Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)

LaMI: Augmenting Large Language Models via Late Multi-Image Fusion (2026.acl-short)

Challenge:	a lack of evaluation grounds for self-consistency on symbolic reasoning is unclear . however, it is unclear whether it improves performance on non-math questions involving encyclopedic knowledge.
Approach:	They establish a knowledge recall split for the popular MMLU benchmark by applying a data-driven heuristic from prior work.
Outcome:	The proposed knowledge recall split achieves an 89% accuracy on the MMLU benchmark.

Copied to clipboard

Guy Yariv, Idan Schwartz, Yossi Adi, Sagie Benaim

Full-Duplex-Bench-v2: A Multi-Turn Evaluation Framework for Duplex Dialogue Systems with an Automated Examiner (2026.acl-short)

Challenge:	Large Language Models lack visual grounding on visual reasoning, despite training on text alone.
Approach:	They propose a late multi-image fusion method that augments LLMs with test-time visual signals.
Outcome:	Using a late multi-image fusion method, the proposed model outperforms LLMs on visual reasoning and matches VLMs in vision-based tasks.

Copied to clipboard

Guan-Ting Lin, Shih-Yun Shan Kuan, Jiatong Shi, Kai-Wei Chang, Siddhant Arora, Shinji Watanabe, Hung-yi Lee

Dialogue Systems & Conversational Agents

Challenge:	Full-duplex speech agents are often half-duplice, alternating turns between user and system.
Approach:	They propose a streaming framework that integrates with an examiner that enforces staged goals under two pacing setups.
Outcome:	The framework reports fluency, multi-turn instruction following, and task-specific competence.

Situated Embedding Models for Context-Aware Dense Retrieval (2026.acl-short)

Copied to clipboard

Junjie Wu, Jiangnan Li, Yuqing Li, Lemao Liu, Liyan Xu, Jiwei Li, Dit-Yan Yeung, Jie Zhou, Mo Yu

Representation Learning

Challenge:	Existing embedding models are not well-equipped to encode situated context effectively, i.e., situating a chunk’s meaning within its context.
Approach:	They propose to represent short chunks in a way that is conditioned on a broader context window to enhance retrieval performance.
Outcome:	The proposed model outperforms state-of-the-art embedding models on a book-plot retrieval dataset.

Big AI is Accelerating the Metacrisis: What Can We Do? (2026.acl-short)

Copied to clipboard

Steven Bird

Challenge:	LLM engineering is at the core of the problem of ecological, meaning, and language crises . big AI is fueling global crises and creating wealth and power for a handful of individuals and corporations while causing existential harm to life on earth.
Approach:	et al., 2025, p162ff) argue that big AI is escalating global crises and creating a metacrisis.
Outcome:	the field of natural language processing is at the core of the problem . it is being leveraged to create unprecedented wealth and power for a handful of individuals and corporations while causing existential harm to life on earth.

From Factuality to Meta-Factivity: A Cognitive Blueprint for Trustworthy LLMs (2026.acl-short)

Copied to clipboard

Liu Daohuan, Xia Lun, Yuer Wang, Jiaoyang Su, Xuri Tang

Attention Sinks Are Provably Necessary in Softmax Transformers: Evidence from Trigger-Conditional Tasks (2026.acl-short)

Challenge:	Current evaluation paradigms on Event Factuality Prediction (EFP) focus on static classification tasks and shortcut learning and unreliable reasoning.
Approach:	They propose a meta-factivity framework that moves evaluation beyond surface recognition to belief trajectory reasoning and epistemic regulation.
Outcome:	The proposed framework shifts from event factuality to meta-factivity . the proposed framework lays the groundwork for a more rigorous benchmark for explainable self-governance .

Copied to clipboard

Yuval Ran-Milo

A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation (2026.acl-short)

Challenge:	Xiao et al., 2024) show that softmax models display an attention sink . he argues that normalization over a probability simplex must force attention to collapse onto a stable anchor to realize a default state.
Approach:	They show that normalization over a trigger-conditional behavior necessarily induces a sink in softmax self-attention models.
Outcome:	The proposed model can solve a task with no sink in softmax models.

Copied to clipboard

Yuval Ran-Milo, Hila Ofek, Shahar Mendel

Is a Document Educational or Just Wikipedia-Style? — Pitfalls of Classifier-Based Quality Filtering (2026.acl-short)

Challenge:	Xiao et al., 2025) show a tendency to allocate disproportionate attention mass to early (often first) positions independent of semantic content.
Approach:	They find that Transformers display an attention sink: disproportionate attention to the first position.
Outcome:	The proposed sinks are found in GPT-2–style models with learned query biases and absolute positional embeddings.

Copied to clipboard

Mateusz Klimaszewski, Piotr Andruszkiewicz

On the Hidden Objective Biases of Group-based Reinforcement Learning (2026.acl-short)

Challenge:	Large Language Models (LLMs) are pre-trained on massive data corpora, and the quality of these corporales is one of the main factors in achieving stateof-the-art performance.
Approach:	They propose to use Wikipedia-style reformatting to alter a model's quality assessment and enable low-quality content to surpass filtering thresholds.
Outcome:	The proposed model would reverse filtering decision for approximately 7% of evaluated documents, thereby admitting content into the pre-training corpus that would otherwise have been excluded.

Copied to clipboard

Aleksandar Fontana, Marco Simoni, Giulio Rossolini, Paolo Mori, Andrea Saracino

Challenge:	Recent studies have reported unexpected behaviors during training, including lengthrelated biases, formatting tokens, and reward hacking in multi-objective settings.
Approach:	They propose to analyze group-based reinforcement learning methods within a unified surrogate formulation.
Outcome:	The proposed methods exhibit structural mismatches between reward optimization and the underlying training objective.

