–relation alignment with an Unlabeled Attachment Score (UAS)-style metric modified for causal masking in decoder-only models. |
| Outcome: | The authors show that head alignments are overlapped and specialized . the head alignment is measurable in large language models trained on raw text . |
A Scalable Framework for Automated NER Annotation Correction in Low-Resource Languages (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing NER benchmarks lack quality annotations, resulting in poor performance. |
| Approach: | They propose a frequency-based iterative approach that leverages self-training and a dual-threshold mechanism to enhance inference confidence. |
| Outcome: | The proposed approach improves NER performance on three datasets with a high number of missing annotations. |
Can ChatGPT Really Understand Modern Chinese Poetry? (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent studies have focused on poetry generation and translation, but their scope has been limited to evaluation and analysis of experimental results without addressing fundamental issues of comprehension. |
| Approach: | They propose a framework for evaluating ChatGPT's understanding of modern poetry . they evaluated the interpretations of unpublished modern Chinese poems by different poets . |
| Outcome: | The proposed framework is based on the evaluation of unpublished poems by poets and shows that its interpretations align with the original poets’ intents in over 73% of the cases. |
Knowing What’s Missing: Assessing Information Sufficiency in Question Answering (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to assess whether a given context contains sufficient information fail on factual questions. |
| Approach: | They propose a framework that asks a model to reason about what information is missing . this framework generates more accurate sufficiency judgments while articulating any information gaps . |
| Outcome: | The proposed framework produces more accurate sufficiency judgments while clearly articulating any information gaps. |
The Curse of Verbalization: How Presentation Order Constrains LLM Reasoning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Xu et al., 2025) found that LLMs struggle when programs execute in an unaligned order. |
| Approach: | They propose to use esoteric programming languages to evaluate LLMs' reasoning abilities. |
| Outcome: | The proposed model improves reasoning performance across state-of-the-art models by restructuring problems to align the presentation order with the order of utilization. |
PATS: Personality-Aware Teaching Strategies with Large Language Model Tutors (2026.findings-eacl)
Copied to clipboard
Donya Rooein, Sankalan Pal Chowdhury, Mariia Eremeeva, Yuan Qin, Debora Nozza, Mrinmaya Sachan, Dirk Hovy
| Challenge: | pedagogical theories are not aligned with teaching strategies for educational tasks . quiet students may be disengaged or not thinking critically because they do not speak up . |
| Approach: | They propose a taxonomy that links pedagogical methods to personality profiles to map teaching strategies to student personality traits. |
| Outcome: | The proposed model improves the use of less common, high-impact strategies such as role-playing . the model also increases the use less common strategies such role-players . |
Mitigating Causal Bias in LLMs via Potential Outcomes Framework and Actual Causality Theory (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models exhibit significant causal hallucination, but evaluation of their document-level ECI performance is lacking. |
| Approach: | They propose to use Large Language Models to evaluate their document-level ECI performance . they propose a framework to mitigate the causal bias associated with using LLMs . |
| Outcome: | The proposed framework significantly reduces the causal bias associated with using LLMs on ECI while also achieving superior performance. |
JuriFindIT: an Italian legal retrieval dataset (2026.findings-eacl)
Copied to clipboard
| Challenge: | Statutory article retrieval (SAR) targets retrieval of legislative provisions relevant to a natural language question. |
| Approach: | They propose a pipeline that integrates dense encoders with an heterogeneous legislative graph . they propose statutory article retrieval (SAR) is the first SAR dataset for the italian legal domain . |
| Outcome: | The proposed pipeline improves over existing approaches. |
Towards Fair and Efficient De-identification: Quantifying the Efficiency and Generalizability of De-identification Approaches (2026.findings-eacl)
Copied to clipboard
| Challenge: | a recent study has not examined their generalizability between formats, cultures, and genders. |
| Approach: | They evaluate large language models (LLMs) and small LLMs at clinical de-identification . they show that smaller models achieve comparable performance while substantially reducing inference cost . |
| Outcome: | The proposed models outperform larger models in de-identification tasks with limited data . the models can be fine-tuned with limited datasets to outperformed larger models . |
How Many Ratings per Item are Necessary for Reliable Significance Testing? (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for estimating model reliability are based on a few output responses per item. |
| Approach: | They propose a method to determine whether an existing dataset has enough responses per item to assure reliable null hypothesis statistical testing. |
| Outcome: | The proposed method can help researchers make better decisions about how to collect data for AI evaluation. |
QFrBLiMP: a Quebec-French Benchmark of Linguistic Minimal Pairs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Specifically, these minimal pairs are created by manually modifying sentences extracted from an official online resource maintained by a Québec government institution. |
| Approach: | They propose to use the Quebec-French Benchmark of Linguistic Minimal Pairs to evaluate LLMs’ linguistic knowledge of prominent grammatical phenomena in Quebec-french. |
| Outcome: | The proposed corpus evaluates LLMs’ linguistic knowledge of prominent grammatical phenomena in Quebec-French. |
QueerGen: How LLMs Reflect Societal Norms on Gender and Sexuality in Sentence Completion Task (2026.findings-eacl)
Copied to clipboard
| Challenge: | Autoregressive Language Models (ARLMs) partially mitigate these patterns, while closed-access ARLMs tend to produce more harmful outputs for unmarked subjects. |
| Approach: | They examine whether explicit information about a subject’s gender or sexuality influences LLM responses across three subject categories: queer-marked, non-queer-mark, and the normalized "unmarked" category. |
| Outcome: | The proposed models reproduce normative social assumptions, but the form and degree of bias depend on model characteristics, which may redistribute—but not eliminate—representational harms. |
Efficient Table Retrieval and Understanding with Multimodal Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Tabular data is often captured in image form across a wide range of real-world scenarios. |
| Approach: | They propose a framework that enables MLLMs to answer queries over large tables. |
| Outcome: | The proposed framework outperforms existing methods by 7.0% in retrieval recall and 6.1% in answer accuracy on a newly constructed dataset with 48,504 unique tables. |
FedReFT: Federated Representation Fine-Tuning with All-But-Me Aggregation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Representation Fine-Tuning (ReFT) adapts large pre-trained models by updating only a small subset of parameters. |
| Approach: | They propose a method that uses sparse intervention layers to steer hidden representations directly to capture rich semantic information. |
| Outcome: | The proposed approach outperforms PEFTs on commonsense reasoning, arithmetic reasoning, and GLUE benchmarks while maintaining a high parameter efficiency. |
RiddleBench: A New Generative Reasoning Benchmark for LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) show remarkable capabilities, but complex reasoning skills require deeper investigation. |
| Approach: | They propose a benchmark of 1,737 puzzles to test reasoning beyond simple pattern matching. |
| Outcome: | The proposed model performs poorly when faced with reordered constraints or irrelevant information. |
Language Model-Driven Data Pruning Enables Efficient Active Learning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing data pruning methods for active learning are expensive and time-consuming. |
| Approach: | They propose a plug-and-play data pruning strategy that leverages language models to prune the unlabeled pool. |
| Outcome: | The proposed pruning strategy outperforms existing pruning methods on translation, sentiment analysis, topic classification, and summarization tasks on diverse datasets. |
HARM: Learning Hate-Aware Reward Model for Evaluating Natural Language Explanations of Offensive Content (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing reward models for explaining hate speech are optimized for broad notions of safety, but they assign lower scores to contextually rich explanations. |
| Approach: | They propose a reward model that integrates interpretable signals to better align reward scores with the needs of hate speech explanation. |
| Outcome: | The proposed model outperforms general-purpose baselines and improves pair-wise preference. |
MATH-IDN: A Multilingual Mathematical Problem Solving Dataset Featuring Local Languages in Indonesia (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models excel at mathematical reasoning in English, but their performance in low-resource languages remains underexplored. |
| Approach: | They propose a multilingual benchmark for mathematical problem solving in Indonesian, Javanese, Sundanese, and Buginese with English as a reference. |
| Outcome: | The proposed model reveals significant performance gaps in low-resource languages, particularly Buginese, and highlights key limitations in current multilingual reasoning capabilities. |
Parameter-Efficient Routed Fine-Tuning: Mixture-of-Experts Demands Mixture of Adaptation Modules (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing Parameter-Efficient Fine-Tuning (PEFT) strategies that focus on specialized experts are not effective for Mixture-of-Experts (MoE). |
| Approach: | They propose to integrate a dynamic routing mechanism among specialized experts in Mixture-of-Experts (MoE) . |
| Outcome: | Extensive experiments on commonsense and math reasoning tasks validate the performance and efficiency of the proposed routed approach. |
