SYSTEMATIC ANALYSIS

Leak, Cheat, Repeat: Data Contamination and Evaluation Malpractices in Closed-Source LLMs (2024.eacl-long)

Challenge:	Recent approaches on non-extractive commonsense QA show increased performance . attention-based injection seems to be preferable for knowledge integration .
Approach:	They propose to use attention-based injection to integrate knowledge into commonsense QA models.
Outcome:	The proposed methods show that attention-based injection is preferable for knowledge integration, and that the degree of domain overlap plays a crucial role in determining model success.

Copied to clipboard

Simone Balloccu, Patrícia Schmidtová, Mateusz Lango, Ondrej Dusek

A Systematic Analysis of Base Model Choice for Reward Modeling (2025.emnlp-main)

Challenge:	Lack of access to model details has raised concerns about data contamination among researchers.
Approach:	They conduct the first systematic analysis of work using OpenAI’s GPT-3.5 and GPT-4, the most prominently used LLMs today, in the context of data contamination.
Outcome:	The proposed models have been exposed to 4.7M samples from 263 benchmarks during the first year after their release.

Copied to clipboard

Kian Ahrabian, Pegah Jandaghi, Negar Mokhberian, Sai Praneeth Karimireddy, Jay Pujara

Challenge:	Reinforcement learning from human feedback (RLHF) and reward modeling are key to training powerful large language models (LLMs).
Approach:	They propose to combine RLHF and reward modeling to boost model selection . they also demonstrate that a small set of benchmarks could be combined to boost the model selection.
Outcome:	The results show that the model selection can be improved by up to 14% compared to the most common (default) choice.

Do Multi-hop Readers Dream of Reasoning Chains? (D19-58)

Copied to clipboard

Haoyu Wang, Mo Yu, Xiaoxiao Guo, Rajarshi Das, Wenhan Xiong, Tian Gao

Question Answering

Challenge:	Existing models for multihop reasoning are limited in their performance . multi-hop reasoning requires the ability to gather information from multiple passages .
Approach:	They propose a method that provides the full reasoning chain of multiple passages instead of just one final passage where the answer appears.
Outcome:	The proposed model improves on existing models by providing the full reasoning chain of multiple passages instead of just one final passage where the answer appears.

An Empirical Study on the Overlapping Problem of Open-Domain Dialogue Datasets (2022.lrec-1)

Copied to clipboard

Yuqiao Wen, Guoqing Luo, Lili Mou

Natural Language Interfaces

Challenge:	Existing benchmark datasets for open-domain dialogue generation are advancing the field . overlapping between training and test sets can cause fake performance .
Approach:	They analyze dailyDialog and OpenSubtitles to find out how overlapping can be exploited to obtain fake state-of-the-art performance.
Outcome:	The proposed datasets are cleaned and set up for future research.

VehicleWorld: A Highly Integrated Multi-Device Environment for Intelligent Vehicle Interaction (2025.findings-emnlp)

Copied to clipboard

Jie Yang, Jiajun Chen, Zhangyue Yin, Shuo Chen, Yuxin Wang, Yiran Guo, Yuan Li, Yining Zheng, Xuanjing Huang, Xipeng Qiu

Challenge:	Traditional Function Calling (FC) approaches operate statelessly, requiring multiple exploratory calls to build environmental awareness before execution, leading to inefficiency and limited error recovery.
Approach:	They propose a state-based function call approach that maintains explicit system state awareness and implements direct state transitions to achieve target conditions.
Outcome:	The proposed approach outperforms traditional function calling approaches, achieving superior execution accuracy and reduced latency.

MALLM: Multi-Agent Large Language Models Framework (2025.emnlp-demos)

Copied to clipboard

Jonas Becker, Lars Benedikt Kaesberg, Niklas Bauer, Jan Philip Wahle, Terry Ruas, Bela Gipp

Out of Distribution, Out of Luck: Process Rewards Misguide Reasoning Models (2026.eacl-short)

Challenge:	Multi-agent debate (MAD) has demonstrated the ability to augment collective intelligence by scaling test-time compute and leveraging expertise.
Approach:	They propose an open-source framework that enables systematic analysis of multi-agent debates.
Outcome:	The proposed framework enables systematic analysis of multi-agent debate components.

Copied to clipboard

Alexey Dontsov, Anton Korznikov, Andrey V. Galichin, Elena Tutubalina

WHOSe Heritage: Classification of UNESCO World Heritage Statements of "Outstanding Universal Value” with Soft Labels (2021.findings-emnlp)

Challenge:	80% of reasoning model outputs respond to formatting artifacts rather than mathematical content.
Approach:	They evaluate process reward models that provide step-level feedback during inference . they identify distinct reward prediction patterns that differentiate reasoning from non-reasoning model outputs .
Outcome:	The proposed model fails to enhance and sometimes degrade reasoning model performance.

Copied to clipboard

Nan Bai, Renqian Luo, Pirouz Nourian, Ana Pereira Roders

MAPS: A Multilingual Benchmark for Agent Performance and Security (2026.findings-eacl)

Challenge:	Evaluating and justifying Outstanding Universal Value (OUV) is essential for each site inscribed in the WHL . manual annotation of heritage values and attributes from multi-source textual data is knowledge-demanding and time-consuming.
Approach:	They propose to use NLP to build a classifier on a dataset containing Statements of OUV.
Outcome:	The proposed model can reach 94.3% accuracy on a dataset containing Statements of OUV . the study is promising to be further developed and applied in heritage research and practice.

Copied to clipboard

Omer Hofman, Jonathan Brokman, Oren Rachmil, Shamik Bose, Vikas Pahuja, Toshiya Shimizu, Trisha Starostina, Kelly Marchisio, Seraphina Goldfarb-Tarrant, Roman Vainshtein

Multilinguality

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.

Music Audio-Visual Question Answering Requires Specialized Multimodal Designs (2026.findings-acl)

Copied to clipboard

Wenhao You, Xingjian Diao, Wenjun Huang, Chunhui Zhang, Keyi Kong, Weiyi Wu, Chiyu Ma, Zhongyu Ouyang, Tingxuan Wu, Ming Cheng, Soroush Vosoughi, Jiang Gui

RECIPE-TKG: From Sparse History to Structured Reasoning for LLM-based Temporal Knowledge Graph Completion (2026.eacl-long)

Challenge:	Music audio-visual question answering presents unique challenges with dense audio-visual content, intricate temporal dynamics, and the need for domain-specific knowledge.
Approach:	They analyze Music AVQA datasets and analyze their results to identify key design patterns . they propose concrete future directions for incorporating musical priors .
Outcome:	The proposed architectures are critical for success in Music AVQA, the authors argue . they suggest concrete future directions for incorporating musical priors .