StructMem: Structured Memory for Long-Horizon Behavior in LLMs (2026.acl-short)

Copied to clipboard

Buqiang Xu, Yijun Chen, Jizhan Fang, Ruobin Zhong, Yunzhi Yao, Yuqi Zhu, Lun Du, Shumin Deng

Z3D: Zero-Shot 3D Visual Grounding from Images (2026.acl-short)

Challenge:	Existing memory systems lack structure and efficiency in capturing relationships between events.
Approach:	They propose a structure-enriched hierarchical memory framework that preserves event-level bindings and induces cross-event connections.
Outcome:	The proposed framework preserves event-level bindings and induces cross-event connections while reducing token usage, API calls, and runtime compared to prior memory systems.

Copied to clipboard

Nikita Drozdov, Andrey Lemeshko, Nikita Gavrilov, Anton Konushin, Danila Rukhovich, Maksim Kolodiazhnyi

Low-Resource NLP

Challenge:	3D visual grounding (3DVG) aims to localize objects in a 3D scene based on natural language queries.
Approach:	They propose a zero-shot 3D visual grounding pipeline that operates on multi-view images without geometric supervision and without object priors.
Outcome:	Experiments on ScanRefer and Nr3D show that the proposed method outperforms existing methods.

Improving Retrieval-Augmented Generation without Taxonomy-based Error Categorization (2026.acl-short)

Copied to clipboard

Gongbo Zhang, Yifan Peng, Chunhua Weng

Information Retrieval

Challenge:	Recent work implicitly assumes reliable critic feedback and focuses on planning strategies, while paying limited attention to the robustness of the error correction process itself.
Approach:	They propose a response-action learning paradigm that maps flawed RAG outputs to error-mitigating action plans without explicit criticism.
Outcome:	The proposed model improves the factual accuracy of large language model outputs without explicit error categorization.

Deep Kernel Fusion for Transformers (2026.acl-short)

Copied to clipboard

Zixi Zhang, Zhiwen Mo, Yiren Zhao, Robert D. Mullins

Anchoring Depends on Confidence and Post-Training in Language Models (2026.acl-short)

Challenge:	Agentic LLM inference with long contexts is limited by memory bandwidth rather than compute.
Approach:	They propose a deeply fused kernel that cuts HBM traffic and boosts cache reuse.
Outcome:	The proposed kernel delivers 13.2% speedup on H100 and 9.7% on A100 over SGLang.

Copied to clipboard

Hillary N. Owusu, Naomi H. Feldman

LLMs Underperform Graph-Based Parsers on Supervised Relation Extraction for Complex Graphs (2026.acl-short)

Challenge:	Existing work has demonstrated the presence of anchoring bias in large language models . Existing research does not predict when a model will be most susceptible to anchoring .
Approach:	They analyze anchoring bias as a function of model confidence and accuracy . they find that incorrect models resist anchoring as effectively as accurate ones .
Outcome:	The findings suggest that anchoring resistance is a structural property of uncertainty rather than knowledge correctness.

Copied to clipboard

Paolo Gajo, Domenic Rosati, Hassan Sajjad, Alberto Barrón-Cedeño

Evolutionary Strategies at Scale lead to Catastrophic Forgetting (2026.acl-short)

Challenge:	Relation extraction is a core NLP task which involves extracting [head, relation, dependent] RDF triples from text.
Approach:	They evaluate four large language models against a graph-based parser on six relation extraction datasets with sentence graphs of varying sizes and complexities.
Outcome:	The graph-based parser outperforms the LLMs on six relation extraction datasets with sentence graphs of varying sizes and complexities.

Copied to clipboard

Immanuel Abdi, Akshat Gupta, Micah Mok, Alex Lu, Nicholas Lee, Gopala Anumanchipalli

Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data (2026.acl-short)

Challenge:	ES has been shown to improve performance on specific tasks, but it is accompanied by significant forgetting of prior abilities.
Approach:	They propose to use Evolutionary Strategies to train gradient-free algorithms to improve performance.
Outcome:	The proposed algorithm achieves performance numbers closer to GRPO for math and reasoning tasks, but forgets prior abilities.

Copied to clipboard

Zhenwen Liang, Yujun Zhou, Sidi Lu, Xiangliang Zhang, Haitao Mi, Dong Yu

Reliable Use of Lemmas via Eligibility Reasoning and Section-Aware Reinforcement Learning (2026.acl-short)

Challenge:	Strong base models saturate benchmarks, resulting in weaker performance, a paradox . a new approach to Reinforcement Learning (RL) is needed to improve performance .
Approach:	They propose a method that uses constrained uniform top-k sampling to flatten the local optimization landscape by sampling uniformly from constrained high-confidence candidates.
Outcome:	Experiments show that the proposed approach prevents policy degeneration and boosts out-of-domain generalization.

Copied to clipboard

Zhikun Xu, Xiaodong Yu, Ben Zhou, Jiang Liu, Jialian Wu, Ze Wang, Ximeng Sun, Hao Chen, Zicheng Liu

Robustness in NLP

Challenge:	Recent large language models (LLMs) perform strongly on mathematical benchmarks but often import conclusions without validating assumptions.
Approach:	They propose a model that encodes a lemma specification and trains with reinforcement learning and section-aware loss masking to assign penalty to the section responsible for errors.
Outcome:	The proposed model performs well on benchmarks but often misapplyes lemmas . the model is able to encode the specification and train with reinforcement learning .

