Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)

Language Directions in Multilingual LLMs: A Layer-wise Diagnostic Study of Token Alignment and Pretraining Imprint (2026.acl-srw)

Challenge:	Recent election cycles have seen a large number of false information spread across social media and news platforms.
Approach:	They propose a framework for automated claim verification using Large Language Models and Retrieval Augmented Generation.
Outcome:	The proposed frameworks are based on large-scale models and new methods such as Retrieval Augmented Generation (RAG).

Copied to clipboard

JaeSeong Kim, Suan Lee

Emergence of Minimal Circuits for Indirect Object Identification in Attention-Only Transformers (2026.acl-srw)

Challenge:	Using a unified probing framework, we analyze six multilingual LLMs across five languages.
Approach:	They analyze multilingual representations across five languages and analyze their behavior . they find that accuracy rises by +73.5 to +80.7 points from L0 to L1 on average .
Outcome:	The proposed framework enables a consistent and substantial early jump in accuracy across models . the token–language alignment measures where vocabulary sharing peaks .

Copied to clipboard

Rabin Adhikari

Why Large Language Models can Secretly Outperform Embedding Similarity in Information Retrieval (2026.acl-srw)

Challenge:	Large language models are difficult to reverse engineer because of their internal operations.
Approach:	They train small attention-only transformers on a symbolic version of the Indirect Object Identification task.
Outcome:	The proposed model with only two attention heads achieves perfect IOI accuracy despite lacking MLPs and normalization layers .

Copied to clipboard

Matei Benescu, Ivo Pascal de Jong

Detecting Hallucinations in Large Language Models via Internal Attention Divergence Signals (2026.acl-srw)

Challenge:	Recent studies show that similarity is a short-sighted interpretation of relevance . LLM-based Relevance Judgment Systems (LLM-RJS) can overcome this limitation .
Approach:	They propose that LLM-Based Relevance Judgment Systems can overcome short-sighted interpretation of relevance by embedding similarity instead of similarity.
Outcome:	The proposed methods outperform Neural Embedding Retrieval Systems by overcoming similarity limitation.

Copied to clipboard

Gijs van Dijk

Reflection in the Dark: Exposing and Escaping the Black Box in Reflective Prompt Optimization (2026.acl-srw)

Challenge:	Existing methods conflate fluency with correctness or require substantial computational overhead.
Approach:	They propose a single-pass uncertainty quantification method that uses attention matrices to estimate uncertainty without requiring repeated sampling or external models.
Outcome:	The proposed method performs well across multiple datasets, task types, and model families and is highly predictive of answer correctness.

Copied to clipboard

Shiyan Liu, Qifeng Xia, Qiyun Xia, Yisheng Liu, Xinyu Yu, Rui Qu

Thesis Proposal: LLMs post-training for multilingual medical tasks. Instruction-Tuning, Continual-Pretraining or Reasoning? (2026.acl-srw)

Challenge:	Automatic prompt optimization (APO) is a powerful paradigm for improving LLM performance without manual prompt engineering.
Approach:	They propose a framework that decouples hypothesis generation from prompt rewriting . they propose VISTA framework that recovers accuracy to 87.57% on same defective seed .
Outcome:	The proposed framework outperforms baselines on GSM8K and AIME2025 on a defective seed.

Copied to clipboard

Pietro Ferrazzi, Alberto Lavelli, Bernardo Magnini

Peek2: Regex-free Byte-level Byte-Pair Encoding Pretokenizer for LLM Inference on Edge Devices (2026.acl-srw)

Challenge:	Adapting Large Language Models to the medical domain remains an active area of research .
Approach:	They propose to compare three common adaptation approaches to adapt large language models to the medical domain.
Outcome:	The proposed models are built on top of foundational LLMs and rely on different post-training methodologies for domain and task performance.

Copied to clipboard

Liu Zai, Iraklis A. Klampanos

Syntactic Text Processing

Challenge:	Pretokenization is a crucial, sequential pass in Byte-level BPE tokenizers . little work has been done to optimize it for edge-side inference .
Approach:	They propose a pretokenization algorithm with linear time complexity and trivial memory usage that is suited for edge scenarios.
Outcome:	The proposed pretokenization algorithm improves microbenchmarking throughput by 2.48 and delivers 1.14 improvement in overall throughput across the entire Byte-level BPE encoding process.

Annotation Entropy Predicts Per-Example Learning Dynamics in LoRA Fine-Tuning (2026.acl-srw)

Copied to clipboard

Brady Steele

Semantic Contrastive Adaptation for Multimodal Figurative Language Understanding (2026.acl-srw)

Challenge:	Annotator disagreement on tasks like natural languageinference (NLI) reflects genuine linguistic ambiguity, but most fine-tuning recipes treat every example as equallylearnable.
Approach:	They ask whether annotator disagreement on tasks like natural languageinference (NLI) reflects genuine linguistic ambiguity.
Outcome:	The proposed method predicts learning dynamics on contested examples under LoRA.

Copied to clipboard

Ayaan Siddiqui

Think Less, Code Better: Probing When Chain-of-Thought Hurts and How to Route Around It (2026.acl-srw)

Challenge:	Existing models for understanding figurative language in images perform well on literal recognition but fail on multimodal figurativ benchmarks.
Approach:	They propose a model that adapts to idiomatic and figurative language using literal alignment bias rather than limited model capacity.
Outcome:	The proposed model generalizes across five idiom-rich languages despite being trained on English supervision.

Copied to clipboard

Rajarshi Ghoshal, Debadri Basak, Salma Emad Mahmoud Abdelhalim, Pratibha Kaur Arora

Probing Functional Correctness in Diffusion Language Models (2026.acl-srw)

Challenge:	Chain-of-Thought (CoT) prompting is the dominant strategy for eliciting step-by-step reasoning in large language models, but its effect on code generation is poorly understood.
Approach:	They develop a chain-of-thought (CoT) prompting router that selects among 12 prompt styles via a single 84 ms forward pass.
Outcome:	The proposed model outperforms CoT in small models with a 84 ms forward pass.

Copied to clipboard

Guan-Ming Chiu, Jeng-Yue Liu

Thesis Proposal: On the Granularity-Robustness Trade-off in Text-Derived Knowledge Graphs (2026.acl-srw)

Challenge:	Diffusion language models generate text by iteratively denoising all tokens in parallel, but when and where hidden states encode whether output will be functionally correct remains unknown.
Approach:	They present the first probing study of Diffusion language models to train classifiers on hidden states to predict functional correctness.
Outcome:	The proposed model generates all tokens simultaneously, but when and where does it encode correctness?

Copied to clipboard

Surawat Pralomram

TokLens: A Multilingual Lens on Tokenizer Quality for LLMs (2026.acl-srw)

Challenge:	Retrieval-augmented generation (RAG) based on dense embeddings is a dominant paradigm for text retrieval, but many real-world applications require attribute-specific querying.
Approach:	They propose a query-driven framework for constructing and retrieving knowledge graphs from text using dense embeddings.
Outcome:	The proposed framework combines the robustness of dense retrieval with the explicit queryability of symbolic representations.

Copied to clipboard

Guan-Ming Chiu

Phase Transitions in Affective Meaning Divergence: The Hidden Drift Before the Break (2026.acl-srw)

Challenge:	TokLens is an open-source toolkit for evaluating tokenizer quality across languages . authors evaluated 24 tokenizers from major LLM families across 15 typologically diverse languages - a gap that is stark in Japanese .
Approach:	They evaluate 24 tokenizers from major LLM families across 15 typologically diverse languages and correlate these metrics with downstream performance.
Outcome:	The proposed tokenizers produce 56x more tokens per word in Japanese than in English . the newer tokenizer Qwen2.5 and Gemma-2 reduce this gap to under 4x .

Copied to clipboard

Napassorn Litchiowong

Linguistic Theories

Challenge:	a couple's locutionary act is identical, but the illocutionary force received diverges from the force intended.
Approach:	They propose a logit best-response map with a hysteretic collapse of repair coordination . they propose lexical divergence variance, AMD variance, dialog-act repair variance .
Outcome:	The proposed model shows that derailing conversations exhibit critical-slowing-down signatures across multiple levels.