Copied to clipboard

Ömer Faruk Akgül, Feiyu Zhu, Yuxin Yang, Rajgopal Kannan, Viktor Prasanna

Where to start? Analyzing the potential value of intermediate models (2023.emnlp-main)

Challenge:	Temporal Knowledge Graphs (TKGs) represent dynamic facts as timestamped relations between entities. Large Language Models (LLMs) have sparked interest in using pretrained generative models for TKG completion.
Approach:	They propose a framework that allows for rule-based multi-hop sampling and contrastive fine-tuning to shape relational compatibility.
Outcome:	Experiments show that RECIPE-TKG outperforms prior LLM-based methods across input regimes.

Copied to clipboard

Leshem Choshen, Elad Venezian, Shachar Don-Yehiya, Noam Slonim, Yoav Katz

Are My Optimized Prompts Compromised? Exploring Vulnerabilities of LLM-based Optimizers (2026.eacl-long)

Challenge:	a finetuned model may be better base models than the vanilla pretrained model . this scheme, often referred to as intertraining, is the focus of the present work .
Approach:	They propose a scheme to analyze the potential intertraining gain independently for the target dataset and for a base model being considered as a starting point.
Outcome:	The proposed model is strong even if training data was not aligned with target dataset.

Copied to clipboard

Andrew Zhao, Reshmi Ghosh, Vitor R. Carvalho, Emily Lawton, Keegan Hines, Gao Huang, Jack W. Stokes

On the Transferability of Minimal Prediction Preserving Inputs in Question Answering (2021.naacl-main)

Challenge:	Recent studies have focused on poisoning during supervised fine-tuning, RLHF, or inference-time time optimization.
Approach:	They propose a simple fake reward attack that requires no access to the reward model and significantly increases vulnerability.
Outcome:	The proposed attack reduces the fake reward ASR from 0.23 to 0.07 without degrading utility.

Copied to clipboard

Shayne Longpre, Yi Lu, Chris DuBois

Natural Language Interfaces

Challenge:	Recent work establishes the presence of short, uninterpretable input fragments that yield high confidence and accuracy in neural models.
Approach:	They investigate competing hypotheses for the existence of MPPIs in question answering . they discover a perplexing invariance of MPIs to random training seed, model architecture, pretraining, and training domain.
Outcome:	The proposed model performance is higher than comparable short queries.

Specialization through Collaboration: Understanding Expert Interaction in Mixture-of-Expert Large Language Models (2026.eacl-long)

Copied to clipboard

Yuanbo Tang, Naifan Zhang, Yan Tang, Meixuan Chen, Shuhan Huang, Tingyu Cao, Yang Li

Texts or Images? A Fine-grained Analysis on the Effectiveness of Input Representations and Models for Table Question Answering (2025.findings-acl)

Challenge:	Mixture-of-Experts (MoE) based large language models are popular for multitasking . however, whether each expert can specialize to a task remains unclear .
Approach:	They propose to use a dictionary learning approach to analyze expert collaboration mechanisms in MoE LLMs.
Outcome:	The proposed model outperforms existing methods by 2.5% while enabling 50% expert reduction.

Copied to clipboard

Wei Zhou, Mohsen Mesgar, Heike Adel, Annemarie Friedrich

Robust Multi-bit Natural Language Watermarking through Invariant Features (2023.acl-long)

Challenge:	Existing studies on table question answering use either serialized text or images . lack of controlled setups limits fine-grained distinctions between these approaches.
Approach:	They propose a method selecting table representations dynamically and compare their performance to previous studies.
Outcome:	The proposed method improves performance 10% compared to using both representations indiscriminately.

Copied to clipboard

KiYoon Yoo, Wonhyuk Ahn, Jiho Jang, Nojun Kwak

Is “moby dick” a Whale or a Bird? Named Entities and Terminology in Speech Translation (2021.emnlp-main)

Challenge:	Recent years have witnessed a proliferation of valuable original natural language contents found in subscription-based media outlets, web novel platforms, and outputs of large language models.
Approach:	They propose a multi-bit watermarking framework that embeds information and extracts watermarks in a robust manner despite possible corruption by following a well-known proposition from image watermark.
Outcome:	The proposed model improves on the previous work on payload and robustness by +16.8% point on average on four datasets, three corruption types, and two corruption ratios.

Copied to clipboard

Marco Gaido, Susana Rodríguez, Matteo Negri, Luisa Bentivogli, Marco Turchi

Machine Translation

Challenge:	Among rare words, named entities and domain-specific terms are crucial . previous studies have neglected these important words due to limited options .
Approach:	They propose a benchmark to evaluate automatic translation systems for rare words . named entities and domain-specific terms are crucial for their translation .
Outcome:	The proposed benchmark is based on European Parliament speeches annotated with NEs and terminology.

Why Do Document-Level Polarity Classifiers Fail? (2021.naacl-main)

Copied to clipboard

Karen Martins, Pedro O.S Vaz-de-Melo, Rodrygo Santos

VishBox v2: A Multi-Agent System for Adaptive Voice Phishing Simulation (2026.acl-industry)

Challenge:	a new method to characterize, quantify and measure the impact of hard instances is proposed . a method to label hard instances can shed light on why and when classifiers fail, authors say .
Approach:	They propose a method to characterize, quantify and measure the impact of hard instances in polarity classification of movie reviews.
Outcome:	The proposed method can quantify the impact of hard instances in polarity classification . it can shed light on why and when classifiers fail, the authors say .

Copied to clipboard

Sungmi Park, Daon Choi, Yoonmo Yang, Hong Yunyi, Heedou Kim

Dialogue Systems & Conversational Agents

Challenge:	Existing alternatives lack round-level granularity and controllability, making it difficult to analyze how victim profiles and psychological states shape tactic effectiveness.
Approach:	They propose a multi-agent architecture that generates structured phishing simulations grounded in crime-script procedures and persuasion principles.
Outcome:	The proposed framework captures tactic concentration, vulnerability transitions, and web-search-induced procedural disruptions across 571 rounds.

E-Bench: Towards Evaluating the Ease-of-Use of Large Language Models (2025.coling-main)

Copied to clipboard

Zhenyu Zhang, Bingguang Hao, Jinpeng Li, Zekai Zhang, Dongyan Zhao

Does Object Grounding Really Reduce Hallucination of Large Vision-Language Models? (2024.emnlp-main)

Challenge:	E-Bench is a framework for easy-to-use research on large language models.
Approach:	They propose to evaluate the ease-of-use of large language models and construct an E-Bench . they simulate human use from synonymous and typographical perturbations .
Outcome:	The proposed model is able to resist synonymous expressions and typos and improves performance.

Copied to clipboard

Gregor Geigle, Radu Timofte, Goran Glavaš

Training Language Models to Generate Text with Citations via Fine-grained Rewards (2024.acl-long)

Challenge:	Large vision-language models (LVLMs) often hallucinate and produce captions that mention concepts that cannot be found in the image.
Approach:	They propose to add grounding objectives to captions that explicitly align image regions or objects to text spans to reduce hallucination.
Outcome:	The proposed evaluation protocol reduces the amount of hallucination in LVLMs by adding grounding objectives.