Attention Under Attack: Analog Noise Effects and Mechanistic Vulnerabilities in Transformer Models (2026.acl-short)

Copied to clipboard

Mafizur Rahman, Lijun Qian

ReproEvalCard: A Reporting Standard for Reproducible Evaluation of LLM Pipelines (2026.acl-short)

Challenge:	Analog in-memory computing (AIMC) provides substantial efficiency gains for transformer inference but introduces hardware-induced noise that can distort attention behavior.
Approach:	They present the first fine-grained analysis of analog vulnerability in pretrained transformers . query (Q), key (K), and value (V) projections are most sensitive components .
Outcome:	The proposed analysis shows that query (Q), key (K), and value (V) projections are the most sensitive components .

Copied to clipboard

Priyaranjan Pattnayak, Apoorv Bhatia

Test-Time Reasoners Are Strategic Multiple-Choice Test-Takers (2026.acl-short)

Challenge:	Existing evaluation standards for multistage pipelines are inconsistent, leaving the reproducibility and independent validation of published evaluations unclear.
Approach:	They propose a lightweight reporting standard that specifies the minimum artifacts required to reproduce and validate LLM evaluations.
Outcome:	The proposed standard audits 55 pipeline-based LLM papers published between 2022 and 2025 and quantifies the availability of reproducibility-critical evaluation artifacts.

Copied to clipboard

Nishant Balepur, Atrey Desai, Rachel Rudinger

MARCH: Multi-Agent Radiology Clinical Hierarchy for CT Report Generation (2026.acl-short)

Challenge:	Large language models (LLMs) give reasoning before answering, excelling in multiple-choice question answering (MCQA) . but, some studies find that LLMs sans reasoning fail in MCQA without using the question, i.e., choices-only.
Approach:	They propose to use reasoning LLMs to separate problematic data from less problematic strategies by examining reasoning traces.
Outcome:	The proposed models perform well in multiple-choice question answering without the question, but they fail to use the question.

Copied to clipboard

Yi Lin, Yihao Ding, Yonghui Wu, Yifan Peng

Text Generation

Challenge:	Automated 3D radiology report generation suffers from clinical hallucinations and lacks the iterative verification characteristic of clinical workflows.
Approach:	They propose a multi-agent framework that emulates the professional hierarchy of radiology departments and assigns specialized roles to distinct agents.
Outcome:	The proposed framework outperforms state-of-the-art models in clinical fidelity and linguistic accuracy on the RadGenome-ChestCT dataset.

Prefix Parsing is Just Parsing (2026.acl-short)

Copied to clipboard

Clemente Pasti, Andreas Opedal, Timothy J. O’Donnell, Ryan Cotterell, Tim Vieira

Green & Sustainable NLP

Challenge:	Existing prefix parsers are typically tied to particular parsing algorithms.
Approach:	They propose a prefix grammar transformation that reduces prefix parsing to ordinary parsers . they propose enabling prediction of the next token by computing the next-token weight vector .
Outcome:	The proposed method reduces prefix parsing to ordinary parsers without modification . the transformed grammar is only a small factor larger than the input .

Privacy-preserving Prosody Representation Learning (2026.acl-short)

Copied to clipboard

Kevin Everson, Mari Ostendorf

Enhancing Linguistic Competence of Language Models through Pre-training with Language Learning Tasks (2026.acl-short)

Challenge:	Acoustic-prosodic cues are known to carry speaker information, exposing users to privacy breaches . a new self-supervised learning approach addresses privacy concerns .
Approach:	They propose a self-supervised approach to learning prosody representations that incorporates speaker disentanglement strategies.
Outcome:	The proposed model outperforms raw prosody and HuBERT-base baselines on three tasks . it achieves strong speaker disentanglement without adverse impact on prosody-related downstream tasks compared with baselines .

Copied to clipboard

Atsuki Yamaguchi, Maggie Mi, Nikolaos Aletras

Rethinking Data Mixing from the Perspective of Large Language Models (2026.acl-short)

Challenge:	Language models (LMs) are pre-trained on raw text datasets to generate text sequences token-by-token.
Approach:	They propose a framework that integrates Language Learning Tasks alongside standard next-token prediction to stimulate the acquisition of morphological, syntactic, and semantic knowledge.
Outcome:	The proposed framework improves performance on linguistic competence benchmarks while maintaining competitive performance on reasoning tasks.

Copied to clipboard

Yuanjian Xu, Tianze Sun, Changwei Xu, XinLong Zhao, Jianing Hao, Ran Chen, Yang Liu, Ruijie Xu, Stephen Chen, Guang Zhang

Translation or Recitation? Calibrating Evaluation Scores for Machine Translation of Extremely Low-Resource Languages (2026.acl-short)

Challenge:	Existing methods to mix data with LLMs have relied on domain definitions derived from intuition.
Approach:	They propose a reweighting framework that restructures data scheduling as a graph-constrained optimization problem.
Outcome:	The proposed framework achieves competitive performance on GPT-2 models.

Copied to clipboard

Danlu Chen, Ka Sing He, Jiahe Tian, Chenghao Xiao, Zhaofeng Wu, Taylor Berg-Kirkpatrick, Freda Shi

Multilinguality

Challenge:	Existing studies show that performance across low-resource settings is variable, resulting in a significant barrier for the MT community.
Approach:	They propose to use FRED Difficulty Metrics to contextualize reported performance across different language pairs to determine whether breakthroughs reported in other contexts are artifacts of benchmark collection.
Outcome:	The proposed metrics explain a significant portion of result variability rather than model capability.