MAPRO: Recasting Multi-Agent Prompt Optimization as Maximum a Posteriori Inference (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks. |
| Approach: | They propose a framework that optimizes MAS prompts as a maximum a posteriori problem and then iteratively updates agent prompts. |
| Outcome: | The proposed framework surpasses manual and automated benchmarks in multiple tasks and provides general guidelines for building more reliable and principled multi-agent systems in the future. |
Debiasing Large Language Models via Adaptive Causal Prompting with Sketch-of-Thought (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing prompting methods for Large Language Models (LLMs) suffer from excessive token usage and limited generalisability across diverse reasoning tasks. |
| Approach: | They propose an Adaptive Causal Prompting with Sketch-of-Thought framework that leverages structural causal models to infer the causal effect of a query on its answer. |
| Outcome: | The proposed framework outperforms existing prompting baselines in terms of accuracy, robustness, and computational efficiency. |
ExpressivityBench: Can LLMs Communicate Implicitly? (2026.findings-eacl)
Copied to clipboard
Joshua Tint, Som Sagar, Aditya Taparia, Kelly Raines, Bimsara Pathiraja, Caleb Liu, Ransalu Senanayake
| Challenge: | a new study evaluates the expressivity of large language models for communicating implicitly . authors: models can express tone, identity, and intent beyond literal meanings . phrasing and tone of a message can convey a number of topics beyond literal contexts - authors . |
| Approach: | They propose a framework to evaluate the expressivity of large language models . they use a social-linguistic grader to validate their models against human judgments . |
| Outcome: | The proposed framework quantifies how well LLM-generated text communicates target properties without explicit mention across nine tasks spanning emotion, identity, and tone. |
Rethinking Schema Linking: A Context-Aware Bidirectional Retrieval Approach for Text-to-SQL (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent methods focus on improving SQL generation but neglect retrieval of relevant schema elements. |
| Approach: | They propose a context-aware bidirectional schema retrieval framework that treats schema linking as a standalone problem. |
| Outcome: | The proposed framework improves schema recall while reducing false positives. |
PEAR: Planner-Executor Agent Robustness Benchmark (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies examine isolated attack surfaces or specific scenarios, leaving a lack of holistic understanding of MAS vulnerabilities. |
| Approach: | They propose a benchmark to evaluate the utility and vulnerability of planner–executor MAS. |
| Outcome: | The proposed benchmark evaluates planner–executor MAS on a widely adopted design. |
Toward Safe and Human-Aligned Game Conversational Recommendation via Multi-Agent Decomposition (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing systems for conversational recommender systems (CRS) have strong results in movies, but games present distinct challenges . MATCHA framework provides specialized agents for intent parsing, tool-augmented retrieval, multi-LLM ranking, and stronger safety. |
| Approach: | They propose a framework for conversational recommender systems that assigns specialized agents for intent parsing, tool-augmented retrieval, multi-LLM ranking and risk control. |
| Outcome: | MATCHA outperforms baselines on real user request dataset, improves Hit@5 by 20%, reduces popularity bias by 24%, and achieves 97.9% adversarial defense. |
Linguistic Cues for LLM-based Implicit Discourse Relation Classification (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have been successful in many NLP tasks, but they struggle to capture subtle lexical relations between arguments. |
| Approach: | They propose a strategy that enriches arguments with explicit lexical-level semantic cues before fine-tuning. |
| Outcome: | The proposed approach improves F1 scores in cross-domain scenarios by more than 10 points compared to baselines. |
SpARK: An Embarrassingly Simple Sparse Watermarking in LLMs with Enhanced Text Quality (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for detecting and monitoring generated text face a trade-off between the quality of the generated text and the effectiveness of the watermarking process. |
| Approach: | They propose a new type of LLM watermark, Sparse WatermARK, which uses watermarks to a small subset of generated tokens distributed across the text. |
| Outcome: | The proposed method outperforms existing methods in detectability and quality while maintaining generated text quality. |
Pretraining Language Models for Diachronic Linguistic Change Discovery (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models are increasingly used as knowledge discovery tools . historical linguistics and literary studies often construct arguments on the basis of distinctions between phenomena like time-period or genre. |
| Approach: | They propose to use LLMs to train large language models over modest historical corpora without allowing contamination from anachronistic data. |
| Outcome: | The proposed model better respects historical divisions and is more computationally efficient compared to the standard approach of fine-tuning an existing LLM. |
Improving Language Identification for Code-Switched Speech: The Pivotal Role of Accented English (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing models fail to identify English spoken with the accent of the matrix (dominant) language. |
| Approach: | They propose to fine tune existing LID models with accented English to improve code-switched LID . they use a metric that captures relative ranking of identified languages often overlooked by traditional metrics. |
| Outcome: | The proposed model can be fine tuned with small amounts of accented English without degrading performance on monolingual speech. |
Reducing Hallucinations in Language Model-based SPARQL Query Generation Using Post-Generation Memory Retrieval (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are susceptible to hallucinations and out-of-distribution errors when generating KG elements, such as Uniform Resource Identifiers (URIs). |
| Approach: | They propose a SPARQL query-generating framework that uses natural language placeholders and a non-parametric memory module to retrieve and resolve the correct KG URIs. |
| Outcome: | The proposed framework significantly enhances query correctness across various LLMs, datasets, and distribution shifts while achieving the near-complete suppression of URI hallucinations. |
Jailbreaking Safeguarded Text-to-Image Models via Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Text-to-image models generate harmful content when unsafe prompts are submitted . authors propose a method to jailbreak text-to image models with safety guardrails . |
| Approach: | They propose a method to jailbreak text-to-image models with safety guardrails . they use a fine-tuned large language model to generate adversarial prompts based on unsafe prompts. |
| Outcome: | The proposed method bypasses safety guardrails and outperforms existing no-box attacks . the proposed method generates adversarial prompts efficiently after fine-tuning the model . |
BSCodec: A Band-Split Neural Codec for High-Quality Universal Audio Reconstruction (2026.findings-eacl)
Copied to clipboard
| Challenge: | Neural audio codecs have enabled high-fidelity reconstruction of speech, music and sound . however, speech-optimized codec systems suffer degradation on music or sound if they ignore spectral differences . |
| Approach: | They propose a neural audio codec that splits the spectral dimension into separate bands and compresses each band independently. |
| Outcome: | Experimental results show that BSCodec achieves better reconstruction quality on music and sound compared to existing codecs. |
Crafting Adversarial Inputs for Large Vision-Language Models Using Black-Box Optimization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing white-box jailbreak methods require full model accessibility and require computational costs. |
| Approach: | They propose a black-box jailbreak attack using Zeroth-Order optimization using ZO-SPSA. |
| Outcome: | The proposed method achieves highest jailbreak success rate on three LVLMs, including InstructBLIP, LLaVA and MiniGPT-4. |
SALT: Step-level Advantage Assignment for Long-horizon Agents via Trajectory Graph (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable capabilities, but their application to complex, multi-step, and long-horizon tasks remains challenging. |
| Approach: | They propose a framework that provides a finer-grained advantage assignment derived solely from outcome rewards. |
| Outcome: | The proposed framework provides a finer-grained advantage assignment, derived solely from outcome rewards. |
UniToolBench: A Benchmark for Tool-Augmented LLMs in Cross-Domain, Universal Task Automation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing benchmarks that focus on manually curated tool graphs lack scalability and diversity across domains. |
| Approach: | They propose a large-scale, cross-domain benchmark to evaluate LLMs' ability to reason over and utilize interconnected tools for automation. |
| Outcome: | The proposed benchmark incorporates automated tool graph construction by formulating link prediction as a probabilistic task, instead of relying on categorical LLM outputs. |
Benchmarking the Energy Savings with Speculative Decoding Strategies (2026.findings-eacl)
Copied to clipboard
Rohit Dutta, Paramita Koley, Soham Poddar, Janardan Misra, Sanjay Podder, Naveen Balani, Saptarshi Ghosh, Niloy Ganguly
| Challenge: | Existing studies on speculative decoding have focused on the energy requirements of these models, despite their utility and utility. |
| Approach: | They propose to analyze the energy requirements of speculative decoding strategies and analyze how various factors influence the energy optimizations. |
| Outcome: | The proposed approach reduces decoding time while offloading a substantial portion of the sequential generation to a smaller, more efficient model. |
Thunder-NUBench: A Benchmark for LLMs’ Sentence-Level Negation Understanding (2026.findings-eacl)
Copied to clipboard
| Challenge: | Negation is a fundamental linguistic phenomenon that poses ongoing challenges for Large Language Models (LLMs) Current benchmarks treat negation as a minor detail within broader tasks, such as natural language inference. |