Copied to clipboard

Chengyu Huang, Zeqiu Wu, Yushi Hu, Wenya Wang

RATSD: Retrieval Augmented Truthfulness Stance Detection from Social Media Posts Toward Factual Claims (2025.findings-naacl)

Challenge:	Recent Large Language Models (LLMs) are prone to hallucination and their outputs often contain incorrect or unverifiable claims.
Approach:	They propose a training framework using fine-grained rewards to teach LLMs to generate highly supportive and relevant citations while ensuring the correctness of their responses.
Outcome:	The proposed training framework outperforms existing methods on QA datasets and surpasses GPT-3.5-turbo on LLaMA-2-7B.

Copied to clipboard

Zhengyuan Zhu, Zeyu Zhang, Haiqi Zhang, Chengkai Li

Sentiment Analysis

Challenge:	Social media provides a valuable lens for assessing public perceptions and opinions.
Approach:	They propose a method that leverages large language models with retrieval-augmented generation to analyze tweets in relation to claims.
Outcome:	The proposed method outperforms state-of-the-art methods on a new dataset . it shows that it outperformed existing methods and achieves a significant increase in Macro-F1 score on TSD-CT.

Be a Multitude to Itself: A Prompt Evolution Framework for Red Teaming (2024.findings-emnlp)

Copied to clipboard

Rui Li, Peiyi Wang, Jingyuan Ma, Di Zhang, Lei Sha, Zhifang Sui

Mitigating Hallucinations in Multimodal Spatial Relations through Constraint-Aware Prompting (2025.findings-naacl)

Challenge:	Large Language Models (LLMs) have gained increasing attention for their capacity to generate harmful content.
Approach:	They propose a scalable evolution framework to evolve red teaming prompts across breadth and depth dimensions, facilitating automatic generation of numerous high-quality and diverse red team prompts.
Outcome:	The proposed framework surpasses existing red teaming methods on attack success rate and diversity.

Copied to clipboard

Jiarui Wu, Zhuo Liu, Hangfeng He

Knowing the Facts but Choosing the Shortcut: Understanding How Large Language Models Compare Entities (2026.eacl-long)

Challenge:	Existing research has explored methods to enhance the performance of large vision-language models in spatial relations.
Approach:	They propose a constraint-aware prompting framework to reduce spatial relation hallucinations by incorporating two types of constraints into the prompt.
Outcome:	The proposed framework improves on three widely-used spatial relation datasets.

Copied to clipboard

Hans Hergen Lehmann, Jae Hee Lee, Steven Schockaert, Stefan Wermter

CodeQA: A Question Answering Dataset for Source Code Comprehension (2021.findings-emnlp)

Challenge:	Large Language Models (LLMs) are increasingly used for knowledge-based reasoning tasks, yet understanding when they rely on genuine knowledge versus superficial heuristics remains challenging.
Approach:	They propose to ask LLMs to compare numerical attributes to find out which country has the highest population, France or Germany.
Outcome:	The proposed model comparisons show that heuristics override principled reasoning for larger models, while smaller models show no discrimination.

Copied to clipboard

Chenxiao Liu, Xiaojun Wan

Question Answering

Challenge:	False. a free-form question answering dataset can serve as a useful research benchmark for source code comprehension.
Approach:	They propose a free-form question answering dataset for source code comprehension . they implement syntactic rules and semantic analysis to transform code comments into question-answer pairs.
Outcome:	The proposed dataset can serve as a useful research benchmark for source code comprehension.

Sparse Latents Steer Retrieval-Augmented Generation (2025.acl-long)

Copied to clipboard

Chunlei Xin, Shuheng Zhou, Huijia Zhu, Weiqiang Wang, Xuanang Chen, Xinyan Guan, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun

Revisiting Query Variation Robustness of Transformer Models (2024.findings-emnlp)

Challenge:	In this study, we uncover interpretable latents that govern RAG behavior in large language models . Sparse Autoencoders are used to control large language model (LLM) behavior .
Approach:	They leverage Sparse Autoencoders within the LLaMA Scope to uncover latents that govern RAG behaviors.
Outcome:	The proposed model can be used to control large language models without architectural modifications.

Copied to clipboard

Tim Hagen, Harrisen Scells, Martin Potthast

ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model (2025.emnlp-main)

Challenge:	Despite their proficiency with natural language, transformer-based large language models are not robust to query variations such as typos and paraphrases.
Approach:	They extend their findings to include more recent large language models . they find that instruct-LLMs are more robust to query variations .
Outcome:	The proposed model can be prompted for robustness by a set of instruction-tuned LLMs.

Copied to clipboard

Zhongyi Zhou, Yichen Zhu, Minjie Zhu, Junjie Wen, Ning Liu, Zhiyuan Xu, Weibin Meng, Yaxin Peng, Chaomin Shen, Feifei Feng, Yi Xu

Analysing the role of lexical and temporal information in turn-taking through predictability (2026.eacl-long)

Challenge:	Recent advances in vision-language-action models prioritize robotic action mastery . however, models trained on visual-text pairs struggle to interpret multimodal data .
Approach:	They propose a framework that integrates multimodal data after initial control mastery and a Mixture-of-Experts architecture to minimize task interference.
Outcome:	The proposed framework surpasses state-of-the-art vision-language-action (VLA) methods on multimodal understanding benchmarks and achieves six times higher performance on visual question-answering datasets.

Copied to clipboard

Sean Leishman, Sarenne Wallbridge, Peter Bell

Dialogue Systems & Conversational Agents

Challenge:	Existing evaluations of spoken dialogue systems do not address which information sources drive predictions.
Approach:	They examine the role of lexical-temporal features on the predictability of turn structure by examining PairwiseTurnGPT, a full-duplex model of spoken dialogue transcripts.
Outcome:	The proposed model can produce fluent conversational output, but it does not guarantee realistic turn-taking behaviour.

Sparse Growing Transformer: Training-Time Sparse Depth Allocation via Progressive Attention Looping (2026.findings-acl)

Copied to clipboard

Yao Chen, Yilong Chen, Yinqi Yang, Junyuan Shang, Zhenyu Zhang, Zefeng Zhang, Shuaiyi Nie, Shuohuan Wang, Yu Sun, Hua Wu, Haifeng Wang, Tingwen Liu

Do LLMs Know and Understand Domain Conceptual Knowledge? (2025.findings-emnlp)

Challenge:	Existing approaches to increasing effective depth of LLMs rely on parameter reuse, extending computation through recursive execution.
Approach:	They propose a training-time sparse depth allocation framework that progressively increases depth for a small subset of parameters as training evolves.
Outcome:	The proposed model outperforms existing approaches to increasing the effective depth of language models while reducing training FLOPs overhead from approximately 16–20% to only 1–3% relative to a standard Transformer backbone.