Sycophantic Anchors: Localizing and Quantifying User Agreement in Reasoning Models (2026.acl-srw)

Copied to clipboard

Jacek Duszenko, Przemyslaw Kazienko, Jan Kocon

NEAT-IR: Neural Explainable Analysis Tool for Information Retrieval (2026.acl-srw)

Challenge:	sycophancy is a behavior that infiltrates the chain-of-thought, leading models to generate plausible-sounding justifications for incorrect answers.
Approach:	They introduce sycophantic anchors that commit models to user agreement . they find scophancy leaves a stronger mechanistic footprint than correct reasoning .
Outcome:	The proposed framework outperforms text-only baselines at high commitment levels and predicts commitment strength from activations.

Copied to clipboard

Lev Sukherman, Artem Frenk, Nina Klimenkova, Connor Jason

Explainability & Interpretability in NLP

Challenge:	Neural IR models achieve strong performance but remain difficult to interpret.
Approach:	They propose a black-box analysis framework that explains ColBERT’s ranking behavior using 26 classical IR features.
Outcome:	The proposed framework preserves ColBERT’s rankings nearly perfectly, yet only explain R2 0.28 of score variance.

BanglaSocialBench: A Benchmark for Evaluating Sociopragmatic and Cultural Alignment of LLMs in Bangladeshi Social Interaction (2026.acl-srw)

Copied to clipboard

Tanvir Ahmed Sijan, S. M. Golam Rifat, Pankaj Partha, Md. Tanjeed Islam, Md Musfique Anwar

Challenge:	Large Language Models have demonstrated strong multilingual fluency, but not socially appropriate language use is guaranteed.
Approach:	They propose a benchmark to evaluate sociopragmatic competence in Bangla through context-dependent language use rather than factual recall.
Outcome:	The proposed benchmark evaluates sociopragmatic competence in Bangla through context-dependent language use rather than factual recall.

Interpretability of LLM Classifiers via the Rational Inattention Theory with Application to Hate Speech Detection (2026.acl-srw)

Copied to clipboard

Yuan Zhao, Ali Abdi

The Shape of Vulnerability: How Adversarial Perturbations Reshape the Topology of Language Model Latent Spaces (2026.acl-srw)

Challenge:	Large language models (LLMs) perform well on text classification, but their decision strategies need to be better understood.
Approach:	They propose an extended rational inattention model that parameterizes linguistic noise and information processing cost and provides an interpretable behavioral framework for black-box LLM classifiers.
Outcome:	The proposed model provides an interpretable behavioral framework for black-box LLM classifiers.

Copied to clipboard

Angelina Tsai, Shreya Subramanian, Catherine Liu, Kimberly Lopez, Leif Zinn-Brooks, Alexia E. Schulz, Adaku Uchendu

Robustness in NLP

Challenge:	Large Language Models (LLMs) have unprecedented capabilities, but they pose security concerns . current adversarial attacks exploit vulnerabilities in the embedding space of language models, allowing attackers to bypass safety guardrails and cause significant harmful consequences.
Approach:	They propose to use topological data analysis to characterize how adversarial perturbations act on text inputs by computing persistent homology metrics from attention maps across different model architectures.
Outcome:	The proposed visualizations show that adversarial perturbations alter higher-dimensional topological features in ways that distinguish them from clean, non-adversarial inputs.

LLM-based Literal Example Generation for Japanese Multiword Expressions (2026.acl-srw)

Copied to clipboard

Mio Ohashi, Hajime Kiyama, Zhidong Ling, Mamoru Komachi

Presentation Slide Translation and Layout Error Correction by LLMs (2026.acl-srw)

Challenge:	Existing work on Japanese multiword expressions has focused on detecting idiomatic usages in context, leaving literal readings underrepresented.
Approach:	They propose to use corpus non-literal usages as contrastive cues for controlled prompting . they compare their results to a test that compares model predictions with human judgments .
Outcome:	The proposed model provides more accurate literal examples than prompts that include no hints or literal information.

Copied to clipboard

Futo Kajita, Nobuyori Nishimura, Takehito Utsuro, Naoki Muto, Chee Siang Leow, Hiromitsu Nishizaki

Machine Translation

Challenge:	Existing translation tools suffer from layout errors due to text expansion during translation . a new approach to translating Japanese slides into English is proposed to overcome this issue .
Approach:	They propose a framework to translate Japanese slides into English and correct layout errors by using multimodal LLMs with slide images and XML structures.
Outcome:	The proposed method outperforms baselines and achieves 4.1% layout error rate and over 80% success rate.

Constructing a Japanese Rap Lyric Generation Model with GRPO (2026.acl-srw)

Copied to clipboard

Hayato Ogawa, Daisuke Kawahara

Challenge:	Rap is a vocal style rooted in Hip-Hop culture, characterized by producing rhymes in synchrony with a rhythmic beat.
Approach:	They propose a method for generating Japanese rap lyrics with a large language model . the model's rhyming behavior is improved by using existing Japanese rhapsodysts as training data.
Outcome:	The proposed method improves outputs that receive moderate or high human ratings on rhyme-related criteria.

Tracking the Evolution of Foresight Signals in News Data: The Case of the European Electric Vehicle Market (2026.acl-srw)

Copied to clipboard

Karine Navasartian

Challenge:	Recent studies have focused on the identification of signals in web-based textual data, but little attention has been paid to how they evolve over time and gain relevance as they become more visible.
Approach:	They examine whether tracking the temporal dynamics of signals can improve their assessment for strategic decision-making.
Outcome:	The proposed model shows that burst dynamics tend to surface signal consolidation rather than the early detection of weak signals.

Debiasing Logical Fallacy Detection for Real-World Robustness via Counterfactually Augmented Data (2026.acl-srw)

Copied to clipboard

Navyansh Singh

Bring Linguistics Back to Cryptanalysis: Using Attestation to Break the Advanced Encryption Standard (2026.acl-srw)

Challenge:	Logical fallacy detection models often over-flag valid reasoning due to spurious correlations.
Approach:	They propose to augment CAD models with counterfactually-augmented data to debias them . they found that the models often over-flag valid reasoning due to spurious correlations .
Outcome:	The proposed approach reduces false positive rate by 58% on a 300-sample set.

Copied to clipboard

Madeline Boese

Numerical Reasoning

Challenge:	Current DES cryptanalysis methods use mathematical methods to find the most likely approximate key in the entire probability space.
Approach:	They propose an algorithm that eliminates AES keys using linguistic attestation . the algorithm proposes trimming the probability space by removing unattested keys .
Outcome:	The proposed algorithm eliminates AES keys using linguistic attestation . the proposed algorithm can be analyzed in seconds with current DES algorithms .

Garden Path Recovery in Causal and Masked Language Models (2026.acl-srw)

Copied to clipboard

Sanjan Baitalik, Rajashik Datta

Confidence as a Tie-Breaker: Reassessing Multilingual Hedging Bias in LLM-as-a-Judge Evaluation (2026.acl-srw)

Challenge:	a linguistics study of garden-path sentences shows that recovery dynamics are important for linguistic evaluation . causal models show larger within-model disambiguation effects than masked models overall .
Approach:	They propose to compare garden-path recovery in causal and masked language models . they use 100 English garden- path/control pairs spanning three constructions .
Outcome:	The proposed model shows that decoder-only models exhibit sharper disruption at the point of syntactic revision, while encoders appear comparatively buffered at the disambiguator due to right-context access.

Copied to clipboard

Rajashik Datta, Sanjan Baitalik

BanglaSTEM: A Parallel Corpus and Term-Weighted Evaluation for Technical Bangla-English Translation (2026.acl-srw)

Challenge:	LLM judges are often used to score generated answers, but their decisions may be affected by surface style rather than semantic correctness.
Approach:	They propose a benchmark to examine multilingual hedging effects in LLM evaluation . they find that when two answers are equally correct, the judge prefers the assertive answer .
Outcome:	The proposed benchmark shows that judge preferences are based on style and style . the results highlight that benchmark auditing is a key requirement for judge-bias research .