CheckMIABench: Firm Foundations For Membership Inference Attacks on Language Models (2026.acl-short)

Copied to clipboard

Jeffrey George Wang, Jason Wang, Marvin Li, Seth Neel

Luring as a Proxy: Evaluating Corpus Transferability for Cybergrooming Detection (2026.acl-short)

Challenge:	Membership inference attacks are a canonical way to assess a machine learning model’s privacy properties.
Approach:	They propose a framework for principled evaluation of membership inference attacks against large language models by leveraging the insight that training data before and after a fixed point during training are drawn from the same distribution.
Outcome:	The proposed framework can be used to evaluate membership inference attacks against large language models.

Copied to clipboard

Shiying Fan, Mareike Bassenge, Martin Steinebach

From Narrow Unlearning to Emergent Misalignment in LLMs (2026.acl-short)

Challenge:	Prior research has noted that cybergrooming often involves the use of luring communication strategies to manipulate minors.
Approach:	They examine the potential transferability of corpora from luring contexts for cybergrooming detection.
Outcome:	The proposed model can be generalized across domains and performs better on corpora with salient toxicity and distinctive stylistic features.

Copied to clipboard

Erum Mushtaq, Anil Ramakrishna, Satyapriya Krishna, Sattvik Sahai, Prasoon Goyal, Kai-Wei Chang, Tao Zhang, Rahul Gupta

Challenge:	Recent work shows that fine-tuning on insecure code data can trigger an emergent misalignment (EMA) phenomenon .
Approach:	They extend their study by demonstrating that EMA can arise from narrow refusal unlearning . they perform refusal unLearning on Cybersecurity and Safety concept and evaluate EMA .
Outcome:	The proposed model can generate malicious responses even to unrelated prompts . the proposed model is able to restore alignment across the affected domains while having lower refusal rates.

Reliable Evaluation Protocol for Low-Precision Retrieval (2026.acl-short)

Copied to clipboard

Kisu Yang, Yoonna Jang, Hwanseok Jang, Kenneth Choi, Isabelle Augenstein, Heuiseok Lim

Information Retrieval

Challenge:	Recent studies have shown that low-precision methods can improve performance, but they introduce high variability in the results based on tie resolution.
Approach:	They propose a retrieval evaluation protocol designed to reduce tie variation . high-precision scoring and tie-aware retrieval metrics are proposed to reduce this variability .
Outcome:	The proposed retrieval evaluation protocol reduces tie-induced instability and recovers expected scores and ranges on 12 retrieval datasets.

When More Words Say Less: Decoupling Length and Specificity in Image Description Evaluation (2026.acl-short)

Copied to clipboard

Rhea Kapur, Robert D. Hawkins, Elisa Kreiss

Visual Data in NLP

Challenge:	Vision-language models are increasingly used to produce textual descriptions of visual content.
Approach:	They propose to disentangle description specificity from description length . they find people prefer more specific descriptions regardless of length based on their own subjective preferences .
Outcome:	The proposed model shows that people prefer more specific descriptions regardless of length.

Selective Span-Level Unlearning for Large Language Models (2026.acl-short)

Copied to clipboard

Chaewon Yoon, Dongjun Kim, Hyun-Je Song

Calibrated? Not for Everyone: How Sexual Orientation and Religious Markers Distort LLM Accuracy and Confidence in Medical QA (2026.acl-short)

Challenge:	Existing selective methods that focus on identifying token-level or span-level unlearning targets are misaligning unlearning objectives with the model’s internal behavior.
Approach:	They propose a selective method that uses model-intrinsic information to identify token-level or span-level unlearning targets within a text rather than entire sequences.
Outcome:	The proposed method achieves comparable unlearning performance while significantly better preserving retained knowledge.

Copied to clipboard

Alberto Testoni, Iacer Calixto

Question Answering

Challenge:	a growing body of research suggests that social descriptors can influence LLM-generated clinical recommendations.
Approach:	They examine whether social descriptors of a patient distort uncertainty signals and model accuracy.
Outcome:	The presence of social identity cues affects the reliability of confidence signals, the authors show . incorporating sociodemographic attributes alters outputs in clinical trial matching and QA .

DIXITWORLD: Evaluating Multimodal Abductive Reasoning in Vision-Language Models with Multi-Agent Dixit Gameplay (2026.acl-short)

Copied to clipboard

Yunxiang MO, Tianshi Zheng, Qing Zong, Jiayu Liu, Baixuan Xu, Yauwai Yim, Chunkit Chan, Jiaxin Bai, Yangqiu Song

UERLens: Understanding Event Relations in Large Language Models (2026.acl-short)

Challenge:	Existing evaluations of multimodal abductive reasoning are limited to static, single-agent tasks.
Approach:	They propose a multiagent evaluation suite that deconstructs the current evaluations of multimodal abductive reasoning in vision–language models.
Outcome:	The evaluation suite is based on two core components: DixitArena and DixitsBench.

Copied to clipboard

Yong Guan, Zhiyuan Li, Shaoru Guo

GOLEMcoref: A Multilingual Coreference Dataset of Fiction (2026.acl-short)

Challenge:	Existing studies on event relation extraction (ERE) have focused on improving model performance.
Approach:	They propose an interpretability framework for understanding event relations in large language models . they first construct a counterfactual dataset that includes causal, temporal, and sub-event relations .
Outcome:	The proposed framework improves event relation extraction by leveraging internal features to train a lightweight classifier.