Copied to clipboard

Sijia Shen, Feiyan Jiang, Peiyan Wang, Yubo Feng, Yuchen Jiang, Chang Liu

How are Prompts Different in Terms of Sensitivity? (2024.naacl-long)

Challenge:	Concept sememe tree is a hierarchical structure that represents lexical meaning by combining sememes and their relationships.
Approach:	They introduce a Neighbor Semantic Structure (NSS) and a Chain-of-Thought prompting method to evaluate the effectiveness of various Large Language Models (LLMs) in generating concept sememe trees.
Outcome:	The proposed method guides LLMs through an analysis of a term’s intrinsic core concepts, essential attributes, and semantic relationships, enabling the generation of concept sememe trees.

Copied to clipboard

Sheng Lu, Hendrik Schuff, Iryna Gurevych

Evaluating Subtitle Segmentation for End-to-end Generation Systems (2022.lrec-1)

Challenge:	In-context learning (ICL) has become one of the most popular learning paradigms due to the rapid development of large language models (LLMs).
Approach:	They propose a prompt analysis based on sensitivity and introduce sensitivity-aware decoding which incorporates sensitivity estimation as a penalty term in the standard greedy decoding.
Outcome:	The proposed approach is particularly useful when information in the input is scarce.

Copied to clipboard

Alina Karakanta, François Buet, Mauro Cettolo, François Yvon

Challenge:	Subtitle segmentation can be evaluated with sequence segmentation metrics against a human reference, but cannot be applied when systems generate outputs different than the reference, e.g. with end-to-end subtitling systems.
Approach:	They propose to use Sigma to evaluate subtitle segmentation against a human reference and a boundary projection method to disentangle the effect of good segmentation from text quality.
Outcome:	The proposed method disentangles the effect of good segmentation from text quality and is compared with existing metrics.

Corpora Generation for Grammatical Error Correction (N19-1)

Copied to clipboard

Jared Lichtarge, Chris Alberti, Shankar Kumar, Noam Shazeer, Niki Parmar, Simon Tong

Machine Translation

Challenge:	Grammatical Error Correction (GEC) is a computational task that requires large amounts of data to solve.
Approach:	They propose two approaches to generate large parallel datasets for GEC using publicly available Wikipedia edit histories using minimal filtration heuristics and round-trip translation through bridge languages.
Outcome:	The proposed methods yield similar sized parallel corpora with around 4B tokens and are far ahead of the state-of-the-art on the CoNLL ‘14 benchmark and the JFLEG task.

How Far Does BERT Look At: Distance-based Clustering and Analysis of BERT’s Attention (2020.coling-main)

Copied to clipboard

Yue Guan, Jingwen Leng, Chao Li, Quan Chen, Minyi Guo

How Many Answers Should I Give? An Empirical Study of Multi-Answer Reading Comprehension (2023.findings-acl)

Challenge:	Recent work on multi-head attention mechanism shows heuristics and clues in analyzing various aspects of the mechanism.
Approach:	They propose to cluster attention heatmaps into significantly different patterns through unsupervised clustering on top of a set of proposed features.
Outcome:	The proposed features can explain and calibrate different attention heads in Transformer models.

Copied to clipboard

Chen Zhang, Jiuheng Lin, Xiao Liu, Yuxuan Lai, Yansong Feng, Dongyan Zhao

LIBERTy: A Causal Framework for Benchmarking Concept-Based Explanations of LLMs with Structural Counterfactuals (2026.findings-acl)

Challenge:	Despite recent progress in multi-answer MRC, there is no systematic analysis of how this phenomenon arises and how to better address it.
Approach:	They develop a taxonomy to categorize commonly-seen multi-answer MRC instances and examine how well different paradigms deal with different types of multi-announced questions.
Outcome:	The proposed taxonomy categorizes commonly-seen multi-answer instances and analyzes how well different paradigms deal with different types of multi-announced instances.

Copied to clipboard

Gilat Toker, Nitay Calderon, Ohad Amosy, Roi Reichart

Explainability & Interpretability in NLP

Challenge:	Concept-based explanations quantify how high-level concepts influence model behavior . existing benchmarks rely on costly human-written counterfactuals that serves as imperfect proxy .
Approach:	They propose a framework for constructing datasets containing structural counterfactual pairs . they use a structured Causal Model to generate a concept-based explanation .
Outcome:	The proposed framework compares concept-based explanations to causal effects estimated from counterfactuals.

Conspiracy Theories and Where to Find Them on TikTok (2025.acl-long)

Copied to clipboard

Francesco Corso, Francesco Pierri, Gianmarco De Francisci Morales

Linguistic Theories

Challenge:	Existing studies on TikTok's potential to promote and amplify harmful content have not been conducted.
Approach:	They analyze a longitudinal dataset of 1.5M videos shared in the U.S. over three years and evaluate the effects of TikTok’s Creativity Program for monetization.
Outcome:	The proposed model achieves high precision in detecting harmful content, but its overall performance is comparable to fine-tuned traditional models such as RoBERTa.

Rethinking Model Selection and Decoding for Keyphrase Generation with Pre-trained Sequence-to-Sequence Models (2023.emnlp-main)

Copied to clipboard

Di Wu, Wasi Ahmad, Kai-Wei Chang

Extract, Denoise and Enforce: Evaluating and Improving Concept Preservation for Text-to-Text Generation (2021.emnlp-main)

Challenge:	Keyphrase generation is a longstanding task in NLP with widespread applications.
Approach:	They propose a likelihood-based decode-select algorithm for seq2seq PLMs that improves greedy search by an average of 4.7% semantic F1 across five datasets.
Outcome:	The proposed algorithm improves greedy search by an average of 4.7% semantic F1 across five datasets.

Copied to clipboard

Yuning Mao, Wenchang Ma, Deren Lei, Jiawei Han, Xiang Ren

Text Generation

Challenge:	Existing models for text-to-text generation do not explicitly focus on important concepts in the input and output.
Approach:	They propose a framework to automatically extract, denoise, and enforce important input concepts as lexical constraints.
Outcome:	The proposed framework performs comparably or better than its unconstrained counterpart on automatic metrics and receives better ratings in the human evaluation.

Probing the Need for Visual Context in Multimodal Machine Translation (N19-1)

Copied to clipboard

Ozan Caglayan, Pranava Madhyastha, Lucia Specia, Loïc Barrault

Adaptive Meta-learner via Gradient Similarity for Few-shot Text Classification (2022.coling-1)

Challenge:	Current work on multimodal machine translation (MMT) suggests that the visual modality is either unnecessary or only marginally beneficial.
Approach:	They propose to use the visual modality to combine visual and textual information to generate better translations by partially depriving models from source-side textual context.
Outcome:	The proposed model can combine visual and textual information to generate better translations under limited textual context.

Copied to clipboard

Tianyi Lei, Honghui Hu, Qiaoyang Luo, Dezhong Peng, Xu Wang

SGPVT: Self-Generated Proximal Visual Tokens for Mitigating Proximal Collateral Damage in MLLM Unlearning (2026.acl-long)

Challenge:	Existing methods for few-shot text classification suffer from overfitting due to the lack of matching between the few amount of samples and complicated models.
Approach:	They propose a method to improve model generalization ability to a new task by leveraging a meta-learner via gradient similarity method.
Outcome:	The proposed method improves few-shot text classification performance on several benchmarks.