Copied to clipboard

Kazi Reyazul Hasan, A. B. M. Alim Al Islam, Muhammad Abdullah Adnan

Disentangling the Effects of Unlearning in Measuring Parametric Faithfulness of Chain-of-Thought (2026.acl-srw)

Challenge:	Large language models excel at technical problem solving in English but struggle when questions are posed in Bangla.
Approach:	They propose a dataset of 5,000 Bangla-English sentence pairs to align technical terms . they use OCR to extract matching passages from bilingual textbooks .
Outcome:	The proposed pipeline extracts matching passages from bilingual textbooks and uses them to align sentences and mark technical terms.

Copied to clipboard

Ryo Mitsuhashi, Gaku Morio, Ayana Niwa, Masahiro Kaneko, Kentaro Inui, Terufumi Morishita, Yuta Koreeda, Yasuhiro Sogawa

FedPAGR: Federated Prototype Alignment via Geometric Refinement for Heterogeneous Architectures (2026.acl-srw)

Challenge:	Chain-of-Thought (CoT) has been debated as a model's faithfulness to internal reasoning process.
Approach:	They propose to use unlearning to measure parametric faithfulness of models by adjusting for unintended artifacts of unlearning.
Outcome:	The proposed metric accounts for the unintended artifacts of unlearning and shows that it is non-negligible.

Copied to clipboard

Kris Prasad, Md Abdullah Al Hafiz Khan

Sentiment Analysis of Yelp Review Dataset: A Comparative Study of Machine Learning Methods (2026.acl-srw)

Challenge:	Federated learning with heterogeneous client architectures can be difficult due to semantic drift and poor inter-class separation.
Approach:	They propose a framework where heterogeneous clients exchange class prototypes with a central server and refine them through a geometric regularization objective.
Outcome:	The proposed framework achieves highest ensemble accuracy across all four image datasets and highest local test accuracy on low-class and clinical tasks.

Copied to clipboard

Krishna Thakar, Mohamed Abu Sheha, Emmanuel Thompson

Sentiment Analysis

Challenge:	Existing methods for sentiment analysis are inconsistent and require manual processing.
Approach:	They use natural language processing and machine learning to classify Yelp reviews' sentiments.
Outcome:	The proposed model outperforms other models on Yelp reviews.

Semantic Span Annotation: An Exploratory Study of LLM Annotation (2026.acl-srw)

Copied to clipboard

Tejas Goyal, Dhriti Krishnan, Anuj Gupta, Jaromir Savelka

Thesis Proposal: An Explainable Multimodal Framework for Detecting Harmful Content in Code-Switched Children’s Media (2026.acl-srw)

Challenge:	Structured span extraction research is siloed by context length, annotation task, and domain . Identifying a span within a natural language text and affixing it with a semantic label has been considered a core task in NLP .
Approach:	They propose a framework for structured span annotation that integrates five datasets under a common JSONL format with character-level offsets.
Outcome:	The proposed framework can generalize across four domains under three prompting configurations.

Copied to clipboard

Juliana Isabelle A. Guillermo, Jasper Kyle Catapang, Nathaniel Oco

Test-Time Strategies for More Efficient and Accurate Agentic RAG (2026.acl-srw)

Challenge:	Current content moderation systems fail to protect children from harmful content, especially in under-resourced, code-switched settings.
Approach:	They propose to integrate a fine-tuned classifier with an LLM-powered module that synthesizes the classifier’s internal evidential signals to generate faithful, human-readable rationales for each decision.
Outcome:	The proposed framework integrates a fine-tuned classifier for accurate, scalable detection with an LLM-powered module that synthesizes the classifier’s internal evidential signals to generate faithful, human-readable rationales for each decision.

Copied to clipboard

Abhinav Sharma, Brian Zhang, Deepti Guntur, Zhiyang Zuo, Shreyas Chaudhari, Wenlong Zhao, Franck Dernoncourt, Puneet Mathur, Ryan A. Rossi, Nedim Lipka

Eye Movement Features Can Predict Human Preferences on Machine-Generated Texts (2026.acl-srw)

Challenge:	Retrieval-Augmented Generation (RAG) systems face challenges with complex, multi-hop questions.
Approach:	They propose to integrate contextualization module and de-duplication module to improve the accuracy of retrieved documents and to reduce the number of turns by 10.5%.
Outcome:	The proposed approach achieves a 5.6% increase in EM score and reduces the average number of turns by 10.5% compared to the baseline.

Copied to clipboard

Xiaoshan He, Xiaoqun Liu, Haodong He, Yu Wang, Yang Xu

Text Generation

Challenge:	Existing studies on eye movement in text quality assessment are limited . eye-movement features are important predictors of human judgments of text quality, but are costly and inconsistent.
Approach:	They propose to capture eye-movement features during screen reading of LLM-generated text using a dataset that includes eye-motion recordings, reading-time measurements, and post-reading evaluations.
Outcome:	The proposed dataset shows that eye-movement features can significantly improve models over other probabilistic metrics, including negative log-likelihood (NLL).

Thesis Proposal: Diagnosing and Mitigating Semantic Interference in Script-Sharing Low-Resource Language Models: A Case Study on Square Bai Script (2026.acl-srw)

Copied to clipboard

Jingting Zheng, Deyi Xiong

Does Locality Cost in Polish Medical Text Classification? Duplicate-Aware Evaluation of Federated Learning (2026.acl-srw)

Challenge:	Multilingual models cover more languages than ever, yet script-sharing low-resource languages remain vulnerable to failures driven by script and dominant-language priors.
Approach:	They propose to diagnose semantic interference in Square Bai Script, a script-sharing low-resource language . they will compare adaptation strategies under realistic compute constraints and estimate when shared-script transfer helps or harms adaptation.
Outcome:	The proposed model underperforms in a script-sharing low-resource setting because Bai is low-source and because it treats visible overlap as safe transfer by default.

Copied to clipboard

Daniel Cieślak, Andrzej Czyżewski

Analyzing Hate Speech Amplification on Fringe Platforms (2026.acl-srw)

Challenge:	a study of Polish medical text classification shows federated learning is a practical trade-off . centralized training is often framed as a governance concession rather than a genuinely competitive learning protocol.
Approach:	a study of Polish medical text classification shows federated learning is a practical trade-off . authors argue that centralized training is generally superior to centralized learning . they also argue that the results are biased by the granularity of evaluations .
Outcome:	a new study compares federated and centralized training in a duplicate-heavy medical text benchmark in Poland . a similar study shows that federation outperforms centralized learning in the strongest setting .

Copied to clipboard

Anika Ghosh Basu, Humberto Jesus Carlon, Junyi Liu

The Silence of the Facts: Popularity as a Barrier to Machine Unlearning (2026.acl-srw)

Challenge:	a new study examines the factors that determine how hate speech amplifies on fringe platforms . fringe platforms like Gab harbor high volumes of hate speech due to minimal moderation and insular communities .
Approach:	They used a dataset of 5K+ threads and 50K+ responses from four fringe platforms . they found that thread structure and disagreements in early response windows can give up to 74% lift in RMSE .
Outcome:	The proposed model estimates how several features influence hate speech amplification on fringe platforms.

Copied to clipboard

Anna Borisiuk, Andrey Savchenko, Alexander Panchenko, Elena Tutubalina

Challenge:	Existing unlearning methods assume that all facts are equally challenging to forget . large models struggle more to forget popular entities, damaging related knowledge in the process .
Approach:	They build a benchmark to investigate whether fact popularity influences the efficiency of LLM unlearning.
Outcome:	The proposed benchmark compares state-of-the-art models on a set of models of different sizes.

Leakage-Aware User-Level ADHD Signal Classification from Social Media: When Graph Aggregation Helps, and When It Does Not (2026.acl-srw)

Copied to clipboard

Daniel Cieślak, Władysław Średniawa

CAL-Log: Cost-Aware Active Learning with Logarithmic Cognitive Effort Modeling and Online Adaptation to Human Annotation Behavior (2026.acl-srw)

Challenge:	Social media data are longitudinal, usercentered, rich in spontaneous language use.
Approach:	They propose a leakage-aware evaluation framework organized around two controlled axes: evidence budget and leakage control.
Outcome:	The proposed framework compares graph aggregation with other models using psycholinguistic features and semantic tweet embeddings.