| Approach: | They propose a novel benchmark specifically created to assess sentence-level understanding of negation in Large Language Models (LLMs). |
| Outcome: | The proposed benchmark compares standard negation with structurally diverse alternatives, such as local negation, contradiction, and paraphrase. |
What Makes a Good Query? Measuring the Impact of Human-Confusing Linguistic Features on LLM Performance (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are often treated as defects of the model or its decoding strategy. |
| Approach: | They construct a 22-dimension query feature vector covering clause complexity, lexical rarity, anaphora, negation, answerability, and intention grounding. |
| Outcome: | The proposed model covers clause complexity, lexical rarity, anaphora, negation, answerability, and intention grounding, all known to affect human comprehension. |
Completely Modular Fine-tuning for Dynamic Language Adaptation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies on multilingual fine-tuning with a fixed set of languages lack dynamic adaptability to new languages. |
| Approach: | They propose a modular fine-tuning pipeline that enables dynamic language adaptation for LLMs by first training English-centric adapters for each language separately and then merging them for arbitrary-direction translation. |
| Outcome: | The proposed pipeline achieves 86% performance over traditional fine-tuning on four languages, while training only 0.1% parameters and relying on English as a bridge language without catastrophic forgetting. |
A Multi-Task Learning Framework for Modeling Engagement and Topic-Sensitive Responses in Arabic Women’s Discourse (2026.findings-eacl)
Copied to clipboard
| Challenge: | a corpus of 158k arab Facebook posts spanning women's rights, gender debates, and economic empowerment reveals patterns of public opinion that vary dramatically across regional and cultural contexts. |
| Approach: | They propose a multi-task learning framework that learns audience reaction classification and engagement magnitude regression and non-engagement detection. |
| Outcome: | The proposed model achieves a test macro-F1 of 72.4 and weighted-F1. It measures 158k posts across gender issues, legal rights advocacy, gender identity discussions, and economic empowerment. |
We Are What We Repeatedly Do: Improving Long Context Instruction Following (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language model context lengths have increased by at least 1000 in the past seven years . however, longer contexts pose challenges to system instruction following . |
| Approach: | They propose to formalize verifiable instructions to evaluate model compliance . they implement and evaluate six mitigation strategies to enhance instruction compliance in extended contexts. |
| Outcome: | The proposed model performs better in long contexts than in natural language models. |
ConRAS: Contrastive In-context Learning Framework for Retrieval-Augmented Summarization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Despite progress in natural language processing, the potential of contrastive learning remains unexplored. |
| Approach: | They propose a framework that injects contrastive objectives into in-context learning-based retrieval-augmented summarization. |
| Outcome: | The proposed framework outperforms state-of-the-art retrieval-augmented methods on three summarization benchmarks showing that it can distinguish between positive and negative samples without parameter updates. |
Beyond Sampling: Self-Sorting for Long-Context Ranking (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) remain unstable on long-context ranking. |
| Approach: | They propose a method that fuses explicit within-list positions with implicit cross-list preferences to score entities and return a top-k set. |
| Outcome: | Experimental results show that large language models remain unstable on long-context ranking . |
Program-of-Thought Reveals LLM Abstraction Ceilings (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models exhibit reasoning ability when supervised with chain-of-thought (CoT) traces. |
| Approach: | They evaluate large language models with CoT traces and fine-tune them with Program-of-Thought supervision. |
| Outcome: | The proposed model performance degrades sharply under numeric perturbations under isomorphic variants. |
From Numbers to Narratives: Efficient Language Model-Based Detection for Safety-Critical Minority Classes (2026.findings-eacl)
Copied to clipboard
| Challenge: | Tabular datasets with high overall accuracy and poor performance on minority classes are often misaligned . a numbers to narratives framework improves overall accuracy by up to 22.43% in five of six datasets while maintaining computational feasibility. |
| Approach: | They propose a number-to-narrative framework that transforms tabular data into contextually rich descriptions. |
| Outcome: | The proposed framework achieves superior minority class F1-scores in five of six datasets. |
R-GDA: Reflective Guidance Data Augmentation with Multi-Agent Feedback for Domain-Specific Named Entity Recognition (2026.findings-eacl)
Copied to clipboard
| Challenge: | Named Entity Recognition (NER) tasks require data augmentation due to the scarcity of annotated corpora. |
| Approach: | They propose a framework that introduces a multi-agent feedback loop to enhance augmentation quality. |
| Outcome: | The proposed framework improves on SciERC and NCBI-disease datasets and achieves low BERTScore in most cases. |
Enabling Autoregressive Models to Fill In Masked Tokens (2026.findings-eacl)
Copied to clipboard
| Challenge: | Autoregressive (AR) and masked language modeling (MLM) models are incapable of mucked infilling, which is the ability to predict mangled tokens between past and future context. |
| Approach: | They propose a method that leverages the strengths of autoregressive and masked language modeling to achieve state-of-the-art mucked infilling performance. |
| Outcome: | The proposed approach outperforms existing methods on masked infilling tasks. |
Position Encoding with Random Float Sampling Enhances Length Generalization of Transformers (2026.findings-eacl)
Copied to clipboard
| Challenge: | Length generalization is the ability of language models to maintain performance on inputs longer than those seen during pretraining. |
| Approach: | They propose a position encoding strategy that uses random float sampling to generalize to unseen lengths. |
| Outcome: | The proposed strategy can generalize to lengths unseen during training and in benchmarks. |
Open-Domain Safety Policy Construction (2026.findings-eacl)
Copied to clipboard
| Challenge: | Moderation layers are core component of many products built on user-generated content. |
| Approach: | They propose a system that drafts a content moderation policy based on human-written seed domain information. |
| Outcome: | The proposed system outperforms definition-only and in-context learning baselines on openAI undesired content benchmarks and an in-house multimodal advertisement moderation benchmark. |
Think Just Enough: Leveraging Self-Assessed Confidence for Adaptive Reasoning in Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent advances in large reasoning models (LLMs) have shown remarkable capabilities in complex tasks such as mathematical problem solving and code generation. |
| Approach: | They propose a method for optimizing reasoning length via self-assessed confidence. |
| Outcome: | The proposed method improves computational efficiency without compromising answer quality. |
CLICKER: Cross-Lingual Knowledge Editing via In-Context Learning with Adaptive Stepwise Reasoning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing knowledge editing methods are static and fail to propagate edits across languages. |
| Approach: | They propose a KE method that dynamically retrieves only knowledge relevant to a given query and edits it to maintain cross-lingual consistency. |
| Outcome: | The proposed method outperforms static KE methods on a multilingual dataset with semantically similar but irrelevant prompts. |
Show or Tell? Modeling the evolution of request-making in Human-LLM conversations (2026.findings-eacl)
Copied to clipboard
| Challenge: | a new framework to describe request-making segments user input into request content, roles assigned, query-specific context, and task-independent expressions. |
| Approach: | They propose a framework to describe request-making that segments user input into request content, roles assigned, query-specific context, and the remaining task-independent expressions. |
| Outcome: | The proposed framework reveals fundamental and habitual user-LLM interaction patterns beyond individual task completion. |
Multilingual Self-Taught Faithfulness Evaluators (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing faithfulness evaluation approaches are mostly English-focused and require expensive human-labeled training data for fine-tuning specialized models. |
| Approach: | They propose a framework that learns exclusively from synthetic multilingual data while leveraging cross-lingual transfer learning to improve an LLM's general language capabilities. |
| Outcome: | The proposed framework shows that it improves over existing baselines, including state-of-the-art English evaluators and machine translation-based approaches. |
Benchmarking Direct Preference Optimization for Medical Large Vision–Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large vision-language models (LVLMs) are gaining traction in clinical tasks such as diagnostic support, report generation, and medical question answering. |
| Approach: | They present a systematic evaluation of nine DPO variants applied to two leading medical LVLMs. |
| Outcome: | The proposed model improves alignment and reduces severe hallucinations, but yields inconsistent gains over supervised fine-tuning. |
Stay Focused: Problem Drift in Multi-Agent Debate (2026.findings-eacl)
Copied to clipboard
| Challenge: | Multi-agent debates have shown promise for solving knowledge and reasoning tasks, but they are limited when solving complex problems that require longer reasoning chains. |
| Approach: | They propose a method to detect problem drift and propose 'driFTJudge' which mitigates 31% of problem drift cases. |
| Outcome: | The proposed method mitigates 31% of problem drift cases and is based on a set of ten tasks across ten different tasks. |
FLUKE: A Linguistically-Driven and Task-Agnostic Framework for Robustness Evaluation (2026.findings-eacl)