Copied to clipboard

Jiaqi Li, Zhijing Zhang, Jiahui Geng, Sheng Bi, Chuanyi Zhang, Fan Liu, Guilin Qi

Who Gets Which Message? Auditing Demographic Bias in LLM-Generated Targeted Text (2026.findings-acl)

Challenge:	Existing approaches focus on general utility metrics, overlooking the preservation of semantically related concepts.
Approach:	They propose a method that introduces self-generated proximal visual tokens to prevent forgetting vulnerability.
Outcome:	The proposed framework outperforms existing methods in preserving semantically related concepts while achieving effective target unlearning.

Copied to clipboard

Tunazzina Islam

Challenge:	Large language models generate demographically conditioned persuasive texts at scale . authors argue that such capabilities raise questions about fairness and representational bias in automated communication.
Approach:	They propose a framework for evaluating demographic-conditioned targeted messages . they find gender- and age-based asymmetries in male- and youth-targeted messages a .
Outcome:	The proposed framework evaluates generated messages across three dimensions: lexical content, language style, and persuasive framing.

OCR Post Correction for Endangered Language Texts (2020.emnlp-main)

Copied to clipboard

Shruti Rijhwani, Antonios Anastasopoulos, Graham Neubig

Bit-Flip Error Resilience in LLMs: A Comprehensive Analysis and Defense Framework (2025.emnlp-main)

Challenge:	Currently, there is little to no data available to build natural language processing models for endangered languages.
Approach:	They propose a benchmark dataset of transcriptions for scanned books in three critically endangered languages and a method to improve OCR in these data-scarce settings.
Outcome:	The proposed method reduces the recognition error rate by 34% across the three endangered languages.

Copied to clipboard

Yuhang Chen, Zhen Tan, Ajay Kumar Jaiswal, Huaizhi Qu, Xinyu Zhao, Qi Lin, Yu Cheng, Andrew Kwong, Zhichao Cao, Tianlong Chen

Investigating Cross-Modal Skill Injection: Scenarios, Methods, and Hyperparameters (2026.acl-long)

Challenge:	Bit-flip errors (BFEs) are hardware faults where individual bits in memory or processing units are unintentionally flipped.
Approach:	They propose a novel defense strategy to mitigate bit-flip errors (BFEs) they propose bfe protection and a self-correction mechanism to minimize performance degradation .
Outcome:	The proposed defense strategy minimizes performance degradation while significantly improving robustness against BFEs.

Copied to clipboard

Zhiyu Xu, Lean Wang, Yuanxin Liu, Lei Li, Hao Zhou, Fandong Meng, Jie Zhou, Xu Sun

Do All Languages Cost the Same? Tokenization in the Era of Commercial Language Models (2023.emnlp-main)

Challenge:	Existing research lacks systematic analysis of the applicability and methodology of cross-modal skill injection.
Approach:	They investigate the applicability and methodology of cross-modal skill injection by integrating a domain-expert LLM into a VLM.
Outcome:	The proposed method enables transfer of domain-specific expertise from Large Language Models (LLMs) to VLMs without incurring additional training data requirements or significant computational overhead.

Copied to clipboard

Orevaoghene Ahia, Sachin Kumar, Hila Gonen, Jungo Kasai, David Mortensen, Noah Smith, Yulia Tsvetkov

Why Multimodal In-Context Learning Lags Behind? Unveiling the Inner Mechanisms and Bottlenecks (2026.acl-long)

Challenge:	Language models have evolved from being research prototypes to commercialized products offered as web APIs.
Approach:	They conduct a systematic analysis of the cost and utility of OpenAI’s language model API on multilingual benchmarks in 22 typologically diverse languages.
Outcome:	The proposed language model API performs poorly on multiple languages and speakers of a large number of languages are overcharged while obtaining poorer results.

Copied to clipboard

Yu Wang, Sharon Li

Bridging Robustness and Generalization Against Word Substitution Attacks in NLP via the Growth Bound Matrix Approach (2025.findings-acl)

Challenge:	In-context learning (ICL) enables models to adapt to new tasks via inference-time demonstrations.
Approach:	They propose a simple inference-stage enhancement method that reinforces task mapping transfer.
Outcome:	The proposed method strengthens task mapping transfer in multimodal models . it performs comparable to text-only ICL in zero-shot settings but degrades significantly under few-shot demonstrations.

Copied to clipboard

Mohammed Bouri, Adnane Saoud

MOLE: Metadata Extraction and Validation in Scientific Papers Using LLMs (2025.findings-emnlp)

Challenge:	Recent studies have shown that adversarial examples can alter models' predicted sentiment due to their sensitivity to specific word choices.
Approach:	They propose a regularization technique to improve NLP model robustness by reducing the impact of input perturbations on model outputs.
Outcome:	The proposed method outperforms state-of-the-art methods in adversarial defense.

Copied to clipboard

Zaid Alyafeai, Maged S. Al-shaibani, Bernard Ghanem

Information Extraction & Text Mining

Challenge:	Metadata extraction relies heavily on manual annotation of documents.
Approach:	They propose a framework that leverages Large Language Models to automatically extract metadata attributes from scientific papers covering datasets of languages other than Arabic.
Outcome:	The proposed framework automates the extraction of metadata attributes from Arabic scientific papers using large language models.

Mitigating Sequential Dependencies: A Survey of Algorithms and Systems for Generation-Refinement Frameworks in Autoregressive Models (2025.findings-emnlp)

Copied to clipboard

Yunhai Hu, Zining Liu, Zhenyuan Dong, Tianfan Peng, Bradley McDanel, Sai Qian Zhang

Challenge:	Sequential dependencies present a fundamental bottleneck in deploying large-scale autoregressive models .
Approach:	They analyze methods based on generation strategies and refinement mechanisms . they examine deployment strategies across computing environments and explore applications spanning text, images, and speech generation.
Outcome:	The proposed frameworks can be used to improve the quality of autoregressive models.

Transparent and Scrutable Recommendations Using Natural Language User Profiles (2024.acl-long)

Copied to clipboard

Jerome Ramos, Hossein A. Rahmani, Xi Wang, Xiao Fu, Aldo Lipani

Human Interest Framing across Cultures: A Case Study on Climate Change (2025.coling-main)

Challenge:	Recent advances in Large Language Models (LLMs) rely on implicit or explicit feedback from users to suggest new items, resulting in a lack of transparency and a user's ability to scrutinize and modify their preferences.
Approach:	They propose to use a natural language (NL) user profile to summarize a user's preferences and then use it to fine-tune a LLM using only NL profiles to make transparent and scrutable recommendations.
Outcome:	The proposed model performs on two benchmarking rating prediction datasets and is comparable to existing models.