Copied to clipboard

Andreas Van Cranenburgh, Xiaoyan Yang, null Alvanita, Cecilia Nicole Di Domenico, Maria Ferragud, Arianna Graciotti, Byungjun Kim, Seonyeong Park, Noa Visser Solissa, Xiaoyu Zhou, Federico Pianzola

Language-Aware Token Boosting: LLM Language Confusion Reduction Without Tuning (2026.acl-short)

Challenge:	Despite considerable progress, most research still focuses predominantly on English . fictional texts bring additional challenges not covered by standard benchmark datasets .
Approach:	They present a multilingual coreference dataset of 827k fanfiction tokens in 7 languages . they discuss their annotation scheme and language-specific challenges .
Outcome:	The proposed dataset includes full stories of diverse lengths, ranging from 500 to 17k words.

Copied to clipboard

Trapoom Ukarapol, Pakhapoom Sarapat, Nut Chukamphaeng

Pref-CTRL: Preference Driven LLM Alignment using Representation Editing (2026.acl-short)

Challenge:	Existing approaches to reduce language confusion rely on fine-tuning . Existing methods rely only on fine tuning to mitigate this issue .
Approach:	They propose a tuning-free paradigm for reducing language confusion by applying targeted perturbations to tokens associated with the desired language and Adaptive Language-Aware Token Boosting.
Outcome:	The proposed methods improve multilingual alignment while maintaining the summarization quality without additional fine-tuning.

Copied to clipboard

Imranul Ashrafi, Inigo Jauregi Unanue, Massimo Piccardi

Representation Learning

Challenge:	Recent work suggests that test-time alignment methods are inefficient because they require a large number of computational resources.
Approach:	They propose a preference-based training framework that uses a multi-objective value function to better reflect the structure of preference data.
Outcome:	The proposed framework outperforms RE-Control and shows greater generalization on out-of-domain datasets.

Dark & Stormy: Modeling Humor in Sentences from the Bulwer-Lytton Fiction Contest (2026.acl-short)

Copied to clipboard

Venkata S Govindarajan, Laura Biester

Commonsense Reasoning

Challenge:	a corpus of "bad" humor sentences from the Bulwer-Lytton Fiction Contest 1 is presented . standard humor detection models perform poorly on corpus, and these sentences combine features common in existing humor datasets with metaphor, metafiction and simile.
Approach:	They propose to analyze a corpus of "bad" humor sentences from the Bulwer-Lytton Fiction Contest . they use literary devices to synthesize contest-style sentences that imitate the form but exaggerate the effect .
Outcome:	The proposed corpus of sentences from the Bulwer-Lytton Fiction Contest 1 is analyzed . it shows that the sentences combine features common in existing humor datasets with metaphor, metafiction and simile.

The Personalization Trap: How User Memory Alters Emotional Reasoning in LLMs (2026.acl-short)

Copied to clipboard

Xi Fang, Weijie Xu, Yuchong Zhang, Scott Nickleach, Stephanie Eckman, Chandan K. Reddy

Neuro-Symbolic Agentic Reinforcement Learning for Long-Term Original Character Companionship and Interaction (2026.acl-short)

Challenge:	Using long-term memory, large language models can embed social hierarchies into their emotional reasoning.
Approach:	They evaluate 15 large language models on validated emotional intelligence tests to examine how user memory affects emotional intelligence.
Outcome:	The results show that the models with advantaged profiles receive more accurate emotional interpretations.

Copied to clipboard

Zhenhan Huang

Knowledge Representation

Challenge:	Existing LLM-based agents that are optimized by prompting or supervised fine-tuning exhibit a generalization gap in long-horizon, socially rich interactions.
Approach:	They propose a framework that formalizes OC companion agents’ interactions as a POMDP and decomposes the agent into three sub-policies optimized via closed-loop RL from AI feedback with verifiable rewards in a graph-constrained action space.
Outcome:	The proposed framework formalizes OC companion agents’ interactions as a POMDP and decomposes the agent into three sub-policies (Router, Memory, and Persona) with verifiable rewards in a graph-constrained action space.

Taming Extreme Tokens: Covariance-Aware GRPO with Gaussian-Kernel Advantage Reweighting (2026.acl-short)

Copied to clipboard

Cheng Wang, Qin Liu, Wenxuan Zhou, Muhao Chen

On the Rejection Criterion for Proxy-based Test-time Alignment (2026.acl-short)

Challenge:	Excessive exploitation can cause the model to become overconfident in its suboptimal solutions, thereby limiting its capabilities to explore novel reasoning strategies.
Approach:	They propose a method that dynamically down-weights extreme token-level updates via a Gaussian kernel and reduces the instability caused by the trade-off.
Outcome:	The proposed method improves downstream performance across reasoning benchmarks and stabilizes entropy as training progresses.

Copied to clipboard

Ayoub Hammal, Pierre Zweigenbaum, Caio Corro

Challenge:	Recent work suggests that test-time alignment methods rely on a small aligned model as a proxy that guides the generation of a larger base model.
Approach:	They propose a rejection criterion based on a conservative confidence bet for test-time alignment methods that use a small aligned model as a proxy to guide the generation of a larger base model.
Outcome:	The proposed approach outperforms previous work on several datasets.

Defense Against Knowledge Poisoning Attack on GraphRAG (2026.acl-short)

Copied to clipboard

Havva Alizadeh Noughabi, Fattane Zarrinkalam, Ali Dehghantanha

Robustness in NLP

Challenge:	Existing GraphRAGs expose a new attack surface: corpus-level knowledge poisoning can corrupt query-specific subgraphs and steer the generator toward incorrect answers.
Approach:	They propose a defense layer between retriever and generator that decomposes multi-hop questions into ordered subqueries and monitors hop-wise execution for poisoning-induced inconsistencies.
Outcome:	The proposed defense layer decomposes multi-hop questions into ordered subqueries and monitors hop-wise execution for poisoning-induced inconsistencies.