Copied to clipboard
Yulia Otmakhova, Thinh Hung Truong, Rahmad Mahendra, Zenan Zhai, Rongxin Zhu, Daniel Beck, Jey Han Lau
| Challenge: | FLUKE introduces controlled variations across linguistic levels and leverages large language models with human validation to generate modifications. |
| Approach: | They propose a framework for assessing model robustness through systematic minimal variations of test data. |
| Outcome: | The proposed framework evaluates models and LLMs across six diverse NLP tasks and shows that they are more robust to natural, fluent modifications than base models. |
Sparse Brains are Also Adaptive Brains: Cognitive-Load-Aware Dynamic Activation for LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing sparsity methods lack adaptivity to contextual or model structural demands or incur prohibitive computational overhead. |
| Approach: | They propose a Cognitive-Load-Aware Dynamic Activation framework that synergizes statistical sparsity with semantic adaptability. |
| Outcome: | The proposed framework achieves 20% average speedup with less than 2% accuracy degradation outperforming Griffin and TT. |
PATCH: Mitigating PII Leakage in Language Models with Privacy-Aware Targeted Circuit PatcHing (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing defense mechanisms to mitigate PII leakage are limited by existing defenses . a new approach, PATCH, identifies and edits PI I circuits to reduce leakage . |
| Approach: | They propose to use PATCH: Privacy-Aware Targeted Circuit Patching to identify PII leakage circuits in language models to reduce leakage. |
| Outcome: | The proposed approach reduces leakage by up to 65% and can reduce residual leakage to as low as 0.01%. |
Argument Component Segmentation with Fine-Tuned Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Argument Mining (AM) aims to identify and interpret argumentative structures in unstructured text. |
| Approach: | They propose a fine-grained, paired-tag annotation schema that distinguishes between relevant and surrounding content. |
| Outcome: | The proposed approach performs comparable to human expert annotators across multiple benchmark datasets. |
DuFFin: A Dual-Level Fingerprinting Framework for LLMs IP Protection (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models are valuable intellectual property due to the computational cost of training. |
| Approach: | They propose a dual-level fingerprinting framework that extracts trigger patterns and knowledge-level signatures to verify black-box ownership. |
| Outcome: | The proposed framework verifies the copyright of protected LLMs on their variants, achieving an IP-ROC greater than 0.99. |
The Art of Saying "Maybe": A Conformal Lens for Uncertainty Benchmarking in VLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent advances in large vision-language models have led to remarkable progress in complex visual understanding across scientific and reasoning tasks. |
| Approach: | They evaluate 18 state-of-the-art vision-language models across 6 multimodal datasets with 3 distinct scoring functions and develop instruction-guided likelihood proxies for closed-source models lacking token-level logprob access. |
| Outcome: | The proposed model is able to achieve higher accuracy on multimodal benchmarks while performing poorer on reasoning tasks. |
Diagnosis of Dysarthria Severity and Explanation Generation Using XAI-Enhanced CLINIC-GENIE on Diadochokinetic Tasks (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent deep learning approaches for dysarthria impairment severity lack interpretability essential for clinical applications. |
| Approach: | They propose a deep neural network classifier that integrates acoustic and speech embeddings with Clinically Explainable Acoustic Features (CEAFs) and a module that transforms CEAFs and their Shapley values into intuitive natural language explanations. |
| Outcome: | The proposed model achieves a balanced accuracy of 0.952 (17.3% improvement over using CEAFs alone) and certified speech-language pathologists rated explanations with an average fidelity score of 4.94, confirming enhanced clinical utility. |
A Comprehensive Evaluation of Multilingual Chain-of-Thought Reasoning: Performance, Consistency, and Faithfulness Across Languages (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent work has examined final-answer accuracy in multilingual settings, but the behavior of thinking traces, i.e., the intermediate steps that lead to the final answer, remains underexplored. |
| Approach: | They propose to measure language compliance, answer accuracy, and answer consistency when LRMs are explicitly instructed or prompt-hacked to think in a target language. |
| Outcome: | The proposed model improves in English and other high-resource languages while relying on traces to varying degrees. |
ORSO QGen: Odds-Ratio Steerable Optimization for Controlling Question Generation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for question generation rely on supervised fine-tuning . scarcity of datasets with multiple fine-grained attributes limits ability of models to generalize or transfer control to new combinations of attributes. |
| Approach: | They propose a framework to enhance attribute sensitivity in question generation models . they propose supervised fine-tuning with explicit attribute labels to produce questions that fit predefined characteristics. |
| Outcome: | The proposed framework improves attribute sensitivity while maintaining quality of output. |
Leveraging Digitized Newspapers to Collect Summarization Data in Low-Resource Languages (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent studies suggest that summarization in English may be solved, or even "dead" However, there are no accessible, high-quality summarizing datasets in under-represented languages. |
| Approach: | They propose a method for collecting naturally occurring summaries via front-page teasers, where editors summarize full length articles. |
| Outcome: | The proposed method is suited to varying linguistic resources and is available in seven languages. |
Let’s Simplify Step by Step: Guiding LLM Towards Multilingual Unsupervised Proficiency-Controlled Sentence Simplification (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models demonstrate limited capability in proficiency-controlled sentence simplification when simplifying across large readability levels. |
| Approach: | They propose a framework that decomposes complex simplifications into manageable steps through dynamic path planning, semantic-aware exemplar selection, and chain-of-thought generation with conversation history for coherent reasoning. |
| Outcome: | The proposed framework reduces computational steps while improving simplification effectiveness on five languages across two benchmarks. |
LogToP: Logic Tree-of-Program with Table Instruction-tuned LLMs for Controlled Logical Table-to-Text Generation (2026.findings-eacl)
Copied to clipboard
Yupian Lin, Guangya Yu, Cheng Yuan, Huan Du, Hui Luo, Yuang Bian, Jingping Liu, Zhidong He, Wen Du, Tong Ruan
| Challenge: | Existing LLMs are difficult to achieve satisfactory results in table-related tasks. |
| Approach: | They propose to develop a specialized logical table-to-text generation model that can be used for table-related tasks. |
| Outcome: | The proposed model achieves state-of-the-art on a Logic2Text dataset. |
IRPO: Implicit Policy Regularized Preference Optimization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent DPOs introduce additional hyperparameters, reducing feasibility for LLM fine-tuning. |
| Approach: | They propose an algorithm that regularizes the reward against a reference policy without extra hyperparameters to address suboptimal outcomes. |
| Outcome: | The proposed algorithm outperforms baseline algorithms with the same hyperparameter complexity while maintaining training simplicity. |
DFPE: A Diverse Fingerprint Ensemble for Enhancing LLM Performance (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) exhibit inconsistent performance across diverse domains. |
| Approach: | They propose a method that systematically constructs subject-adaptive ensembles by balancing model diversity and competence. |
| Outcome: | The proposed method achieves 17.1% gain over the best single model, reaching 71.4% accuracy on the MMLU-pro benchmark. |
Ranking Human and LLM Texts Using Locality Statistics (2026.findings-eacl)
Copied to clipboard
| Challenge: | The paper extends the Data Movement Distance (DMD) metric defined to measure the locality in computer memory to text by defining a new term designed to better characterize low-frequency tokens. |
| Approach: | They propose to define a normalized version of the Data Movement Distance (nDMD) term is designed to better characterize low-frequency tokens. |
| Outcome: | The proposed normalized version outperforms baselines and improves performance on the English subset of the M4 dataset and the GenAI detection shared task. |
MangaVQA and MangaLMM: A Benchmark and Specialized Model for Multimodal Manga Understanding (2026.findings-eacl)
Copied to clipboard
Jeonghun Baek, Kazuki Egashira, Shota Onohara, Atsuyuki Miyai, Yuki Imajuku, Hikaru Ikuta, Kiyoharu Aizawa
| Challenge: | Manga is a richly multimodal narrative form that blends images and text in complex ways. |
| Approach: | They propose two benchmarks for multimodal manga understanding: mangaOCR and mangaVQA . mangaVQ consists of 526 high-quality, manually constructed question-answer pairs . |
| Outcome: | The proposed model is finetuned from the open-source LMM Qwen2.5-VL . it compares with proprietary models such as GPT-4o and Gemini 2.5 to evaluate its performance . |
Hierarchical User Intent Inference with Knowledge Graph Grounding (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing large language models lack structured grounding and do not capture nuanced intent expression. |
| Approach: | They propose a Hierarchical Intent Inference framework that first predicts fine-grained aspect ratings and then generates natural language intent statements guided by contextual subgraphs retrieved from a domain-specific knowledge graph. |
| Outcome: | The proposed framework outperforms strong LLM and encoder-based baselines on a hotel review dataset. |
Improving the OOD Performance of Closed-Source LLMs on NLI Through Strategic Data Selection (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to improve robustness require changing the fine-tuning process or large-scale data augmentation, which are infeasible or cost prohibitive for closed-source models. |