Copied to clipboard

Gisela Vallejo, Christine de Kock, Timothy Baldwin, Lea Frermann

Challenge:	Human Interest (HI) framing is a narrative strategy that injects news stories with a relatable, emotional angle and a human face to engage the audience.
Approach:	They perform a systematic analysis of HI stories to understand its role in climate change reporting in English-speaking countries from four continents.
Outcome:	The proposed approach has shown to capture and retain readership and enhance political engagement of the population.

Improving the Robustness of Large Language Models via Consistency Alignment (2024.lrec-main)

Copied to clipboard

Yukun Zhao, Lingyong Yan, Weiwei Sun, Guoliang Xing, Shuaiqiang Wang, Chong Meng, Zhicong Cheng, Zhaochun Ren, Dawei Yin

What Do LLMs Learn First? Asymmetric Learning Dynamics of Input Complexity and Output Ambiguity in Preference Alignment (2026.acl-long)

Challenge:	Large language models have shown tremendous success in following user instructions and generating helpful responses, but their robustness is still far from optimal.
Approach:	They propose a two-stage training framework that helps a model generalize on following instructions via similar instruction augmentations.
Outcome:	The proposed training framework improves diversity and aligns the model with human expectations by differentiating subtle differences in similar responses.

Copied to clipboard

Mengyang Li, Jingwen Wang, Pinlong Zhao

Challenge:	Existing methods treat all preference pairs uniformly during training.
Approach:	They propose a training framework that maintains separate, adaptive pacing schedules for each dimension.
Outcome:	The proposed training framework outperforms curriculum baselines by 2.1% and 0.21 points . it achieves 42.3% length-controlled win rate on AlpacaEval 2.0 and 7.66 on MT-Bench .

BackdoorAgent: A Unified Framework for Backdoor Attacks on LLM-based Agents (2026.findings-acl)

Copied to clipboard

Yunhao Feng, Yige Li, Yutao Wu, Yingshui Tan, Yanming Guo, Yifan Ding, Kun Zhai, Xingjun Ma, Yu-Gang Jiang

Knowing When to Quit: Diagnosing and Training LLMs to Abort Futile Reasoning (2026.findings-acl)

Challenge:	Large language model (LLM) agents execute tasks through multi-step workflows that combine planning, memory, and tool use.
Approach:	They propose a modular framework that provides a unified view of backdoor threats in LLM agents.
Outcome:	The proposed framework provides a unified, agent-centric view of backdoor threats in LLM agents.

Copied to clipboard

Xinyan Guan, Jiali Zeng, Chunlei Xin, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun, Fandong Meng

Curing Miracle Steps in LLM Mathematical Reasoning with Rubric Rewards (2026.acl-long)

Challenge:	Large language models generate costly yet semantically void reasoning on beyond-capability tasks . the dominant failure mode is specious reasoning, superficially valid outputs with subtle hallucinations .
Approach:	They propose a capability-aligned reinforcement learning approach that aligns model behavior with capability boundaries.
Outcome:	The proposed model reduces futile reasoning while maintaining performance across tasks.

Copied to clipboard

Youliang Yuan, Qiuyang Mang, Jingbang Chen, Hong Wan, Xiaoyuan Liu, Junjielong Xu, Jen-tse Huang, Wenxuan Wang, Wenxiang Jiao, Pinjia He

Exploring Large Language Models for Classical Philology (2023.acl-long)

Challenge:	Existing models are susceptible to reward hacking, leading to a substantial overestimation of a model's reasoning ability.
Approach:	They propose a Rubric Reward Model that rewards the entire reasoning trajectory against problem-specific rubrics.
Outcome:	The proposed model outperforms outcome-only supervision on four math benchmarks and boosts Verified Pass@1024 from 26.7% to 62.6% and reduces the incidence of Miracle Steps by 71%.

Copied to clipboard

Frederick Riemenschneider, Anette Frank

Incorporating Diverse Perspectives in Cultural Alignment: Survey of Evaluation Benchmarks Through A Three-Dimensional Framework (2025.emnlp-main)

Challenge:	Recent advances in NLP have led to the creation of powerful language models for many languages including Ancient Greek and Latin.
Approach:	They propose to use encoder-only and encoder decoder architectures to create four models for Ancient Greek that vary along two dimensions for tasks of interest for Classical languages.
Outcome:	The proposed models improve on existing models of Ancient Greek and Latin and provide a large pre-training corpus for Ancient Greek to support the creation of a larger, comparable model zoo for Classical Philology.

Copied to clipboard

Meng-Chen Wu, Si-Chi Chin, Tess Wood, Ayush Goyal, Narayanan Sadagopan

Challenge:	Large Language Models (LLMs) serve diverse global audiences, making it critical for responsible AI deployment across cultures.
Approach:	They propose a framework that conceptualizes alignment along three dimensions: Cultural Group, Cultural Elements and Awareness Scope.
Outcome:	The proposed framework reveals critical gaps between benchmarks and real-world cultural biases . region dominates cultural group representation, social and political relations dominates coverage . majority of datasets adopt majority-focused Awareness Scope approaches .

Leveraging Large Language Models for NLG Evaluation: Advances and Challenges (2024.emnlp-main)

Copied to clipboard

Zhen Li, Xiaohan Xu, Tao Shen, Can Xu, Jia-Chen Gu, Yuxuan Lai, Chongyang Tao, Shuai Ma

Less Is More? Examining Fairness in Pruned Large Language Models for Summarising Opinions (2025.emnlp-main)

Challenge:	introducing Large Language Models (LLMs) has opened new avenues for assessing generated content quality, e.g., coherence, creativity, and context relevance.
Approach:	They propose a taxonomy for organizing existing LLM-based evaluation metrics and a structured framework to understand and compare them.
Outcome:	The proposed taxonomy offers a framework to understand and compare LLM-based evaluation methods.

Copied to clipboard

Nannan Huang, Haytham M. Fayek, Xiuzhen Zhang

Annotating the Annotators: Analysis, Insights and Modelling from an Annotation Campaign on Persuasion Techniques Detection (2025.findings-acl)

Challenge:	reducing the size of LLMs through post-training pruning has been studied, but its impact on model fairness remains unexplored.
Approach:	They propose a pruning method that removes parameters that are redundant for input processing but influential in output generation.
Outcome:	The proposed pruning method can maintain or improve fairness across models and tasks where existing methods have limitations.

Copied to clipboard

Davide Bassi, Dimitar Iliyanov Dimitrov, Bernardo D’Auria, Firoj Alam, Maram Hasanain, Christian Moro, Luisa Orrù, Gian Piero Turchi, Preslav Nakov, Giovanni Da San Martino

Challenge:	Existing annotation campaigns based on heuristic guidelines have not been thoroughly discussed.
Approach:	They propose a probabilistic model for optimizing intervention scheduling to reduce the cost of an expert oversight in annotation tasks.
Outcome:	The proposed model advocates for an expert oversight in annotation tasks and periodic quality audits to reduce costs.