Copied to clipboard

Vihanga Supasan Kariyakaranage, Banuka Athuraliya

Text Classification

Challenge:	Standard uncertainty sampling assumes that annotating a 500-word document requires the same effort as a 50-word tweet, leading to suboptimal resource allocation when documents vary in length.
Approach:	They propose a cost-aware AL variant using logarithmic cost modeling where C(x) is the predicted annotation time for document x and L(x), is its token length.
Outcome:	Experiments on ten text classification benchmarks show a 3.3 speedup over BADGE and 3.9 over Entropy sampling to reach F1=0.80, with large effect sizes.

Thesis Proposal: Targeted and Unified Cross-Lingual Unlearning from Multilingual Language Models (2026.acl-srw)

Copied to clipboard

Jan Bronec, Jindřich Helcl

A11y-Compressor: A Framework for Enhancing the Efficiency of GUI Agent Observations through Visual Context Reconstruction and Redundancy Reduction (2026.acl-srw)

Challenge:	Large language models trained on corpora scraped from the web can reproduce sensitive and copyright-protected data.
Approach:	They propose to extend existing benchmarks to multilingual data by compiling parallel translations of question-answer pairs consisting of real-world facts and synthetic personally identifiable information.
Outcome:	The proposed dataset will include translations of question-answer pairs consisting of real-world facts and synthetic personally identifiable information.

Copied to clipboard

Michito Takeshita, Takuro Kawada, Takumi Ohashi, Shunsuke Kitada, Hitoshi Iyatomi

Counterspeech Generation using Small Language Models (2026.acl-srw)

Challenge:	Existing approaches to grounding GUI environments are categorized into image-based and text-based representations.
Approach:	They propose a framework that transforms linearized accessibility trees into compact and structured representations.
Outcome:	The proposed framework reduces input tokens to 22% of the original while improving task success rates by 5.1 percentage points on average.

Copied to clipboard

Abubakar Sadiq Muhammad, Simona Frenda, Gavin Abercrombie

From Graphs to Hypergraphs: Enhancing Aspect-Term Sentiment Analysis via Multi-Level Relational Modeling (2026.acl-srw)

Challenge:	Social media use is growing annually with about 68.5% of the global population active on these platforms as of July 2025.
Approach:	They evaluate SLMs ranging from 100 million to 3 billion parameters using simple prompting strategies as well as fine-tuning, combining automatic and robust human evaluations.
Outcome:	The proposed models generate relevant, coherent, and high-quality counterspeech, suggesting their suitability for efficient and responsible deployments.

Copied to clipboard

Omkar Mahesh Kashyap, Padegal Amit, Madhav Kashyap, Ashwini M Joshi, Shylaja S S

Sentiment Analysis

Challenge:	Existing graph-based approaches to predict sentiment polarity for specific aspect terms rely on predefined pairwise structures to improve expressive capacity.
Approach:	They propose a dynamic hypergraph framework that can be used to generate a single instance-specific hypergraph from contextual token representations.
Outcome:	The proposed framework improves on Lap14, Rest14, and MAMS . it uses a single instance-specific hypergraph constructed directly from contextual token representations .

Probing Bias Formation in Medical LLMs through Activation Steering (2026.acl-srw)

Copied to clipboard

Bayram Ayadi, Annette Hautli-Janisz

Faithfulness Beyond Plausibility: Auditing Human Explanations in Educational Assessment (2026.acl-srw)

Challenge:	Large Language Models specialized for the medical domain achieve high performance on static benchmarks, but are vulnerable to sycophantic confabulation.
Approach:	They propose a framework toward clinical AI systems that are more robust and aligned with expert medical logic.
Outcome:	The proposed framework outperforms static global interventions on a medical prompt with cluster-conditioned dynamic steering.

Copied to clipboard

Ria Talsania, Dhruv Ritesh Shah, Sudhir Dhage

Explainability & Interpretability in NLP

Challenge:	a gap exists between explanation components and how scores are constructed, and whether they reflect how scores were constructed . authors: explanation components are structurally inconsistent and may not be used as post-hoc justifications .
Approach:	They propose to use human tutor grading traces to test whether explanations are reliable . they find that removing rubric-level information leads to substantial changes in reconstructed scores .
Outcome:	The proposed diagnostic measures how explanation components contribute to score interpretation . removing rubric-level information leads to substantial changes in reconstructed scores .

CBAL: Context-Based Agentic Learning for Speaker Diarization Segmentation Refinement (2026.acl-srw)

Copied to clipboard

Odwitiyo Dutta, Dinesh K Vishwakarma

Measuring and Mitigating Shortcut Reliance in Language Models with Probe-Based Representation Entanglement (2026.acl-srw)

Challenge:	Speaker diarization systems produce segmentation errors that degrade transcript readability and downstream applications.
Approach:	They propose a framework that refines segmentation boundaries in diarized scripts . they use a lightweight LLM agent to reason about merge decisions .
Outcome:	The proposed framework achieves 93.4% accuracy across 359 applied merges and reduces segment count by 6.1%.

Copied to clipboard

Divyajot Singh

LAMP-MedQA: A Lightweight Multi-Agent System for Patient-Oriented Medical Question Answering (2026.acl-srw)

Challenge:	Shortcut learning remains a major obstacle to robust NLP systems.
Approach:	They propose to fine-tune Gemma 3 1B Instruct and Llama 3.2 1B on two synthetic sentiment shortcuts in SST-2 and one natural shortcut in MNLI based on lexical overlap.
Outcome:	The proposed model improves on two synthetic sentiment shortcuts and one natural shortcut in MNLI with a 99% shortcut ratio, while Gemma drops from 91.8% to 60.2%.

Copied to clipboard

Jack A. Johnson, Meghali Banerjee, Joseph Crawford, James Welch, Jim Davies, Tingyan Wang

Inference-Time Feedback for Reasoning Controllability in Diffusion Language Models (2026.acl-srw)

Challenge:	Large language models (LLMs) are a promising way to bridge the gap between patient health literacy and access to care.
Approach:	They evaluate a range of open- and closed-source LLMs on a MeDiSumQA dataset . they propose a lightweight multi-agent framework for patient-oriented medical question answering .
Outcome:	The proposed model achieves lower FKGL than zero-shot GPT-5 and highest simplification quality among all models.

Copied to clipboard

Clovis Barbour, Huixin Zhan

Disentangling Linguistic Relatedness from Task Alignment in Cross-Lingual Transfer (2026.acl-srw)

Challenge:	In scientific NLP systems, model outputs often serve as interfaces to downstream systems that assume strict structural requirements.
Approach:	They evaluate machine-checkable controllability along three axes: structural control, iterative correction, and decoding dynamics.
Outcome:	The proposed model can be usefully decomposed into global structure versus local control . the proposed model improves global structure while improving iterative correction .

Copied to clipboard

Ahmed Haj Ahmed, Ruochen Zhang, Alvin C Grissom II

Cross-Lingual Transfer

Challenge:	Large language models (LLMs) have advanced natural language processing, yet their benefits remain concentrated in English and a small number of high-resource languages.
Approach:	They fine-tuned large language models (4B–671B parameters) on Arabic and evaluated zero-shot reading comprehension on Semitic languages and non-Semitic controls.
Outcome:	The results show that models with weak baselines improve across all languages, whereas strong-baseline models show only marginal gains regardless of language family.

PE-QAT: Parameter-Efficient Quantization-Aware Training for Large Language Models (2026.acl-srw)

Copied to clipboard

Shresth Mishra

Fusion Training for Mathematical Generalization in Large Language Models (2026.acl-srw)

Challenge:	Quantization Aware Training (QAT) is expensive to train and unscalable to large models.
Approach:	They propose a parameter-efficient framework targeting per-channel 4-bit weight-activation quantization of large language models.
Outcome:	The proposed framework preserves accuracy within 0.11 percentage points of the full-precision baseline on Llama-2-7B zero-shot tasks while training only 1.26% of total parameters.

Copied to clipboard

Congfeng Cao, Pengyu Zhang, Jelke Bloem

Does Topic Sentiment Cause Perceived Ideology? Comparing Human and LLM Annotations in Political News Articles (2026.acl-srw)

Challenge:	Existing efforts to improve reasoning efficiency have focused on switching between simple queries and complex problems.
Approach:	They analyze the effects of the training schedule and data ratio between thinking and non-thinking modes and construct a benchmark to test their theory.
Outcome:	The proposed model unifies a thinking mode and a non-thinking mode within a single model.