| Approach: | They propose to prioritize more complex examples or replace existing training examples with LLM-generated data to improve performance on OOD NLI datasets. |
| Outcome: | The proposed methods improve performance on difficult OOD datasets while training with synthetic data leads to substantial improvements on easier OOD data. |
MMRA: A Benchmark for Evaluating Multi-Granularity and Multi-Image Relational Association Capabilities in Large Visual Language Models (2026.findings-eacl)
Copied to clipboard
Siwei Wu, King Zhu, Yu Bai, Yiming Liang, Yizhi Li, Haoning Wu, Jiaheng Liu, Ruibo Liu, Xingwei Qu, Xuxin Cheng, Ge Zhang, Wenhao Huang, Chenghua Lin
| Challenge: | Current multimodal benchmarks focus on facts within individual images, but neglect associative relations among multiple images. |
| Approach: | They propose a multi-image relational association task and a MMRA benchmark to evaluate LVLMs. |
| Outcome: | The proposed benchmarks show that entity-level multi-image perception tasks pose greater challenges than image-level tasks. |
COIG-P: A High-Quality and Large-Scale Chinese Preference Dataset for Alignment with Human Values (2026.findings-eacl)
Copied to clipboard
Siwei Wu, JinCheng Ren, Xeron Du, Shuyue Guo, Xingwei Qu, Yiming Liang, Jie Liu, Yunwen Li, Tyler Loakman, Tianyu Zheng, Boyu Feng, Huaqing Yuan, Zili Wang, Jiaheng Liu, Wenhao Huang, Chenglin Cai, Haoran Que, Jian Yang, Yuelin Bai, Zekun Moore Wang, Zhouliang Yu, Qunshu Lin, Ding Pan, Yuchen Eleanor Jiang, Tiannan Wang, Wangchunshu Zhou, Shenzhi Wang, Xingyuan Bu, Minghao Liu, Guoyin Wang, Ge Zhang, Chenghua Lin
| Challenge: | Existing Chinese preference datasets suffer from limited scale, restricted domain coverage, and insufficiently rigorous data validation. |
| Approach: | They propose an LLM-based data annotation pipeline with no human intervention to annotate Chinese preference datasets. |
| Outcome: | The proposed pipeline outperforms existing Chinese preference datasets on AlignBench and Chinese Reward Benchmark. |
Revealing the Numeracy Gap: An Empirical Investigation of Text Embedding Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Text embedding models are widely used in natural language processing but are often benchmarked on tasks that do not require understanding nuanced numerical information in text. |
| Approach: | They evaluate 13 widely used text embedding models and find they struggle to capture numerical details accurately. |
| Outcome: | The proposed models struggle to capture nuanced numerical details accurately, despite being benchmarked on tasks that do not require understanding nuance. |
code-transformed: The Influence of Large Language Models on Code (2026.findings-eacl)
Copied to clipboard
| Challenge: | Using Large Language Models, code generation capabilities have transformed programming practices. |
| Approach: | They analyze 20,000 GitHub repositories linked to arXiv papers published between 2020 and 2025 . they identify measurable trends in the evolution of coding style that align with LLM-generated code . |
| Outcome: | The proposed study examines 20,000 GitHub repositories linked to arXiv papers . it finds that LLMs influence code style, and that they can be observed in real-world code . |
Do LLMs model human linguistic variation? A case study in Hindi-English Verb code-mixing (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing large language models (LLMs) do not reliably classify verb language preferences to match native speaker judgments. |
| Approach: | They investigate whether large language models (LLMs) model linguistic variation by comparing Hindi-English verb code-mixing with English verb karna. |
| Outcome: | The proposed models do not reliably classify verb language preferences to match native speaker judgments, but with specific supervision, some models do predict human preference to an extent. |
ART: Attention-Regularized Transformers for Multi-Modal Robustness (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to enhancing robustness are domain-specific or lack formal guarantees. |
| Approach: | They propose a framework that enhances robustness across modalities by regularizing attention maps under adversarial perturbations. |
| Outcome: | The proposed framework improves robustness across modalities and training on IMDB, QNLI, CIFAR-10, Cifar-100, and Imagenette. |
GRAFF: GRaph-Augmented Fine-grained Fusion for Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to integrate graphs into LLMs compress the graph's structural information into a single token, restricting their ability to capture deep semantic and structural information. |
| Approach: | They propose a method that integrates fine-grained node-level structural information with corresponding text entities to LLMs via a lightweight, structure adapter module. |
| Outcome: | The proposed method outperforms baseline models in graph-based question answering by 10.24%. |
Tackling Distractor Documents in Multi-Hop QA with Reinforcement and Curriculum Learning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing work on retrieval-augmented generation systems has shown that retrievers exhibit imperfect recall and precision, limiting downstream performance. |
| Approach: | They propose a retrieval-augmented generation model that generates answers from larger sets of retrieved contexts. |
| Outcome: | The proposed model generates answers and cites relevant information from larger sets of retrieved contexts. |
RoD-TAL: A Benchmark for Answering Questions in Romanian Driving License Exams (2026.findings-eacl)
Copied to clipboard
Andrei Vlad Man, Răzvan-Alexandru Smădu, Cristian-George Craciun, Dumitru-Clementin Cercel, Florin Pop, Mihaela-Claudia Cercel
| Challenge: | a growing need for tools that support legal education, especially in under-resourced languages such as Romanian . we evaluate the capabilities of large language models and vision-language models in legal education . |
| Approach: | They evaluate the capabilities of Large Language Models and Vision-Language Models in Romanian driving law through textual and visual question-answering tasks. |
| Outcome: | The proposed model improves retrieval performance and QA accuracy in Romanian driving tests. |
FactSelfCheck: Fact-Level Black-Box Hallucination Detection for LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to detect hallucinated content are limited by their tendency to generate factual errors. |
| Approach: | They propose a black-box sampling-based method that enables fine-grained fact-level detection by representing text as interpretable knowledge graphs consisting of facts in the form of triples. |
| Outcome: | The proposed method improves hallucination correction by 35.5% compared to baseline methods while sentence-level SelfCheckGPT yields only 10.6% improvement. |
Punctuations and Predicates in Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent work has shown that LLMs perform tasks in ways that diverge significantly from human reasoning. |
| Approach: | They examine the computational importance of punctuation tokens in large language models . they use zeroing and layer-swapping techniques to examine their necessity and sufficiency . |
| Outcome: | The proposed model differs in GPT-2, DeepSeek, and Gemma in that punctuation is necessary and sufficient in multiple layers . the findings offer new insight into the internal mechanisms of punctuations in LLMs and have implications for interpretability and model analysis. |
Test-time Corpus Feedback: From Retrieval to RAG (2026.findings-eacl)
Copied to clipboard
| Challenge: | Retrieval-augmented generation (RAG) pipelines treat retrieval and reasoning as isolated components, limiting performance on complex tasks. |
| Approach: | They propose to integrate large language models with retrieval to improve query quality . they also propose to use feedback to improve the query, retrieved context, or document pool . |
| Outcome: | The proposed methods bridge IR and NLP perspectives and highlight retrieval as a dynamic, learnable component of end-to-end RAG systems. |
RADAR: A Reasoning-Guided Attribution Framework for Explainable Visual Data Analysis (2026.findings-eacl)
Copied to clipboard
| Challenge: | Multimodal Large Language Models (MLLMs) provide no visibility into which parts of visual data informed their conclusions. |
| Approach: | They propose a semi-automatic approach to attribute reasoning process by highlighting regions in charts and graphs that justify model answers. |
| Outcome: | The proposed method improves attribution accuracy by up to 15 percentage points compared to baseline methods and achieves high semantic similarity with ground truth responses. |
MaskLoRA: Low-Rank Subspace–Induced Token Masking for Efficient and Faithful Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | MASKLORA is a plug-and-play masking mechanism that can be used to mask lowrank subspaces. |
| Approach: | They propose a plug-and-play masking mechanism that transforms PEFT's lowrank subspace into a faithful token selector. |
| Outcome: | The proposed masking mechanism matches full-model accuracy while yielding 1.3-2.6 speedups. |
A Domain-Specific Curated Benchmark for Entity and Document-Level Relation Extraction (2026.findings-eacl)
Copied to clipboard
Marco Martinelli, Stefano Marchesin, Vanessa Bonato, Giorgio Di Nunzio, Nicola Ferro, Ornella Irrera, Laura Menotti, Federica Vezzani, Gianmaria Silvello
| Challenge: | Existing biomedical IE benchmarks are narrow in scope and rely heavily on distantly supervised annotations. |
| Approach: | They propose a benchmark for Information Extraction (IE) that annotates entities, concept-level links, and relations manually from PubMed abstracts. |
| Outcome: | The GutBrainIE benchmark is based on more than 1,600 PubMed abstracts, manually annotated by biomedical and terminological experts with fine-grained entities, concept-level links, and relations. |
What Matters to an LLM? Behavioral and Computational Evidences from Summarization (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly entrusted with the management of information. |
| Approach: | They combine behavioral and computational analyses to find out what LLMs prioritize . they generate length-controlled summaries and derive empirical importance distributions . |
| Outcome: | The proposed model converges on consistent importance patterns and clusters more by family than by size. |
Neural network embeddings recover value dimensions from psychometric survey items on par with human data (2026.findings-eacl)
Copied to clipboard