Judging with Many Minds: Do More Perspectives Mean Less Prejudice? On Bias Amplification and Resistance in Multi-Agent Based LLM-as-Judge (2025.findings-emnlp)

Copied to clipboard

Chiyu Ma, Enpei Zhang, Yilun Zhao, Wenjun Liu, Yaning Jia, Peijun Qing, Lin Shi, Arman Cohan, Yujun Yan, Soroush Vosoughi

VisiPruner: Decoding Discontinuous Cross-Modal Dynamics for Efficient Multimodal LLMs (2025.emnlp-main)

Challenge:	LLM-as-Judge frameworks provide scalable alternative to human evaluation . but the question of how intrinsic biases manifest in these settings remains unexplored .
Approach:	They conduct systematic analysis of four bias types in multi-agent LLM-as-Judge frameworks . they find debate framework amplifies biases sharply after initial debate .
Outcome:	The proposed frameworks amplify biases after debate and show they are stronger in meta-judge scenarios.

Copied to clipboard

Yingqi Fan, Anhao Zhao, Jinlan Fu, Junlong Tong, Hui Su, Yijie Pan, Wei Zhang, Xiaoyu Shen

PuzzleVQA: Diagnosing Multimodal Reasoning Challenges of Language Models with Abstract Visual Patterns (2024.findings-acl)

Challenge:	Multimodal Large Language Models (MLLMs) suffer from significant computational overhead due to the quadratic growth of attention computations with the number of multimodal tokens.
Approach:	They propose a training-free pruning framework that prunes multimodal tokens without a trained pruning method.
Outcome:	The proposed pruning framework outperforms existing token pruning methods and generalizes across diverse MLLMs.

Copied to clipboard

Yew Ken Chia, Vernon Toh, Deepanway Ghosal, Lidong Bing, Soujanya Poria

VocalNet: Speech LLMs with Multi-Token Prediction for Faster and High-Quality Generation (2025.emnlp-main)

Challenge:	recognizing patterns and abstracting concepts are key to general intelligence, we show . state-of-the-art large multimodal models struggle to generalize well to simple abstract patterns .
Approach:	They evaluate large multimodal models with abstract patterns based on colors, numbers, sizes, and shapes.
Outcome:	The proposed model fails to generalize well to simple abstract patterns, the study shows . the model fails on single-concept puzzles, despite its sophistication .

Copied to clipboard

Yuhao Wang, Heyang Liu, Ziyang Cheng, Ronghua Wu, Qunshan Gu, Yanfeng Wang, Yu Wang

Can Large Language Models Infer Causal Relationships from Real-World Text? (2026.acl-long)

Challenge:	Experimental results show VocalNet outperforms existing open-source speech LLMs despite limited training data.
Approach:	They propose a scalable and model-agnostic training framework and a novel multi-token prediction paradigm for speech generation.
Outcome:	The proposed model outperforms open-source speech LLMs while outperforming existing open-sourced models.

Copied to clipboard

Ryan Saklad, Aman Chadha, Oleg V. Pavlov, Raha Moraffah

ArgBench: Benchmarking LLMs on Computational Argumentation Tasks (2026.findings-acl)

Challenge:	Existing work evaluating large language models relies on synthetic or simplified texts with explicit causal relationships.
Approach:	They develop a benchmark to evaluate LLMs' ability to infer causal relationships from texts . they use a dataset of texts with different levels of explicitness and complexity .
Outcome:	The proposed benchmark is the first-ever real-world dataset for this task.

Copied to clipboard

Yamen Ajjour, Carlotta Quensel, Nedim Lipka, Henning Wachsmuth

Rethinking Table Instruction Tuning (2025.findings-acl)

Challenge:	Argumentation skills are an essential toolkit for large language models (LLMs).
Approach:	They propose a benchmark to evaluate the generalizability of five LLM families across 46 computational argumentation tasks.
Outcome:	The proposed benchmark evaluates the generalizability of five LLM families across 46 computational argumentation tasks covering mining arguments, assessing perspectives, evaluating argument quality, reasoning about arguments, and generating arguments.

Copied to clipboard

Naihao Deng, Rada Mihalcea

Protein Large Language Models: A Comprehensive Survey (2025.findings-emnlp)

Challenge:	Existing studies have overlooked the impact of hyperparameters on table understanding abilities . authors show that smaller learning rates and fewer training instances can enhance table understanding while preserving general capabilities.
Approach:	They propose a hyperparameter-based instruction-tuned model for table-related tasks that improves out-of-domain table understanding ability and general capabilities.
Outcome:	The proposed model outperforms existing models on table-related tasks while maintaining strong out-of-domain generalization and general capabilities.

Copied to clipboard

Yijia Xiao, Wanjia Zhao, Junkai Zhang, Yiqiao Jin, Han Zhang, Zhicheng Ren, Renliang Sun, Haixin Wang, Guancheng Wan, Pan Lu, Xiao Luo, Yu Zhang, James Zou, Yizhou Sun, Wei Wang

Hidden in Plain Sight: Reasoning in Underspecified and Misspecified Scenarios for Multimodal LLMs (2025.emnlp-main)

Challenge:	Existing studies focus on specific aspects or applications, but this study provides a comprehensive overview of Protein-specific large language models.
Approach:	This paper proposes a structured taxonomy of state-of-the-art ProteinLLMs . they analyze how they leverage large-scale protein sequence data for improved accuracy .
Outcome:	The proposed model covers their architectures, training datasets, evaluation metrics, and diverse applications.

Copied to clipboard

Qianqi Yan, Hongquan Li, Shan Jiang, Yang Zhao, Xinze Guan, Ching-Chen Kuo, Xin Eric Wang

Beyond Text: Unveiling Privacy Vulnerabilities in Multi-modal Retrieval-Augmented Generation (2025.emnlp-main)

Challenge:	Multimodal large language models are increasingly deployed in open-ended, real-world environments where inputs are messy, underspecified, and not always trustworthy.
Approach:	They evaluate multimodal large language models in real-world environments where inputs are messy, underspecified, and not always trustworthy.
Outcome:	The proposed models fail to detect hidden issues even when they possess the necessary perceptual and reasoning skills.

Copied to clipboard

Jiankun Zhang, Shenglai Zeng, Jie Ren, Tianqi Zheng, Hui Liu, Xianfeng Tang, Hui Liu, Yi Chang

Mitigating Hallucination in Large Vision-Language Models through Aligning Attention Distribution to Information Flow (2025.findings-emnlp)

Challenge:	Large Multimodal Models (LMMs) can process text, images, and audio, but they introduce privacy vulnerabilities.
Approach:	They propose a compositional structured prompt attack to exploit MRAG privacy vulnerabilities . they show that LMMs can generate outputs resembling retrieved content .
Outcome:	The proposed approach generates outputs resembling retrieved content and exposes sensitive information.

Copied to clipboard

Jianfei Zhao, Feng Zhang, Xin Sun, Chong Feng

LLMs (Almost) Never Abstain Under Medical Uncertainty (2026.acl-long)

Challenge:	Decode-Only models propagate information from left to right, but the model's attention still focuses on the visual representations, resulting in hallucinations.
Approach:	They propose to leverage the core information embedded in semantic representations to enhance the model's visual understanding by leveraging the attention distributions.
Outcome:	The proposed method reduces hallucinations by 80% by aligning the attention distribution with the actual information flow.