PExA: Parallel Exploration Agent for Complex Text-to-SQL (2026.acl-short)

Copied to clipboard

Tanmay Parekh, Ella Hofmann-Coyle, Shuyi Wang, Sachith Sri Ram Kothur, Srivas Prasad, Yunmo Chen

Programming Languages in NLP

Challenge:	Recent work in text-to-SQL has explored toolaugmented LLMs, deep planning, and agentic workflows to address complex challenges.
Approach:	They validated a framework for text-to-SQL, Spider 2.0, with 70.2% execution accuracy.
Outcome:	The proposed framework achieves 70.2% execution accuracy on a state-of-the-art benchmark for text-to-SQL, Spider 2.0.

Skill-Aware Data Selection and Fine-Tuning for Data-Efficient Reasoning Distillation (2026.acl-short)

Copied to clipboard

Lechen Zhang, Yunxiang Zhang, Wei Hu, Lu Wang

Experiments or Outcomes? Probing Scientific Feasibility in Large Language Models (2026.acl-short)

Challenge:	Large reasoning models such as DeepSeek-R1 and their distilled variants achieve impressive performance on complex reasoning tasks, yet their costs remain substantial.
Approach:	They propose a skill-centric distillation framework that efficiently transfers reasoning ability to weaker models with two components: (1) Skill-based data selection, which prioritizes examples targeting the student model’s weaker skills, and (2) Skillaware fine-tuning, which encourages explicit skill decomposition during problem solving.
Outcome:	The proposed framework surpasses baselines on Qwen3-4B and Qwend3-8B and focuses on skills emphasized during training.

Copied to clipboard

Seyedali Mohammadi, Manas Gaur, Francis Ferraro

CaBSALLM: Efficient Context-Aware Batch Annotation of Conversational Streams with Large Language Models (2026.acl-short)

Challenge:	Scientific feasibility assessment asks whether a claim aligns with established knowledge and whether experimental evidence could support or refute it.
Approach:	They frame scientific feasibility assessment as a diagnostic reasoning task . given a hypothesis, a model predicts feasible or infeasible and justifies its decision . they evaluate large language models under controlled knowledge conditions .
Outcome:	The results show that providing outcome evidence is more reliable than providing experiment descriptions.

Copied to clipboard

Mohammadsadegh Abolhasani, Reza Mousavi, Paul Jen-Hwa Hu

Dialogue Systems & Conversational Agents

Challenge:	Large-scale annotations of subjective, discourse-dependent social interactions remain a critical bottleneck in computational social science.
Approach:	They propose a pipeline that incorporates lightweight conversational context and a dynamic batching method to improve throughput and scalability.
Outcome:	The proposed pipeline improves throughput and scalability while preserving interpretive depth essential to complex social annotations.

Challenging the Explanation Based on Preceding Tokens: Discovering Transferable Non-Literal Biasing (2026.acl-short)

Copied to clipboard

Yuchen Huang, Junpeng Zhang, Quanshi Zhang

Explainability & Interpretability in NLP

Challenge:	et al. (2017) show that the generated preceding tokens may push the large language model towards the target answer.
Approach:	They find that generated preceding tokens may push large language models towards the target answer . they suggest that the LLM may intentionally use the semantically unrelated tokens to help generation of the target .
Outcome:	The generated preceding tokens may push the large language model towards the target answer . the biased connotations of the target response can also transfer to other prompts .

One-step Nonautoregressive Natural Language Generation with Shortcut Flow Matching Models (2026.acl-short)

Copied to clipboard

Jędrzej Warczyński, Ondrej Dusek, Mateusz Lango

Text Generation

Challenge:	Recent diffusion-based approaches to text generation are inefficient due to the need for multiple denoising steps.
Approach:	They propose a shortcut flow-matching model that learns to directly predict multi-step denoising outcomes in a single step.
Outcome:	The proposed model improves on three datasets and can predict multi-step denoising outcomes in a single step.

SED-SFT: Selectively Encouraging Diversity in Supervised Fine-Tuning (2026.acl-short)

Copied to clipboard

Yijie Chen, Yijin Liu, Fandong Meng

Frame-Semantic Knowledge Injection for Event-Level Inference in LLMs (2026.acl-short)

Challenge:	Existing studies have proposed a new approach to optimize for SFT followed by RL . existing studies have suggested a method to optimize SFT for large language models .
Approach:	They propose a framework that encourages diversity based on token exploration space.
Outcome:	Experiments show that SED-SFT significantly improves generation diversity with a negligible computational overhead increase over CE loss.

Copied to clipboard

Shahid Iqbal Rai, Danilo Croce, Roberto Basili

Exploring Cross-Client Memorization of Training Data in Large Language Models for Federated Learning (2026.acl-short)

Challenge:	Large language models (LLMs) are fluent but often brittle when interpretation depends on external information.
Approach:	They propose a framework that injects frame-semantic knowledge into Large Language Models via LoRA.
Outcome:	The proposed framework can generalize beyond surface cues in large language models.

Copied to clipboard

Tinnakit Udsa, Can Udomcharoenchaikit, Patomporn Payoungkhamdee, Sarana Nutanong, Norrathep Rattanavipanon

FL-MSCL: A Unified Figurative Language Detection Model Driven by Multi-Type Signals and Contrastive Learning (2026.acl-short)

Challenge:	Existing methods to assess memorization in federated learning focus on one sample at a time . centralized learning does not eliminate the risk of memorizing large language models .
Approach:	They propose a framework that quantifies both intra- and inter-client memorization in FL . they use fine-grained cross-sample memorisation measurement across all clients .
Outcome:	The proposed framework quantifies both intra- and inter-client memorization in FL using fine-grained cross-sample memorisation measurement across all clients.