| Challenge: | Embedings from large language models can recover structure of human values . quantitative analysis reveals that SQuID addresses the challenge of obtaining negative correlations between dimensions without domain-specific fine-tuning or training data reannotation. |
| Approach: | They propose to use questionnaire item embeddings to recover human values from PVQ-RR . their results have implications for psychometrics and social science research . |
| Outcome: | The proposed method explains 55% variance in dimension-dimension similarities compared to human data. |
Compositional Reasoning via Joint Image and Language Decomposition (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches typically decompose only language queries, treating images as monolithic inputs. |
| Approach: | They propose a framework that decomposes both images and questions into visual sub-domains with corresponding sub-questions. |
| Outcome: | REDI achieves absolute accuracy improvements of 8.9%, 8.2%, and 16.0% over existing models. |
Better Call CLAUSE: A Discrepancy Benchmark for Auditing LLMs Legal Reasoning Capabilities (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing evaluation methods fail to probe fragile reasoning capabilities of large language models (LLMs) a single undetected defect can have catastrophic consequences, highlighting the need for a new class of benchmarks to stress-test their reliability against nuanced contractual flaws. |
| Approach: | They propose a benchmark to evaluate the fragility of an LLM’s legal reasoning by producing over 7500 real-world perturbed contracts from foundational datasets like CUAD and ContractNLI. |
| Outcome: | The proposed benchmark evaluates LLMs' ability to detect and reason about fine-grained discrepancies from over 7500 real-world perturbed contracts from foundational datasets like CUAD and ContractNLI. |
Token-Wise Kernels (TWiKers) for Vicinity-Aware Attention in Transformers (2026.findings-eacl)
Copied to clipboard
| Challenge: | Token-Wise Kernels (TWiKers) are a novel enhancement to transformers that learn token-specific convolutional kernels applied to the keys or values. |
| Approach: | They propose a transformer enhancement that learns token-specific convolutional kernels applied to the keys or values. |
| Outcome: | The proposed transformers learn token-specific convolutional kernels applied to the keys or values . the results show that content words retain self-focus while function words shift attention toward their neighbors . |
Beyond a Single Extractor: Re-thinking HTML-to-Text Extraction for LLM Pre-training (2026.findings-eacl)
Copied to clipboard
Jeffrey Li, Joshua P Gardner, Doug Kang, Fangping Shi, Karanjeet Singh, Chun-Liang Li, Herumb Shandilya, David Leo Wright Hall, Oncel Tuzel, Percy Liang, Ludwig Schmidt, Hadi Pouransari, Fartash Faghri
| Challenge: | Existing open-source datasets predominantly apply a single fixed extractor to all webpages. |
| Approach: | They propose to take a Union over different extractors to improve model performance . they show that extractor choice can significantly impact downstream task performance based on content type . |
| Outcome: | The proposed approach can increase the token yield of DCLM-Baseline by 71% while maintaining benchmark performance. |
Can Models Help Us Create Better Models? Evaluating LLMs as Data Scientists (2026.findings-eacl)
Copied to clipboard
| Challenge: | Current benchmarks assess LLMs on more isolated capabilities, such as language understanding and question-answering. |
| Approach: | They propose a benchmark to evaluate the ability of large language models (LLMs) to perform feature engineering. |
| Outcome: | The proposed benchmark evaluates the ability of large language models to perform feature engineering, a critical and knowledge-intensive task in data science. |
Distill and Align Decomposition for Enhanced Claim Verification (2026.findings-eacl)
Copied to clipboard
Jabez Magomere, Elena Kochkina, Samuel Mensah, Simerjot Kaur, Fernando Acero, Arturo Oncevay, Charese Smiley, Xiaomo Liu, Manuela Veloso
| Challenge: | Existing methods for complex claim verification struggle to align decomposition quality with verification performance. |
| Approach: | They propose a reinforcement learning approach that optimizes decomposition quality and verifier alignment using Group Relative Policy Optimization. |
| Outcome: | The proposed method outperforms prompt-based approaches and existing methods in six evaluation settings. |
Argument-Based Consistency in Toxicity Explanations of LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to evaluate free-form toxicity explanations are overly relying on input text perturbations. |
| Approach: | They propose a multi-dimensional criterion to evaluate LLMs' reasoning about toxicity . they conduct experiments on three Llama models and an 8B Ministral model . |
| Outcome: | The proposed criterion measures the extent to which LLMs’ free-form toxicity explanations reflect an ideal and logical argumentation process. |
Reasoning Beyond Literal: Cross-style Multimodal Reasoning for Figurative Language Understanding (2026.findings-eacl)
Copied to clipboard
| Challenge: | figurative language is essential for expressing intent, emotion, and perspective . figural language is often dependent on Styles Reasoning, causing incongruities between expressions . |
| Approach: | They propose a framework that induces reasoning capabilities to compact vision–language models . figurative language is essential in expressing intent, emotion, and perspective . |
| Outcome: | The proposed framework can interpret multimodal figurative language, provide transparent reasoning traces, and generalize across multiple figurativ styles. |
QueStER: Query Specification for Generative Keyword-Based Retrieval (2026.findings-eacl)
Copied to clipboard
| Challenge: | Generative retrieval (GR) models can be expensive and brittle out of domain. |
| Approach: | They propose a query specification for gEnerative Keyword-Based Retrieval which bridges GR and query reformulation by learning to generate explicit keyword-based search specifications. |
| Outcome: | The proposed query specification improves over existing queries and maintains strong efficiency. |
Evaluating Sparse Autoencoders for Monosemantic Representation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Sparse autoencoders (SAEs) have been proposed to mitigate polysemanticity, where neurons activate for multiple unrelated concepts. |
| Approach: | They propose a sparse autoencoder to transform dense activations into sparser, more interpretable features by transforming them into sparses. |
| Outcome: | The proposed model reduces polysemanticity and achieves higher concept separability. |
Event Detection with a Context-Aware Encoder and LoRA for Improved Performance on Long-Tailed Classes (2026.findings-eacl)
Copied to clipboard
| Challenge: | Decoder-only models dominate the event detection literature, but their unidirectional attention mechanism has been a roadblock in getting strong performance on embedding. |
| Approach: | They propose to use Macro-F1 as a more representative measure of a model’s ability across the long-tail of event types to improve their models' performance. |
| Outcome: | The proposed model improves on the decoder-only models, showing that low-rank Adaptation can be an effective tool to enhance LLMs’ performance on long-tailed event classes. |
Think Hard Only When Needed: A Hybrid Best-of-N and Beam Search for Efficient Test-Time Compute (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) exhibit remarkable reasoning and planning capabilities, yet their substantial inference-time cost significantly impedes deployment in resourceconstrained applications. |
| Approach: | They propose a hybrid inference pipeline that combines beam search and Best-of-N . THROW generates shorter initial trajectories and evaluates them using PRMs . |
| Outcome: | THROW achieves 1.54 and 14.38 latency speedups and 35.7% and 80.4% token reductions on average compared to Best-of-N and beam search . |
Role-Conditioned Refusals: Evaluating Access Control Reasoning in Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) blur role boundaries by producing unrestricted responses. |
| Approach: | They propose to extend the Spider and BIRD text-to-SQL datasets with real-time PostgreSQl role-based policies at the table and column levels. |
| Outcome: | The proposed model improves refusal precision and lowers false permits. |
NL2Logic: AST-Guided Translation of Natural Language into First-Order Logic with Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Structured reasoning approaches that parse first-order logic rules from natural language lack syntax control and semantic faithfulness. |
| Approach: | They propose a structured reasoning paradigm that parses first-order logic rules from natural language and delegates inference to automated solvers. |
| Outcome: | a proposed framework parses first-order logic rules from natural language and delegates inference to automated solvers. |
Coding Agents with Multimodal Browsing are Generalist Problem Solvers (2026.findings-eacl)
Copied to clipboard
| Challenge: | specialized AI agents with task-specific tools or architectures fail to generalize beyond their intended scope. |
| Approach: | They propose a single-agent system with a modest number of general tools . they propose to generalize across software engineering, deep research and web browsing . |
| Outcome: | The proposed system achieves superior or competitive performance over specialized agents on three benchmarks. |
Quantifying Data Contamination in Psychometric Evaluations of LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies have raised concerns about data contamination from psychometric inventories . however, there is no systematic attempt to quantify the extent of data contamination . |
| Approach: | They propose a framework to measure data contamination in psychometric evaluations of Large Language Models by item memorization, evaluation memorisation and target score matching. |
| Outcome: | The proposed framework evaluates item memorization, evaluation memorisation, and target score matching in 21 models from major families and four widely used psychometric inventories. |
Task-aware Block Pruning with Output Distribution Signals for Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods to estimate block importance rely on representation similarity or computationally expensive sensitivity analyses to estimate task-aware model behavior. |
| Approach: | They propose a novel approach that quantifies block-level uncertainty from the statistics of each block’s early-exited output distribution on a calibration dataset. |