Copied to clipboard

Alessio Cocchieri, Luca Ragazzi, Giuseppe Tagliavini, Gianluca Moro

Challenge:	Medical multiple-choice question answering (MCQA) benchmarks assume large language models should always commit to an answer.
Approach:	They propose a benchmark to evaluate medical abstention under uncertainty . they remove the gold answer and introduce an explicit "I abstain" option . results highlight abstinence as a critical but overlooked dimension of medical decision-making evaluation .
Outcome:	The new benchmark evaluates medical abstention under uncertainty.

Evaluation Pitfalls and Challenges in Multimedia Event Extraction (2026.acl-long)

Copied to clipboard

Philipp Seeberger, Steffen Freisinger, Tobias Bocklet, Korbinian Riedhammer

Information Extraction & Text Mining

Challenge:	Recent work has focused on textual content, but recent work has explored the integration of additional modalities to support more accurate and comprehensive event understanding.
Approach:	They propose to analyze the evaluation pitfalls of multimedia event extraction by combining textual and visual inputs to identify events and their arguments across multiple modalities.
Outcome:	The proposed model overestimates performance and performance of the proposed model in a series of controlled experiments under a strict evaluation framework.

Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts (2026.acl-long)

Copied to clipboard

Haolei Xu, Haiwen Hong, Hongxing Li, Rui Zhou, Yang Zhang, Longtao Huang, Hui Xue, Yongliang Shen, Weiming Lu, Yueting Zhuang

Schoenfeld’s Anatomy of Mathematical Reasoning by Language Models (2026.acl-long)

Challenge:	Existing multimodal Mixture-of-Experts models accurately perceive image content yet fail in subsequent reasoning . Seeing but not thinking phenomenon is a puzzling phenomenon .
Approach:	They propose a routing-guided intervention method that enhances domain expert activation.
Outcome:	The proposed method achieves consistent improvements on visual reasoning tasks.

Copied to clipboard

Ming Li, Chenrui Fan, Yize Cheng, Soheil Feizi, Tianyi Zhou

Unlocking the Black Box of Latent Reasoning: An Interpretability-Guided Approach to Intervention (2026.acl-long)

Challenge:	Large language models expose reasoning traces, yet their underlying cognitive structure and steps remain difficult to identify and analyze beyond surface-level statistics.
Approach:	They propose a framework that explicitly abstracts reasoning traces into functional reasoning steps such as Analysis, Explore, Implement, Verify, etc.
Outcome:	The proposed framework reveals reproducible thinking dynamics and structural differences between reasoning and non-reasoning models, which are not apparent from token-level views.

Copied to clipboard

Shuochen Chang, Tong Bai, Xiaofeng Zhang, Qianli Ma, Qingyang Liu, Zhaohe Liao, Yibo Miao, Li Niu

HookMoE: A learnable performance compensation strategy of Mixture-of-Experts for LLM inference acceleration (2025.emnlp-main)

Challenge:	Existing methods for large language models (LLMs) lack a coherent representation of reasoning steps.
Approach:	They propose a set of latent reasoning interventions that enable latent thinking and decode-time interventions that refine the latent process by imposing the identified geometric and semantic priors.
Outcome:	The proposed models unlock latent capabilities and improve reasoning accuracy without any parameter updates.

Copied to clipboard

Cheng Longkai, Along He, Mulin Li, Xie Xueshuo, Tao Li

Green & Sustainable NLP

Challenge:	Mixture of Experts (MoE) models have been a promising paradigm for scaling model capacity through top-k routing mechanisms.
Approach:	They propose a plug-and-play single-layer compensation framework that strategically inserts a lightweight trainable Hook module immediately preceding selected transformer blocks.
Outcome:	The proposed framework reduces the number of activated experts by more than 50% and achieves a 1.42 inference speed-up during the prefill stage.

IGenBench: Benchmarking the Reliability of Text-to-Infographic Generation (2026.acl-long)

Copied to clipboard

Yinghao Tang, Xueding Liu, Boyuan Zhang, Tingfeng Lan, Yupeng Xie, Jiale Lao, Yiyao Wang, Haoxuan Li, Tingting Gao, Bo Pan, Luoxuan Weng, Xiuqi Huang, Minfeng Zhu, Yingchaojie Feng, Yuyu Luo, Wei Chen

MARD: Module-Aware Reasoning Distillation for Language Models with Adaptive Supervision (2026.acl-long)

Challenge:	Generated infographics may appear correct at first glance but contain easily overlooked issues, such as distorted data encoding or incorrect textual content.
Approach:	They propose to evaluate reliability of text-to-infographic generation using IGenBench . they employ multimodal large language models to verify each question .
Outcome:	The proposed framework decomposes reliability verification into atomic yes/no questions based on a taxonomy of 10 question types.

Copied to clipboard

Wenqi Yang, Jianjun Li, Zhibo Zhang, Mingqian Ding, Yushen Fang

Beyond Single Representations: Multi-Model Embedding Fusion for Stable Text Classification (2026.acl-long)

Challenge:	Multi-step reasoning remains challenging for language models with limited capacity . et al., 2025) demonstrate remarkable reasoning capabilities across diverse tasks .
Approach:	They propose a module-aware reasoning distillation framework that explicitly targets key Transformer components for effective reasoning transfer.
Outcome:	The proposed framework targets key components for effective reasoning transfer . it adopts an offline distillation setting, where a strong teacher model provides reasoning trajectories in advance .

Copied to clipboard

Jiho Gwak, Yuchul Jung

Understanding the Behaviors of Environment-aware Information Retrieval (2026.acl-long)

Challenge:	Existing studies on embedding fusion have not evaluated the effectiveness of individual layers or the impact of combining embeddables from multiple models.
Approach:	They propose to combine embeddings from multiple models to improve performance across NLP tasks.
Outcome:	The proposed method improves performance on low-resource datasets and reduces the impact of any single model as the number of integrated models increases.

Copied to clipboard

Ruifeng Yuan, Chaohao Yuan, David Dai, Yu Rong, Hong Cheng, Hou Pong Chan, Chenghao Xiao

Towards Mitigating Hallucinations in Large Vision-Language Models by Refining Textual Embeddings (2026.findings-acl)

Challenge:	Recent retrieval-augmented generation approaches have demonstrated strong capability in handling complex queries.
Approach:	They propose a branching-based rollout technique that improves training stability . they find different retrievers exhibit distinct optimal query styles .
Outcome:	The proposed method improves training stability and improves retrieval-aware systems.

Copied to clipboard

Aakriti Agrawal, Gouthaman KV, Rohith Aralikatti, Gauri Jagatap, Jiaxin Yuan, Sarvesh Baskar, Vijay Kamarshi, Andrea Fanelli, Furong Huang