Copied to clipboard

Upasana Chatterjee

Understanding Conversational Implicatures in Humans and LLMs (2026.acl-srw)

Challenge:	Existing validation studies assess output-level agreement; none test causal structure of LLM decisions mirrors that of human decisions.
Approach:	They compare topic sentiment and ideology labels using human annotators . they say this is evidence of shortcut learning by fine-tuning on ideology-labeled data .
Outcome:	The proposed model internalises a spurious sentiment–ideology coupling not operative in human judgment for this task.

Copied to clipboard

Daeun Kang

Dialogue Systems & Conversational Agents

Challenge:	Large Language Models (LLMs) interpret conversational implicatures using humans as a baseline . et al.: do LLMs exhibit a human-like sensitivity to pragmatic inference?
Approach:	They adopt a surprisal-based and response-based metric to measure the accuracy of implicatures . they find that LLMs employing the response- based meter exhibit human-like behavior .
Outcome:	The proposed model performs better in the literal condition than in the implied condition . the model differs from humans in its understanding of conversational implicatures .

Thesis Proposal: Self-Adaptive and Epistemic Uncertainty-Guided ASR of Dense Intra-Sentential Code-Switched Speech for African Low-Resource Languages (2026.acl-srw)

Copied to clipboard

Umar Baba Umar, Sulaimon Adebayo Bashir, Abdulmalik Danlami Mohammed, Amina Gogo Tafida

Speech & Audio in NLP

Challenge:	Existing multilingual and pretrained ASR systems improve general recognition accuracy but are weak at switch regions and are sensitive to language imbalance during adaptation.
Approach:	They propose a self-adaptive and epistemic uncertainty-guided framework for African low-resource code-switched ASR using Hausa–English and Hausa-Yorùbá as case studies.
Outcome:	The proposed framework is based on Hausa–English and Hausa-Yorùbá as case studies.

RegTrack: A Fine-Grained Benchmark for Multi-Class Legal Change Detection (2026.acl-srw)

Copied to clipboard

Joe Yu, Kevin Chenhao Li, Julian Ostarek

Validator-Guided Hard Negative Mining for Masked Language Modeling in Low-Resource Ancient Languages (2026.acl-srw)

Challenge:	Existing models lack finegrained classification schemes to determine whether small changes impact legal obligations or merely update formatting.
Approach:	They propose a benchmark for change detection in EU regulations that uses 4,772 manually annotated pairs of structurally distinct provisions mapped to a six-class taxonomy of legal change types.
Outcome:	The proposed framework combines lexical algorithms, dense encoders, and Large Language Models (LLMs) as baselines.

Copied to clipboard

Andrei Voinea

Conformal LLM Routing with Distribution-Free Safety Guarantees (2026.acl-srw)

Challenge:	Pre-trained multilingual models lack exposure to low-resource ancient languages such as Sumerian.
Approach:	They build a hierarchical validator capturing subword, word, and part-of-speech patterns from 4,545 annotated sequences and use it to filter candidates and identify hard negatives for fine-tuning.
Outcome:	The proposed model achieves 18.0% hit@10 accuracy, while the validator alone improves this to 72.8%, while hard negative fine-tuning reaches 78.3%.

Copied to clipboard

Iqtedar Uddin, André Bauer

Supervision versus Demonstration-Based In-Context Learning for Multiword Expression Classification (2026.acl-srw)

Challenge:	Existing input-based routers optimize cost-performance trade-offs but provide no formal bound on how often cheaper model fails . a feasibility analysis across all 10 RouterBench models reveals that routability is jointly model- and task-dependent.
Approach:	They adapt a proactive conformal gate framework to LLM routing to provide cost-performance trade-offs . they find that a third axis is missing: the routability of queries is model-dependent .
Outcome:	The proposed method maintains the target within the tolerance across two benchmarks . it is the first input-based LLM router with distribution-free safety guarantees .

Copied to clipboard

Sercan Karakaş, Yusuf Şimşek

When Models Hesitate: Answer Instability as a Label-Free Uncertainty Signal for LLMs (2026.acl-srw)

Challenge:	idiomatic light verb constructions (LVCs) are challenging for multiword expression processing . they share the same surface form as fully literal verb–object combinations .
Approach:	They frame Turkish LVC detection as a binary classification task . they compare a supervised Turkish encoder baseline to three instruction-tuned LLMs .
Outcome:	The proposed method improves Turkish LVC detection on a controlled set with matched negatives and positives.

Copied to clipboard

Jasper Meynard Arana, Kristine Ann M. Carandang, Ethan Robert Casin, Christian Alis, Christopher Monterola

Thesis Proposal: Toward a Human-Centered and Perspective-Aware Framework for Reproducible ML Evaluation and AI Alignment (2026.acl-srw)

Challenge:	Existing approaches to uncertainty estimation typically require access to internals, additional supervision, or computationally intensive pipelines.
Approach:	They propose to use a label-free uncertainty signal to predict the variability of a model's final answer across repeated stochastic generations of the same prompt to achieve performance competitive with semantic entropy.
Outcome:	The proposed method achieves performance competitive with semantic entropy while requiring no similarity model.

Copied to clipboard

Deepak Pandita, Christopher M. Homan

One Panel Does Not Fit All: Case-Adaptive Multi-Agent Deliberation for Clinical Prediction (2026.acl-srw)

Challenge:	Disagreement arises from subjective human opinion and can vary with one’s identity, beliefs, and social environment.
Approach:	They propose a human-centered framework for reproducible ML evaluation and AI alignment that takes disagreement into account when building human-centric AI systems.
Outcome:	The proposed framework is based on a human-centered and perspective-aware framework for reproducible ML evaluation and AI alignment.

Copied to clipboard

Yuxing Lu, Yushuhong Lin, Jason Zhang

Challenge:	Existing single-agent strategies sample from one role-conditioned distribution, and multi-agend frameworks use fixed roles with flat majority voting, discarding the diagnostic signal in disagreement.
Approach:	They propose a case-adaptive multi-agent panel where an attending-physician agent dynamically assembles a specialist panel tailored to each case’s diagnostic uncertainty.
Outcome:	The proposed model outperforms baseline models on diagnostic prediction and brief hospital course generation using MIMIC-IV.

LLMs for Now, Fine-Tuning for Later: An Ensemble Approach to Data Drift in Domain-Specific Tasks (2026.acl-srw)

Copied to clipboard

Yuxuan Lu, Bingsheng Yao, Shao Zhang, Yisi Sang, Yun Wang, Hansu Gu, Peng Zhang, Tun Lu, Toby Jia-Jun Li, Dakuo Wang

Thesis Proposal: When Does an Agent Know It Is Lost? Confidence Trajectory Analysis for Tool-Using LLMs (2026.acl-srw)

Challenge:	Deploying machine learning models in domain-specific scenarios is challenged by data drift and the scarcity of expert annotations.
Approach:	They propose a system that combines an LLM, an AL-assisted compact model and an automatic switch module to assist the active learning process.
Outcome:	The proposed system achieves 96–98% switch accuracy and outperforms both models used alone.

Copied to clipboard

Zhenjiang Mao

Challenge:	Existing uncertainty quantification methods treat each step in isolation, ignoring how confidence evolves and compounds across a full task trajectory.
Approach:	They propose a framework for trajectory-level confidence analysis in the tool-use agent setting.
Outcome:	The proposed framework will expose early warning signals for agent failure and offer interpretable diagnostic tools for understanding when and why LLM agents lose confidence.

Task Assignment meets Annotator Modeling: Human-LLM Collaborative Annotation with Constraints (2026.acl-srw)

Copied to clipboard

Kei Moriyama, Kouta Nakayama, Yukino Baba

Challenge:	Existing approaches to label annotation are labor-intensive and time-consuming.
Approach:	They propose a framework that estimates per-task accuracy from task features using a learning from crowds model and incorporates these estimations into a linear programming formulation that assigns tasks under practical constraints.
Outcome:	The proposed method achieves comparable accuracy to baseline methods while satisfying given constraints.