| Outcome: | Experiments show that the proposed approach preserves downstream task performance while reducing inference latency and computational cost. |
LARA: LLM-based Agile Power Distribution Network Restoration from Disastrous Events (2026.findings-eacl)
Copied to clipboard
| Challenge: | a large language model generates high-level restoration plans over a compact catalogue of feasible actions. |
| Approach: | They propose a method that generates restoration plans over a catalogue of feasible actions. |
| Outcome: | The proposed model outperforms a time-capped solver on an IEEE 13-node power distribution feeder by 13% while using less than 1% of its wall-clock runtime. |
Evaluating Multi-Hop Reasoning in Large Language Models: A Chemistry-Centric Benchmark (2026.findings-eacl)
Copied to clipboard
Mohammad Khodadad, Ali Shiraee Kasmaee, Mahdi Astaraki, Nicholas Sherck, Hamidreza Mahyar, Soheila Samiee
| Challenge: | a new pipeline for compositional multi-hop reasoning in large language models is being developed . a recent study shows that even state-of-the-art models struggle with compositional reasoning . |
| Approach: | They propose a pipeline that builds benchmarks from proprietary or public data . they use generative reasoning models, chemical named-entity recognition, and external knowledge bases to build knowledge graphs. |
| Outcome: | The proposed pipeline compares state-of-the-art models with and without retrieval augmentation . the pipeline is generalizable with fine-tuning, enabling creation of challenging benchmarks . |
SD-E2: Semantic Exploration for Reasoning Under Token Budgets (2026.findings-eacl)
Copied to clipboard
| Challenge: | Small language models struggle with complex reasoning because exploration is expensive under tight compute budgets. |
| Approach: | They propose a framework that makes exploration explicit by optimizing semantic diversity in generated reasoning trajectories. |
| Outcome: | The proposed framework surpasses Qwen2.5-3B-Instruct and strong GRPO baselines on GSM8K and improves on the harder AIME benchmark to 13.28% vs. base 6.74%. |
How to Contextualize Empirical Data for Risk Analysis with LLMs: A Case Study of Power Outages (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly being considered for high-stakes decision-making, yet their application in statistical risk analysis remains largely underexplored. |
| Approach: | They propose a method for extracting key information from raw data and translating it into structured contextual input within the LLM prompt. |
| Outcome: | The proposed approach significantly improves the LLM’s performance in risk assessment tasks. |
Thinking Beyond the Local: Multi-View Instructed Adaptive Reasoning in KG-Enhanced LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for large language models adopt query-driven iterative reasoning from a local perspective, limiting efficiency and accuracy for complex multi-hop tasks. |
| Approach: | They propose a multi-view instructed adaptive reasoning of LLM on Knowledge Graphs that allows LLMs to plan, evaluate, and adapt reasoning paths from a global perspective. |
| Outcome: | The proposed model overcomes the limitations of local exploration by enabling LLMs to plan, evaluate, and adapt reasoning paths from a global perspective. |
DAMASHA: Detecting AI in Mixed Adversarial Texts via Segmentation with Human-interpretable Attribution (2026.findings-eacl)
Copied to clipboard
| Challenge: | a new framework for mixed authorship detection addresses the challenge of segmenting mixed-authorship text . mixed-authored text detection is a growing concern in the age of advanced large language models . a recent survey highlighted the greater challenges of detecting AI content in realworld settings . |
| Approach: | They propose a framework for mixed authorship detection that integrates stylometric cues, perplexity-driven signals, and structured boundary modeling to accurately segment collaborative human-AI content. |
| Outcome: | The proposed framework improves robustness against adversarial perturbations while revealing limitations. |
FINEST: Improving LLM Responses to Sensitive Topics Through Fine-Grained Evaluation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing evaluation frameworks lack systematic methods to identify weaknesses in LLMs . Existing methods to evaluate LLM responses to sensitive topics are lacking . |
| Approach: | They propose a FINE-grained response evaluation taxonomy for sensitive topics that breaks down helpfulness and harmlessness into errors across three main categories: Content, Logic, and Appropriateness. |
| Outcome: | The proposed model outperforms refinement without guidance on Korean-sensitive questions . FINEST significantly improves the model responses across all three categories . |
Turn-PPO: Turn-Level Advantage Estimation with PPO for Improved Multi-Turn RL in Agentic LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Reinforcement learning (RL) has re-emerged as a natural approach for training interactive LLM agents in real-world environments. |
| Approach: | They propose a variant that operates on a turn-level MDP formulation, instead of the commonly used token-level one. |
| Outcome: | The proposed method is more robust than the widely used GRPO algorithm and more efficient than token-level MDPs. |
Decoding Time Series with LLMs: A Multi-Agent Framework for Cross-Domain Annotation (2026.findings-eacl)
Copied to clipboard
Minhua Lin, Zhengzhang Chen, Yanchi Liu, Xujiang Zhao, Zongyu Wu, Junxiang Wang, Xiang Zhang, Suhang Wang, Haifeng Chen
| Challenge: | Time series data is ubiquitous across various domains, including manufacturing, finance, and healthcare. |
| Approach: | They propose a multi-agent system to generate general and domain-specific annotations for time series data. |
| Outcome: | The proposed system outperforms existing methods on synthetic and real-world datasets. |
Multi-Hall-SA: A Cross-lingual Benchmark for Multi-Type Hallucination Detection in Low-Resource South African Languages (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models generate false or unsupported information, which can be difficult to detect in low-resource languages. |
| Approach: | They propose a cross-lingual benchmark for hallucination detection spanning English and South African languages. |
| Outcome: | The proposed model detects 23.6% fewer hallucinations in South African languages compared to English . human validation confirms the quality and cross-lingual alignment of the model . |
Query4Regex: Verifiable Regex Transformation through Formal Operations from NL and DSL Queries (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) generate structured data, but their ability to precisely manipulate it remains relatively under-explored. |
| Approach: | They propose a benchmark to evaluate verifiable transformations on regexes . they use natural language instructions and a program-like domain-specific language that specifies the sequence of operations to evaluate LLMs. |
| Outcome: | The proposed benchmark compares LLM performance on natural language and DSL queries for regex manipulation. |
SrcMix: Mixing of Related Source Languages Benefits Extremely Low-resource Machine Translation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Multilingual models are widely used for machine translation, but their effectiveness for extremely low-resource languages (ELRLs) is dependent on how related languages are incorporated during fine-tuning. |
| Approach: | They propose a source-side mixing strategy that combines related ELRLs during fine-tuning while constraining the decoder to a single target language. |
| Outcome: | The proposed approach improves performance in high-resource to ELRL translations and in mid-resourced to MT translations. |
IMRNNs: An Efficient Method for Interpretable Dense Retrieval via Embedding Modulation (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing dense retrieval methods rely on static embeddings that obscure bidirectional relationship between queries and documents. |
| Approach: | They propose a framework that augments any black-box dense retrievers with dynamic, bidirectional modulation at inference time. |
| Outcome: | a new framework augments any dense retriever with dynamic, bidirectional modulation at inference time. |
MMUIE: Massive Multi-Domain Universal Information Extraction for Long Documents (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing document-level information extraction systems operate at the sentence level or within narrow domains due to annotation constraints. |
| Approach: | They propose a large-scale universal dataset for multi-domain, document-level information extraction from long texts. |
| Outcome: | The proposed dataset integrates traditional knowledge bases with large language models to extract fine-grained entities, aliases, and relation triples across 34 domains. |
Learning to Judge: LLMs Designing and Applying Evaluation Rubrics (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models are increasingly used as evaluators for natural language generation . human rubrics are often static and misaligned with how models internally represent language quality. |
| Approach: | They propose to use large language models to generate interpretable and task-aware evaluation dimensions and apply them within models. |
| Outcome: | The proposed model improves the semantic coherence and scoring reliability of LLM-defined criteria and their alignment with human criteria. |
PsyProbe: Proactive and Interpretable Dialogue through User State Modeling for Exploratory Counseling (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing approaches to mental health dialogue are reactive and lack systematic user state modeling for proactive therapeutic exploration. |
| Approach: | They propose a dialogue system designed for the exploration phase of counseling that systematically tracks user psychological states through the PPPPPI framework augmented with cognitive error detection. |
| Outcome: | The proposed system outperforms baseline and ablation modes in automatic evaluation and expert evaluation by a certified counselor. |
Learning from Child-directed Speech in Two-language Scenarios: A French-English Case-Study (2026.findings-eacl)
Copied to clipboard
| Challenge: | a systematic study of compact language models with limited computational resources is challenging for many research contexts and real-world applications. |
| Approach: | They extend BabyBERTa to English-French scenarios under strictly sizematched data conditions. |