Copied to clipboard

Lu Shijia, Fumiyo Fukumoto, Huang Xiaoxi, Yoshimi Suzuki

Revisiting Evaluation of Question Answering Systems in Low-Resource Indic Languages: Bridging Human and Metric Alignment (2026.acl-short)

Challenge:	Figurative language recognition challenges distinguishing between fine-grained rhetorical categories . existing approaches are framed as single-category binary classifiers .
Approach:	They propose a framework that integrates prompt-based knowledge injection with supervised contrastive learning to enforce explicit class distinctions.
Outcome:	The proposed framework achieves competitive performance on a four-way sentence-level classification task.

Copied to clipboard

Anuj Kumar, Satyadev Ahlawat, Yamuna Prasad, Virendra Singh

Question Answering

Challenge:	Evaluating Question Answering systems in low-resource Indic languages remains challenging due to the scarcity of annotated data and the lack of reliable evaluation metrics.
Approach:	They propose a language-based multi-aspect evaluation framework for question answering systems . the framework integrates semantic similarity, factual completeness, numerical accuracy and contextual relevance .
Outcome:	The proposed metric is evaluated across eight Indic-language QA tasks using multiple LLMs . Across all settings, it shows stronger agreement with human evaluation .

Data-efficient Targeted Token-level Preference Optimization for LLM-based Text-to-Speech (2026.acl-short)

Copied to clipboard

Rikuto Kotoge, Yuichi Sasaki

How Do Inpainting Artifacts Propagate to Language? (2026.acl-short)

Challenge:	Recent work has shown that text-to-speech (TTS) models generate contextaware pronunciations from raw text without morphological analysis.
Approach:	They propose a preference optimization algorithm that aligns text-to-speech (TTS) outputs with human feedback.
Outcome:	The proposed method improves the challenging Japanese pronunciation accuracy by 39% and reduces CER by 54%.

Copied to clipboard

Pratham Yashwante, Davit Abrahamyan, Shresth Grover, Sukruth Rao

LOGICAL-COMMONSENSEQA: A Benchmark for Logical Commonsense Reasoning (2026.acl-short)

Challenge:	Figure 1 shows representative examples of visual artifacts introduced by diffusion-based inpainting . despite visually plausible reconstructions, localized inpainding artifactors lead to object substitutions, attribute changes, or category-level errors in downstream captions.
Approach:	They propose a diagnostic setup in which masked image regions are reconstructed and then provided to captioning models.
Outcome:	The proposed diagnostic framework can be used to examine how visual artifacts affect language generation in vision-language models.

Copied to clipboard

Obed Junias, Maria Leonor Pacheco

BioHiCL: Hierarchical Multi-Label Contrastive Learning for Biomedical Retrieval with MeSH Labels (2026.acl-short)

Challenge:	LOGICAL-COMMONSENSEQA benchmarks evaluate commonsense reasoning as logical composition over pairs of atomic statements . commonsensible reasoning is central to human cognition and a long-standing challenge in artificial intelligence and natural language understanding.
Approach:	They propose a benchmark that reframes commonsense reasoning as logical composition over pairs of atomic statements using plausibility-level operators.
Outcome:	LOGICAL-COMMONSENSEQA exposes fundamental reasoning limitations and provides a framework for advancing compositional commonsense reasoning.

Copied to clipboard

Mengfei Lan, Lecheng Zheng, Halil Kilicoglu

Dialogue is the Plan: From Interface to Joint Action in Agentic AI (2026.acl-short)

Challenge:	Existing biomedical generative retrievers lack domain semantics and hierarchical relationships among biomedically related texts.
Approach:	They propose a biomedical retrieval model with hierarchical multi-label contrastive learning that leverages hierarchic MeSH annotations to provide structured supervision for multi-labor contrastive training.
Outcome:	The proposed models achieve promising performance on biomedical retrieval, sentence similarity, and question answering tasks while remaining computationally efficient for deployment.

Copied to clipboard

Mert Inan, Malihe Alikhani, Anthony Sicilia

Dialogue Systems & Conversational Agents

Challenge:	Large Language Model agents' language use is often used as an interface for instructing and reporting results.
Approach:	They argue that large language models are often used as an interface for instructingactions and reporting results.
Outcome:	We show that large-scale language models can be used to plan and act, yet their language is often used as an interface for instructing and reporting results.

Late Code Chunking: A Code Chunking Strategy for Repository-Level Code Completion (2026.acl-short)

Copied to clipboard

Seungmin Oh, Eunseok Lee

Programming Languages in NLP

Challenge:	Despite significant advancements in Large Language Models (LLMs), repository-level code completion remains a challenging area.
Approach:	They propose a chunking strategy to improve the semantic understanding of code segments for Large Language Models.
Outcome:	The proposed strategy improves the semantic understanding of code segments for Large Language Models.

Temporal Leakage in Search-Engine Date-Filtered Web Retrieval: A Retrospective Forecasting Case Study (2026.acl-short)

Copied to clipboard

Ali El Lahib, Ying-Jieh Xia, Zehan Li, Yuxuan Wang, Xinyu Pi

Information Retrieval

Challenge:	Search-engine date filters are widely used to enforce pre-cutoff retrieval in retrospective evaluations of search-augmented forecasters.
Approach:	They propose stronger retrieval safeguards or evaluation on frozen, time-stamped web snapshots to prevent post-cutoff leakage.
Outcome:	The proposed approach is unreliable across two major search engines, and the results are inflated.