Dynamic Meta-Metrics: Source-Sentence Conditioned Weighting for MT Evaluation (2026.acl-srw)

Copied to clipboard

Luke Zhang, Justin Vasselli, Aditya Khan, York Hay Ng, En-Shiun Annie Lee

Tonal Salience in Cognitive Decline: In-Context MCI Detection with Multimodal LLMs (2026.acl-srw)

Challenge:	Rather than relying on a single static ensemble or language-specific weighting, DMM adapts the metric combination based on properties of the source segment.
Approach:	They propose a framework for machine translation evaluation that learns source-sentence conditioned combinations of existing metrics.
Outcome:	The proposed framework outperforms linear and Gaussian process-based ensembles across multiple language pairs and introducing soft conditioning yields gains over linear models.

Copied to clipboard

Christopher Song, Abdullah P. Rashed Ahmed

What Moves the Pareto Frontier in Tool-Using Agents? A Compute-Aware Study of ReAct Variants (2026.acl-srw)

Challenge:	a recent study shows that tonal languages like Chinese have a higher classification performance than non-tonal languages like English.
Approach:	a new study examines the differences between tonal and non-tonal language classifications . they hypothesize that the difference is rooted in language typology . early cognitive decline is notoriously difficult to detect .
Outcome:	The proposed method compared to TAUKADIAL audio shows that Chinese and English perform better on Chinese . the findings suggest that language typology should inform the design of audio-based cognitive screening tools .

Copied to clipboard

Rishi N. Simhadri

Challenge:	Tool-using LLM agents are typically compared by accuracy alone, despite deployments being constrained by inference cost.
Approach:	They propose a budgeted evaluation of common strategies for improving ReAct-style tool agents . they compare cumulative accuracy versus token budget on three benchmarks .
Outcome:	The proposed methods show that the most common improvement strategies increase inference cost.

Filling the Long Tail: Structure-Aware Curriculum-Gap Completion for Medical Education with LLMs (2026.acl-srw)

Copied to clipboard

Wenjie Lin

Mechanistic Analysis Of Universality: Numerical Comparison Circuits Across Transformer Architectures (2026.acl-srw)

Challenge:	Medical education resources are dense for common diseases, but sparse for under-covered conditions, atypical presentations, and fine-grained concept distinctions.
Approach:	They propose a task where a model reconstructs missing educational units from a partially specified curriculum graph.
Outcome:	The proposed model predicts omitted concepts, restores missing instructional links, and completes automatically verifiable teaching content.

Copied to clipboard

Arya Bhardia, Julian Ramirez, Siddhanta Verma, Karen Mkrtchyan

How Hard is Math? Using Quantitative Metrics to Measure LLM Alignment to Human Intuitions of Difficulty (2026.acl-srw)

Challenge:	Mechanistic interpretability seeks to identify internal circuits within transformer language models but it is unclear whether they generalize across model families and scales.
Approach:	They propose to identify internal circuits within transformer language models by numerical comparisons.
Outcome:	The proposed model implementations are consistent across architecture and scale, the authors show . their results highlight the need for cross model comparisons to claim generalization of internal circuits.

Copied to clipboard

Micah Helzerman, Steven R Wilson, Cam McLeman

Fine-Grained Semantic Comparison of Legal Documents using LLMs (2026.acl-srw)

Challenge:	Often overlooked is how "difficulty" is operationalized in the context of LLM problem solving tasks.
Approach:	They examine quantitative measurements from LLM-generated solutions and evaluate their inter-correlations with human-annotated difficulty scores.
Outcome:	The proposed model shows that LLMs can solve problems with reasonable accuracy, but performance is poor when generalizing to other benchmarks.

Copied to clipboard

Elisei Rykov, Nikolay Ivanov, Maria Bandulevich, Kseniia Petrushina, Valentin Malykh, Vasily Konovalov, Alexander Panchenko, Ilseyar Alimova

Challenge:	Existing tools for detecting inconsistencies and contradictions in complex regulatory documents rely on character-level diffs.
Approach:	They propose a benchmark to evaluate span-aware semantic comparison of legal documents . legDiff is an annotated pair of legal paragraphs that is automatically generated .
Outcome:	The proposed benchmark evaluates span-aware semantic comparisons of legal documents . it generates synthetic training data that aligns with the manual annotations and mirrors the structure and label distribution of the benchmark .

From Fluent to Useful: Generative AI That Models Purpose, Audience, and Presenter for Scientific Communication (2026.acl-srw)

Copied to clipboard

Ishani Mondal

Challenge:	Modern generative AI produces fluent text,polished slides, and clean diagrams, yet fails when an artifact must serve a specificpurpose for a particular reader, used by aspecificpresenter.
Approach:	They propose to use persona-aware slide generation to help audiences reframe content rather than blanketsimplification to SMART-Editor to model user preferences.
Outcome:	The proposed system is based on persona-aware slide generation, sciDoc2Diagrammer-MAF, and SMART-Editor.

Beyond Discrete Search: Divergent Thinking as Intention Optimization in Latent Space (2026.acl-srw)

Copied to clipboard

Mateusz Bystroński, Grzegorz Piotrowski, Tomasz Jan Kajdanowicz

Information Retrieval

Challenge:	Despite rapid progress in LLM-based code generation, a persistent gap remains between what models can solve and what they solve on a given attempt.
Approach:	They propose a framework that recasts coding as optimization overconditioning contexts that influence the generation of natural-languagesolution intentions.
Outcome:	The proposed framework raises resolution rate of weak, quantized 24B open-weight model to parity with frontier models +25 its size.

Boosting Self-Consistency with Ranking (2026.acl-srw)

Copied to clipboard

Maria Marina, Daniil Moskovskiy, Sergey Pletenev, Mikhail Salnikov, Alexander Panchenko, Viktor Moskvoretskii

LLM-Based Zero-Shot Soft Labeling for Anticipating Disagreement in Negotiation Dialogues (2026.acl-srw)

Challenge:	Existing approaches to improve performance of large language models include self-consistency, RISC, extended reasoning, and iterative self-correction.
Approach:	They propose a test-time scaling technique that uses multiple features to score candidate answers in self-consistency as a ranking problem.
Outcome:	The proposed method achieves better accuracy-efficiency trade-off than standard self-consistency and strong baselines on three datasets.

Copied to clipboard

Ken Watanabe, Katsuhide Fujita

Dialogue Systems & Conversational Agents

Challenge:	Negotiation involves complex emotional and strategic dynamics that pose challenges for AI agents in negotiation dialogues.
Approach:	They propose a zero-shot soft-labeling method using large language models . they also examine the performance of model training on rule-based annotated hard and soft labels .
Outcome:	The proposed method shows a maximum HIT@3 score of 0.87 against rule-based annotated hard labels . failure cases also demonstrated the limitations of rule--based annotation .

Analysis of the Neglect-Zero Effect in Large Language Models (2026.acl-srw)

Copied to clipboard

Jin Tanaka, Daiki Matsuoka, Ryoma Kumon, Hitomi Yanaka

Morphology-Aware Multi-Granularity Representation Learning for Agglutinative Languages (2026.acl-srw)

Challenge:	a neglect-zero effect is a human cognitive bias in language processing . it is unclear whether LLMs also exhibit this effect .
Approach:	They focus on a human cognitive bias called the neglect-zero effect . they propose a paradigm where exposure to a preceding sentence facilitates processing of a subsequent sentence due to their similarity.
Outcome:	The proposed paradigm based on priming facilitates processing of a subsequent sentence due to their similarity to the target.

Copied to clipboard

Zhonghao Zhang, Na Liu, Jiajia Ma, Nier Wu, Guiping Liu

Syntactic Text Processing

Challenge:	Existing methods for learning low-resource agglutinative languages are limited to word and phrase levels.
Approach:	They propose a morphology-aware gated multi-granularity pre-training framework for agglutinative languages . framework leverages morphological knowledge and integrates a word-level encoder to capture contextual semantics .
Outcome:	The proposed framework improves on Mongolian and Turkish agglutinative languages . it leverages morphological knowledge and integrates tagging and segmentation to build fine-grained representations .