| Outcome: | The proposed model extends to English-French scenarios under sizematched data conditions . the results show context-dependent effects of multilingual training . |
DeVisE: Towards the Behavioral Testing of Medical Large Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing evaluations of large language models do not reveal whether their outputs reflect genuine medical reasoning or superficial correlations. |
| Approach: | They propose a framework that probes fine-grained clinical understanding through controlled counterfactuals. |
| Outcome: | The proposed framework is based on demographic and vital signs data from the ICU discharge notes of patients in the intensive care unit (MIMIC-IV). |
Sequence Repetition Enhances Token Embeddings and Improves Sequence Labeling with Decoder-only Language Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Modern language models (LMs) are trained in autoregressive manner, conditioned on the prefix. sequence labeling (SL) tasks assign labels to each individual input token, naturally benefiting from bidirectional context. |
| Approach: | They explore sequence repetition (SR) as a less invasive alternative to decoder-only models . they show that increasing the number of repetitions does not degrade SL performance . |
| Outcome: | The proposed technique improves the quality of token-level embeddings and surpasses encoders and unmasked decoders. |
MEENA (PersianMMMU): Multimodal-Multilingual Educational Exams for N-level Assessment (2026.findings-eacl)
Copied to clipboard
Omid Ghahroodi, Arshia Hemmat, Marzia Nouri, Seyed Mohammad Hadi Hosseini, Doratossadat Dastgheib, Mohammad Vali Sanian, Alireza Sahebi, Reihaneh Zohrabi, Mohammad Hossein Rohban, Ehsaneddin Asgari, Mahdieh Soleymani Baghshah
| Challenge: | Recent advances in large vision-language models have primarily focused on English, with limited attention given to other languages. |
| Approach: | They propose a dataset to evaluate Persian VLMs across scientific, reasoning, and human-level understanding tasks. |
| Outcome: | The proposed model performs well across scientific reasoning, reasoning, and human-level understanding tasks in Persian and English. |
Teaching Old Tokenizers New Words: Efficient Tokenizer Adaptation for Pretrained Models (2026.findings-eacl)
Copied to clipboard
| Challenge: | Extending existing vocabulary is a widely used step in adapting pre-trained language models to new domains or languages. |
| Approach: | They propose to extend a pre-trained tokenizer by continuing the BPE merge learning process on new data. |
| Outcome: | The proposed method improves tokenization efficiency and improves model utilization. |
AGIC: Attention-Guided Image Captioning to Improve Caption Relevance (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for image captioning generate generic captions that are limited in capturing nuanced visual details. |
| Approach: | They propose attention-guided image captioning which amplifies visual regions directly in the feature space to guide caption generation. |
| Outcome: | The proposed approach matches or surpasses state-of-the-art models while achieving faster inference. |
Visual–Linguistic Abductive Reasoning with LLMs for Knowledge-based Visual Question Answering (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent efforts to leverage large language models for reasoning focus on visual perception and language reasoning as separate processes. |
| Approach: | They propose a method that integrates visual and linguistic modalities into interpretable abductive reasoning chains. |
| Outcome: | The proposed method improves performance on AOKVQA, OKVQA and GQA by 2.31% . it uses fuzzy scoring to select the most coherent combination, enabling unified reasoning . |
FactAppeal: Identifying Epistemic Factual Appeals in News Media (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods focus on the content of factual statements and ignore the epistemic structures that confer credibility and persuasive force to these claims. |
| Approach: | They propose a task of Epistemic Appeal Identification to identify whether and how factual statements have been anchored by external sources or evidence. |
| Outcome: | The proposed task identifies whether and how factual statements have been anchored by external sources or evidence. |
Vietnamese Automatic Speech Recognition: A Revisit (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing datasets with low quality and inconsistent annotations are insufficient for high-quality models. |
| Approach: | They propose a pipeline for aggregating and preprocessing high-quality ASR datasets from diverse, potentially noisy, open-source sources. |
| Outcome: | The proposed pipeline provides a foundation for training and evaluating state-of-the-art Vietnamese ASR systems. |
MapCoder-Lite: Distilling Multi-Agent Coding into a Single Small LLM (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing large-scale (> 30 B) models are costly and collapse when downsized to small open-source models. |
| Approach: | They propose a framework for distilling large, multi-agent coding systems into a single 7B model. |
| Outcome: | The proposed framework doubles xCodeEval accuracy and reduces GPU memory and token generation time by 4 compared to a 32B model. |
When Do Language Models Endorse Limitations on Human Rights Principles? (2026.findings-eacl)
Copied to clipboard
Keenan Samway, Miu Nicole Takagi, Rada Mihalcea, Bernhard Schölkopf, Ilias Chalkidis, Daniel Hershcovich, Zhijing Jin
| Challenge: | a recent study evaluated how large language models navigate trade-offs involving the Universal Declaration of Human Rights. |
| Approach: | They evaluate how large language models navigate trade-offs involving the Universal Declaration of Human Rights (UDHR) they use 1,152 synthetically generated scenarios across 24 rights articles and eight languages . |
| Outcome: | The proposed models accept limiting economic, social, and cultural rights more often than political and civil rights, the authors show . their models show significant cross-linguistic variation with elevated endorsement rates of rights-limiting actions in Chinese and Hindi compared to English or Romanian . |
Abstractive Summarization of Bengali Academic Videos Based on Audio Subtitles (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing methods for summarizing educational videos in Bengali are limited due to the rapid growth of educational video content. |
| Approach: | They propose an end-to-end pipeline for the abstractive summarization of Bengali videos . they fine-tuned the BanglaT5 model on a new benchmark dataset . |
| Outcome: | The proposed system preprocesses audio and converts speech to text using Google's Speech Recognition API. |
Active Learning with Non-Uniform Costs for African Natural Language Processing (2026.findings-eacl)
Copied to clipboard
| Challenge: | Annotating datasets for African languages is challenging due to the continent's vast linguistic diversity, complicating development of NLP systems. |
| Approach: | They propose a cost-aware active learning method that integrates BatchBALD acquisition strategy with a 0-1 Knapsack optimization objective to select informative and budget-efficient samples. |
| Outcome: | The proposed method outperforms BALD, BatchBALD, and stochastic sampling variants across cost scenarios on the MasakhaNEWS multilingual news classification benchmark covering 11 African languages. |
CrisiText: A dataset of warning messages for LLM training in emergency communication (2026.findings-eacl)
Copied to clipboard
| Challenge: | Identifying threats and mitigating their potential damage during crisis situations is paramount for safeguarding endangered individuals. |
| Approach: | They present a large-scale dataset for the generation of warning messages across 13 different types of crisis scenarios. |
| Outcome: | The proposed dataset contains more than 400,000 warning messages (spanning almost 18,000 crisis situations) aimed at assisting civilians during and after such events. |
Training-Free Text Emotion Tagging via LLM-Based Best-Worst Scaling (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are often used as automatic annotators for tasks such as Text Emotion Recognition (TER). |
| Approach: | They propose a novel algorithm that leverages Best-Worst Scaling to prompt the LLM to choose the least and most suitable emotions for a given text from several label subsets. |
| Outcome: | The proposed method compares favorably to existing methods and naive prompting approaches in terms of accuracy and calibration. |
Scaling Cultural Resources for Improving Generative Models (2026.findings-eacl)
Copied to clipboard
Hayk Stepanyan, Aishwarya Verma, Andrew Zaldivar, Rutledge Chin Feman, Erin MacMurray van Liemt, Charu Kalia, Vinodkumar Prabhakaran, Sunipa Dev
| Challenge: | generative models have been known to have reduced performance in different global cultural contexts and languages. |
| Approach: | They construct a pipeline to collect and contribute culturally salient, multilingual data . they argue such data can assess the state of the global applicability of generative AI models . |
| Outcome: | The proposed pipeline can assess the state of the global applicability of our models and improve upon cross-cultural gaps. |
Cards Against Contamination: TCG-Bench for Difficulty-Scalable Multilingual LLM Reasoning (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent studies find 25-50% of evaluation datasets appear in training corpora . contamination hinders the possibility to differentiate memorization and reasoning skills. |
| Approach: | They propose a two-player trading card game that is contaminated by a public engine and hidden card implementations to prevent benchmark saturation. |
| Outcome: | The proposed benchmark is based on a new two-player trading card game similar to Magic: The Gathering. |
ILSIC: Corpora for Identifying Indian Legal Statutes from Queries by Laymen (2026.findings-eacl)
Copied to clipboard
| Challenge: | Existing studies have focused on the use of court judgments as input for legal Statute Identification (LSI) however, there is little research to explore the differences between court and laypeople data for LSI. |
| Approach: | They create a corpus of laypeople queries covering 500+ statutes from Indian law . they use court case judgements to compare between the two datasets . |
| Outcome: | The proposed corpus of laypeople queries covers 500+ statutes from Indian law . the results show that models trained on court judgements are ineffective . |