A Shared Geometry of Difficulty in Multilingual Language Models (2026.acl-short)

Copied to clipboard

Stefano Civelli, Pietro Bernardelle, Nicolò Brunello, Gianluca Demartini

T⋆: Progressive Block Scaling for Masked Diffusion Language Models Through Trajectory Aware Reinforcement Learning (2026.acl-short)

Challenge:	Large language models encode problem difficulty as an internal signal that can be linearly decoded from their residuals.
Approach:	They train linear probes on the AMC subset of the Easy2Hard benchmark, translated into 21 languages, and found difficulty-related signals emerge at two distinct stages of the model internals.
Outcome:	The results show that difficulty-related signals emerge at two distinct stages of the model internals, corresponding to shallow (early-layers) and deep (later-layer) representations, that exhibit functionally different behaviors.

Copied to clipboard

Hanchen Xia, Baoyou Chen, Yutang Ge, Guojiang Zhao, Siyu Zhu

Learning from Emptiness: De-biasing Listwise Rerankers with Content-Agnostic Probability Calibration (2026.acl-short)

Challenge:	Autoregressive (AR) modeling via next-token prediction dominates scaling practice and deployed systems.
Approach:	They propose a TraceRL-based curriculum for progressive block-size scaling in masked diffusion language models.
Outcome:	The proposed curriculum outperforms direct large-block TraceRL on two SDAR scales and three benchmarks and retains block-size-specific non-monotone updates while improving accuracy.

Copied to clipboard

Hang Lv, Hongchao Gu, Ruiqing Yang, Liangyue Li, Zulong Chen, Defu Lian, Hao Wang, Enhong Chen

Decoupling Generalization and Adaptation in Meta-Learning for Large Language Models (2026.acl-short)

Challenge:	Existing methods for listwise reranking exhibit intrinsic position bias . existing methods are constrained by an inherent trade-off between efficiency and flexibility .
Approach:	They propose a training-free framework that mechanically decouples positional bias from ranking decisions.
Outcome:	a training-free framework decouples position bias from ranking decisions . evaluations show it outperforms training-based methods and outperformed expensive methods .

Copied to clipboard

Nitin Vetcha, Binqian Xu, Dianbo Liu

What Makes AI Research Replicable? Executable Knowledge Graphs as Scientific Knowledge Representations (2026.acl-short)

Challenge:	Adapting large language models to specific downstream tasks requires multi-step fine-tuning with substantial training data, incurring significant computational overhead.
Approach:	They propose a framework that separates learning generalizable initializations and adaptation through dedicated parameter spaces.
Outcome:	The proposed framework outperforms existing meta-learning and standard multi-task baselines on common-sense reasoning, mathematics, logic, medical and coding benchmarks.

Copied to clipboard

Yujie Luo, Zhuoyun Yu, Xuehai Wang, Yuqi Zhu, Ningyu Zhang, Lanning Wei, Lun Du, Da Zheng, Huajun Chen

Chain-of-Thought Degrades Visual Spatial Reasoning Capabilities of Multimodal LLMs (2026.acl-short)

Challenge:	Existing approaches to replicate AI research are limited by insufficient background knowledge and the limitations of retrieval-augmented generation methods.
Approach:	They propose a pluggable, paper-centric knowledge base that integrates code snippets and technical insights extracted from scientific literature into a verifiable, executable representation.
Outcome:	The proposed knowledge base shows significant performance gains on paperBench when integrated into three agent frameworks with two different LLMs.

Copied to clipboard

Sai Srinivas Kancheti, Aditya Sanjiv Kanade, Vineeth N. Balasubramanian, Tanuja Ganu

Reviving Iterative Refinement in Diffusion-based NER with an Initializer-Restorer Approach (2026.acl-short)

Challenge:	Existing multimodal reasoning models lack generalized spatial intelligence, a new study shows . a critical gap exists in the field of vision-centric reasoning, the authors argue .
Approach:	They evaluate 16 multimodal reasoning models using Chain-of-Though (CoT) based thinking . they find that CoT prompting consistently degrades performance in visual spatial reasoning .
Outcome:	The proposed model hallucinates visual details from textual priors even when the image is absent.

Copied to clipboard

Long Hai Trieu, Phí Minh Hieu, Makoto Miwa

Protein-STORY: Semantic Text-Oriented Representation Yields biologically meaningful Protein embeddings (2026.acl-short)

Challenge:	Named Entity Recognition (NER) is a fundamental task in Information Extraction.
Approach:	They propose a generative paradigm for Named Entity Recognition by modeling NER as a boundary diffusion process.
Outcome:	The proposed model performs better than baseline on ACE2004, GENIA, and CleanCoNLL.

Copied to clipboard

Nabil Ibtehaz, Daisuke Kihara

Representation Learning

Challenge:	Unsupervised representation learning relying on sequence data often overlooks decades of expert-curated biological knowledge stored in textual formats.
Approach:	They propose a pipeline that synthesizes protein embeddings from diverse, multi-source text descriptions and a network architecture that integrates high-fidelity functional and structural insights into a unified representation.
Outcome:	The proposed pipeline outperforms existing models on diverse downstream tasks (+2 pts F1) and enables zero-shot text-prompted protein search.

Diving into the Decoding Space of Non-Autoregressive Models via Lexically Constrained Search (2026.acl-short)

Copied to clipboard

Chenyang Huang, Osmar Zaiane