Processing Inconsistency Predicts Language Competence: LLM Evaluation Without Answer Labels on Turkic Languages (2026.acl-srw)

Copied to clipboard

Ilya Galyukshev, Ilseyar Alimova

TableMBR: Minimum Bayes Risk Table Generation Based on Structural Consistency (2026.acl-srw)

Challenge:	Most languages lack labeled evaluation benchmarks for large language models (LLMs). Creating labeles requires native speakers, domain expertise, and answer annotation.
Approach:	They hypothesize that a model's internal processing signals correlate with its actual accuracy on a language . they extract over 25 processing features per model–language pair and test them .
Outcome:	The proposed model outperforms the model's English/Russian benchmark score on 11 instruction-tuned LLMs across 14 language–script varieties.

Copied to clipboard

Daiki Yoshida, Hiroyuki Deguchi, Yusuke Sakai, Hidetaka Kamigaito, Taro Watanabe

Through the Looking Glass of Multilingual AI: Contrasting Language- and Name Script-Dependent Ethnic Hierarchies in GPT and DeepSeek (2026.acl-srw)

Challenge:	Experimental results show TableMBR outperforms the baseline, achieving relative improvements of up to 15% in F1 on Rotowire and 23% in accuracy on LiveSum.
Approach:	They propose a text-to-table task that generates structured data from unstructured text . they propose 'tableMBR' that maintains structural consistency through minimum Bayes risk decoding .
Outcome:	The proposed method outperforms the baseline and achieves relative improvements in F1 and LiveSum.

Copied to clipboard

Annabella Sakunkoo, Jonathan Sakunkoo

Thesis Proposal: Auditing and Mitigating Demographic Bias in Multi-Stage Retrieval Systems for Criminal Justice Applications (2026.acl-srw)

Challenge:	a recent study found that large language models are biased overwhelmingly Anglocentric . a stereotype perceptual map is a framework for analyzing how ethnic groups are positioned along evaluative dimensions.
Approach:	They use a stereotype perceptual map to examine how ethnic groups are positioned along evaluative dimensions.
Outcome:	The stereotype perceptual map analyzes model behavior across languages, scripts, evaluative domains and models.

Copied to clipboard

Archan Dutta

Contextual Diversity Measure (CDM) for Controllable Story Generation in Large Language Models (2026.acl-srw)

Challenge:	racial descriptors alter embedding similarity scores and retrieval rankings, a new study shows . rife-specific biases can displace relevant records outside top-10 results, the study concludes .
Approach:	They propose to detect, measure, and mitigate racial bias in NLP systems deployed in criminal justice contexts . they propose to develop and evaluate debiasing techniques, validate synthetic findings on authentic law enforcement data .
Outcome:	The proposed research examines how bias propagates across retrieval pipelines . it shows that racial descriptors alter embedding similarity scores and retrieval rankings .

Copied to clipboard

Richard Susilo, Hanna Suominen, Patrik Haslum

Constructing a Japanese Verdict Prediction Dataset for Fact-Checking of LLM-Generated Texts (2026.acl-srw)

Challenge:	Existing studies on controllable text generation focus on controlling attributes such as sentiment, writing style, and writing style.
Approach:	They introduce a metric that quantifies semantic diversity for scenario generation under fixed abstract semantic constraints and validate it through controlled experiments.
Outcome:	The proposed metric achieves excellent discrimination accuracy (100% and 91.9%, respectively), with discriminative power up to 5.5 greater than the best baseline.

Copied to clipboard

Miwa Masano, Hirokazu Kiyomaru, Atsushi Keyaki, Kaito Horio, Rei Minamoto, Ribeka Keyaki, Kouta Nakayama, Hideyuki Tachibana, Daisuke Kawahara

Calibration vs Decision Making: Revisiting the Reliability Paradox in Unlearned Language Models (2026.acl-srw)

Challenge:	Text generated by Large Language Models (LLMs) may contain plausible but incorrect information known as hallucinations.
Approach:	They extend the label set for verdict prediction to capture claim-evidence relationships humans would commonly interpret as supported or refuted.
Outcome:	The proposed system improves F1 by 4 percentage points compared to baseline.

Copied to clipboard

Divyaksh Shukla, Ashutosh Modi

Disentangling Meaning and Language Components in Diverse Multilingual Sentence Embeddings (2026.acl-srw)

Challenge:	Model calibration is commonly used as a proxy for reliability, but low calibration error does not necessarily imply reliable decision rules.
Approach:	They investigate the calibration error gap in generative language models using the TOFU benchmark and attribution-based shortcut detection.
Outcome:	The proposed model calibrations achieve low calibration error compared to pretrained models and retain low calibration despite reduced accuracy on the forget split.

Copied to clipboard

Kanade Nonomura, Keita Fukushima, Risa Kondo, Tomoyuki Kajiwara

Representation Learning

Challenge:	Existing studies have reported language specificity in multilingual sentence embeddings, resulting in language-specific subspaces.
Approach:	They propose to disentangle multilingual sentence embeddings into language-dependent and language-agnostic components to improve cross-lingual similarity estimation.
Outcome:	The proposed methods improve cross-lingual similarity estimation across multiple embeddings.

Linguistically-Informed Evaluation of LLMs on Acceptability Judgments in a Forced-Choice Paradigm (2026.acl-srw)

Copied to clipboard

Ziyue Liu, Nils Reiter

Representing Lean Proofs as Trajectories in Latent Space (2026.acl-srw)

Challenge:	Recent studies have reported close alignment with human judgments and systematic divergences driven by response biases and inconsistencies.
Approach:	They evaluate large language models' ability to perform acceptability judgments in a forced-choice paradigm.
Outcome:	The proposed models approximate human judgments but lack performance on morphosyntactic phenomena and prompt formulations.

Copied to clipboard

Elisaveta Samoylov, Soroush Vosoughi

Evaluation of Multilingual Ability to Use Spatial Deictic Expressions in Vision-Language Models (2026.acl-srw)

Challenge:	Lean proofs are built as sequences of tactic-induced state transitions, but learned models often represent proof steps through tactic strings or raw proof-state text.
Approach:	They train an encoder-only Transformer to learn contextualized representations of Lean proof steps from state changes.
Outcome:	The proposed model yields better held-out next-tactic retrieval than a surface-syntax control . the results provide a promising basis for future trajectory-aware theorem proving .

Copied to clipboard

Kaito Watanabe, Taisei Yamamoto, Tomoki Doi, Hitomi Yanaka

LLM Parameters for Math Across Languages: Shared or Separate? (2026.acl-srw)

Challenge:	Existing studies have focused on the ability of vision-language models to utilize spatial deictic expressions, which depend on the situation of utterance.
Approach:	They develop a benchmark to evaluate the multilingual ability of VLMs to use spatial deictic expressions in four languages.
Outcome:	The proposed models use demonstratives in a different manner from humans, particularly in selecting demonstrative based on distance from the object.

Copied to clipboard

Behzad Shomali, Luisa Victor, Tim Selbach, Ali Hamza Bashir, David Berghaus, Joachim Koehler, Mehdi Ali, Markus Frey

Thesis Proposal: Intentional Inference for Insight Generation (2026.acl-srw)

Challenge:	Existing research on large language models (LLMs) has focused on performance or representational properties, but it remains unclear whether these differences reflect language-specific parameters or a shared mechanism.
Approach:	They propose to localize and compare model parameters that support mathematical reasoning across languages.
Outcome:	The proposed analysis shows that the model parameters in English and lower-resource languages exhibit partial cross-lingual overlap with systematic language-dependent differences.

Copied to clipboard

Kristýna Onderková

Reference-Free Schema Generation for Literature Review Tables via Multi-Faceted Rewards (2026.acl-srw)

Challenge:	Large language models excel at surface-level fluency but struggle with consistent logical inference beyond surface- level patterns.
Approach:	They propose to shift focus from surface-level generation to assumption-aware deeper inferences . authors argue that underspecification leads to unintentional assumptions .
Outcome:	The proposed model underspecification leads to unintentional assumptions, the authors argue . they also examine how to improve reasoning to enable deeper inferences, focusing on code generation and qualitative reasoning.

Copied to clipboard

Sinjoy Saha, Suman Saha, Mahfuza Farooque, Wenpeng Yin

Factual State Discovery Benchmark: Evaluating Fact Elicitation in Polish Tax Law (2026.acl-srw)

Challenge:	Literature review systems generate literature review tables by inferring schemas and values from documents.
Approach:	They propose to use schema generation as a reinforcement learning problem to determine which dimensions to compare a set of papers.
Outcome:	The proposed model improves over the untuned model across intrinsic, reference-based, and LLM-judge metrics and remains competitive with supervised fine-tune models at 5 the parameter count on structural and diversity dimensions.

Copied to clipboard

Mateusz Bystroński, Kamil Tagowski, Denis Janiak, Julia Farganus, Lukasz Augustyniak, Monika Kajdanowicz, Tomasz Jan Kajdanowicz

Challenge:	FSDBench is a benchmark for eliciting all relevant facts through dialogue . missing facts may lead the authority to apply the wrong provision or issue a ruling that is inapplicable to the actual situation.
Approach:	They propose a method to systematically elicit facts through dialogue from a simulated taxpayer . they use 500 narratives from official Polish tax interpretations to test their models .
Outcome:	The proposed model recovers only 77% of facts on easy and hard samples and under 49% on hard samples after 50 turns.

Evolutionary Search for Automated Design of Uncertainty Quantification Methods (2026.acl-srw)

Copied to clipboard

Mikhail Seleznyov, Daniil Korbut, Viktor Moskvoretskii, Oleg Somov, Alexander Panchenko, Elena Tutubalina

Explainability & Interpretability in NLP

Challenge:	Qualitative analysis reveals that different LLMs employ qualitatively distinct evolutionary strategies for automating, interpretable hallucination detector design.
Approach:	They apply LLM-powered evolutionary search to discover unsupervised UQ methods represented as Python programs and apply them to atomic claim verification.
Outcome:	The proposed methods outperform strong manually-designed baselines while generalizing robustly out-of-distribution.

Thesis Proposal: A Normalization-First Framework for Sound, Complete, and Utility-Ready Open Information Extraction (2026.acl-srw)

Copied to clipboard

Chandan Prakash, Pavan Kumar Chittimalli, Arnab Bhattacharya

Mind the Gap: Multilingual Divide in LLM Bias Detection and Reasoning (2026.acl-srw)

Challenge:	Existing approaches to extract relational tuples from text are incomplete and ambiguous . Existing methods rely on predefined schemas to produce t-uples .
Approach:	They propose a normalization-first framework that reframes OIE as a structured semantic transformation pipeline . they formalize soundness, completeness, and usefulness as approximate yet verifiable guarantees over extraction quality .
Outcome:	The proposed framework aims to make OIE usable for downstream reasoning and machine interpretability.

Copied to clipboard

Medha Hira, Prachi Goyal, Raj Maheshwari, Arnav Goel

Multi-Constraint State Tracking with Negation: A Diagnostic Benchmark for LLM World Modeling (2026.acl-srw)

Challenge:	Large Language Models (LLMs) are increasingly deployed in multilingual settings . but most bias evaluation remains English-centric and ignores how bias manifests within reasoning .
Approach:	They evaluate large language models with supervised fine-tuning and preference optimization . they find that bias varies substantially across languages, with consistent degradation in non-English settings .
Outcome:	The proposed model improves in English, Dutch, Spanish, and Turkish using the MBBQ benchmark.

Copied to clipboard

Ayan Sar, Pranav Singh Puri, Sumit Aich, Anurag Kaushish, Tanupriya Choudhury, Ajith Abraham

Learning Shortcut Models for Efficient Recursive Reasoning (2026.acl-srw)

Challenge:	Existing evaluations of Large Language Models (LLMs) lack the ability to maintain and update internal representations of an evolving world despite their ubiquity in real-world language.
Approach:	They propose a diagnostic benchmark for multi-constraint state tracking that evaluates an LLM's ability to maintain consistent world models across sequences of actions involving inventory changes, spatial movement, temporal ordering, and systematic negation.
Outcome:	The proposed model is based on 100,847 questions spanning 12 real-world domains and includes culturally diverse entity names to enable analysis of cross-cultural robustness.

Copied to clipboard

Shiv Shankar

The Confident Liar: Diagnosing Multi-Agent Debate with Log-Probabilities and LLM-as-Judge (2026.acl-srw)

Challenge:	Recent research shows that Transformer-style models can be made more efficient by sharing parameters over blocks.
Approach:	They propose a framework for distilling latent reasoning into a multiscale jump model that enables flexible test-time compute.
Outcome:	Experiments on ARC-AGI show that the proposed model achieves competitive accuracy compared to recursive baselines while requiring fewer sequential updates.

Copied to clipboard

Ali Keramati, Justin Cheok, Jacob Horne, Mark Warschauer

Convergent Demographic Utility Hierarchies: Geometry of Intersectional Values in LLMs (2026.acl-srw)

Challenge:	Multi-agent debate systems are typically evaluated only on whether the final answer is correct, overlooking the quality of the intermediatereasoning that debate produces.
Approach:	They propose a framework that pairs a two-agent debate architecture with anLLM-as-judge that scores each agent’s reasoning along instructionfollowing, justification quality, and evidence grounding, together with acritical-failure flag.
Outcome:	The proposed framework pairs a two-agent debate architecture with anLLM-as-judge that scores each agent’s reasoning along instructionfollowing, justification quality, and evidence grounding, together with acritical-failure flag.

Copied to clipboard

Pravish Sainath

Challenge:	Recent work shows that LLMs develop internally coherent utility functions that emerge with scale.
Approach:	They elicit pairwise preferences across 15 intersectional demographic groups . they fit Thurstonian utility functions to the preference matrices and find a compensatory hierarchy .
Outcome:	elicited preferences show that they encode demographic hierarchies across 15 demographic groups . gender, race, and combinations overestimate the most extreme intersectional gap by 26- 40% .

CRL-Prompt: Contrastive and Reinforcement Learning for Soft Prompt Tuning for Text Classification (2026.acl-srw)

Copied to clipboard

Danila Lapokin, Andrey Savchenko

Optimizing Packing and Shuffling Strategies for Enhanced Performance in Generative Language Models (2026.acl-srw)

Challenge:	Manual prompt engineering is time-consuming, non-scalable, and brittle, while current auto-prompting techniques are far from maturity.
Approach:	They propose a two-stage method for prompt learning of frozen language models, CRL-Prompt, based on soft prompt initialization followed by contrastive and reinforcement-based refinement.
Outcome:	The proposed method achieves consistent improvements over baseline prompt tuning strategies, with gains of up to 2.2% while training fewer than 0.25% of model parameters.

Copied to clipboard

Yanbing Chen, Ruilin Wang, Zihao Yang, Lavender Yao Jiang, Eric Karl Oermann

LLM as a Meta-Judge: Synthetic Data for NLP Evaluation Metric Validation (2026.acl-srw)

Challenge:	Documents are typically concatenated to chunks of maximum sequence length (MSL) and shuffled in chunks (atom-size chunks).
Approach:	They propose to pack and shuff documents in chunks of tokens to prevent overfitting . they also propose to use padding to only include one document per chunk .
Outcome:	The proposed method reduces the risk of overfitting and improves generalizability.

Copied to clipboard

Lukáš Eigler, Jindřich Libovický, David Hurych

Challenge:	Existing evaluation metrics for natural language generation are expensive and time-consuming.
Approach:	They propose a framework that utilizes LLMs to generate synthetic evaluation datasets . they propose meta-correlation to measure alignment between metric rankings and human benchmarks based on synthetic data .
Outcome:	The proposed framework achieves meta-correlations exceeding 0.9 in multilingual QA and replaces human judgment with synthetic evaluation datasets.

Continuous Context Sampling Allows Extending Diversity Boundaries of Large Language Models (2026.acl-srw)

Copied to clipboard

Mateusz Bystroński, Doheon Han, Nitesh V. Chawla, Tomasz Jan Kajdanowicz

One Task Vector is not Enough: A Large-Scale Study for In-Context Learning (2026.acl-srw)

Challenge:	Large language models exhibit a persistent limitation: repeated generations from the same prompt tend to be semantically similar.
Approach:	They propose to construct a conditioning distribution from a small set of diverse anchor generations and use it to condition an LLM's generation distribution.
Outcome:	The proposed framework significantly expands the model's reachable semantic range by constructing a conditioning distribution from a small set of diverse anchor generations.

Copied to clipboard

Pavel Tikhonov, Ivan Oseledets, Elena Tutubalina