DEEPSEEK

SenseRel: A Sense-Level Benchmark for Denotational and Connotational Meaning Relations (2026.acl-long)

Challenge:	Existing work on bias evaluation includes benchmark datasets and automated detection methods.
Approach:	They propose an open-source web application for side-by-side comparison of LLM outputs with real-time bias analysis.
Outcome:	The open-source application compares LLM outputs with real-time bias analysis.

Copied to clipboard

Pierluigi Cassotti, Naomi Baes, Stefano De Pascale, Jáder Martins Camboim de Sá, Francesco Periti, Nick Haslam, Dirk Geeraerts, Nina Tahmasebi

Measuring Linguistic Competence of LLMs on Indigenous Languages of the Americas (2026.eacl-short)

Challenge:	Polysemy enables a single word to convey multiple related meanings . a word's sense is extended to new contexts and concepts, a process called semantic change is gradual .
Approach:	They propose a benchmark for modeling semantic relations between word senses . they use a model that distinguishes denotational and connotationally related aspects of meaning .
Outcome:	The proposed model is able to distinguish between denotational and connotationalist aspects of meaning . it is compared with models with GPT-4o, Llama 3.1, and DeepSeek .

Copied to clipboard

Justin Vasselli, Arturo Martínez Peguero, Frederikus Hudi, Haruki Sakajo, Taro Watanabe

Challenge:	Existing benchmarks for linguistic knowledge of Indigenous languages of the Americas focus on high- and medium-resource languages with substantial digital presence.
Approach:	They propose a framework for probing large language models’ linguistic knowledge of Indigenous languages of the Americas using zero-shot prompting and few-shot probing.
Outcome:	The proposed framework evaluates models from five major families on 13 Indigenous languages including Bribri, Guarani, and Nahuatl.

Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyond (2025.acl-industry)

Copied to clipboard

Liang Wen, Yunke Cai, Fenrui Xiao, Xin He, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Tanglifu Tanglifu, Xiaowei Lv, Haosheng Zou, Yongchao Deng, Shousheng Jia, Xiangzheng Zhang

A Benchmark and Evaluation of Automated Language of Study Extraction from Computational Linguistics Publications (2026.eacl-srw)

Challenge:	Experimental results show that opensource curriculum training is more effective when distinct datasets are available for different training stages.
Approach:	They propose an opensource suite for training long reasoning models using publicdata and models.
Outcome:	The proposed model outperforms DeepSeek-R1-DistillQwen-32B models in math reasoning.

Copied to clipboard

Henry Gagnier, Ashwin Kirubakaran

The Evolution of Gen Alpha Slang: Linguistic Patterns and AI Translation Challenges (2025.acl-srw)

Challenge:	Language of study extraction is an aspect of computational linguistics papers that is useful for analyses of trends and diversity in computational linguists.
Approach:	They propose to benchmark and evaluate automated language of study extraction from computational linguistics papers.
Outcome:	The proposed language extraction benchmarks show that they can extract languages from papers with accuracy without high computational costs.

Copied to clipboard

Ishita Ishita, Radhika Mamidi

Machine Translation

Challenge:	Generation Alpha (born 2010-2024) exhibits unique linguistic behaviours influenced by rampant online communication and platform-specific cultures.
Approach:	They construct a comprehensive slang corpus from online platforms and evaluate four AI translation systems on over 100 sling terms.
Outcome:	The proposed translation systems outperform four existing translation models on over 100 slang terms.

DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models (2024.acl-long)

Copied to clipboard

Damai Dai, Chengqi Deng, Chenggang Zhao, R.x. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y. Wu, Zhenda Xie, Y.k. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, Wenfeng Liang

Culture Cartography: Mapping the Landscape of Cultural Knowledge (2025.emnlp-main)

Challenge:	Mixture-of-Experts (MoE) architectures face challenges in ensuring expert specialization . despite the promising performance, scaling language models to an extremely large scale is associated with exceedingly high computational costs.
Approach:	They propose an architecture that allows for ultimate expert specialization by segmenting experts into mN ones and activating mK from them.
Outcome:	The proposed architecture achieves comparable performance with GShard with 2B parameters and computation.

Copied to clipboard

Caleb Ziems, William Barr Held, Jane Yu, Amir Goldberg, David Grusky, Diyi Yang

Challenge:	Large Language Models (LLMs) can empower users to be more knowledgeable, productive, and creative, but their utility is often diminished for under-represented groups and cultures.
Approach:	They propose a methodology that operationalizes a mixed-initiative approach to finding culture-specific knowledge that is salient to in-group users but unknown to LLMs.
Outcome:	The proposed method improves the accuracy of LLMs on culturally-competent language models by 19.2%.

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

Copied to clipboard

Annabella Sakunkoo, Jonathan Sakunkoo

Multilinguality

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.

Bias in the East, Bias in the West: A Bilingual Analysis of LLM Political Bias on U.S.- and China-Related Issues (2026.findings-eacl)

Copied to clipboard

Ying Ying Lim, Paul Röttger

Evaluating AI for Finance: Is AI Credible at Assessing Investment Risk Appetite? (2025.emnlp-industry)

Challenge:	Large language models (LLMs) can exhibit political biases, which creates a risk of undue influence on LLM users and public opinion.
Approach:	They use a dataset of 36k real-time test prompts to measure LLM political bias on U.S. and Chinese issues.
Outcome:	The proposed model origin and prompt language influence bias on 60 political issues.

Copied to clipboard

Divij Chawla, Ashita Bhutada, Duc Anh Do, Abhinav Raghunathan, Vinod Sp, Cathy Guo, Dar Win Liew, Prannaya Gupta, Rishabh Bhardwaj, Rajat Bhardwaj, Soujanya Poria

From Reasoning to Answer: Empirical, Attention-Based and Mechanistic Insights into Distilled DeepSeek R1 Models (2025.emnlp-main)

Challenge:	Our analysis was conducted on proprietary systems and open-weight models . FINRISKEVAL analyzed 1,720 profiles spanning a broad spectrum of possible risk categories .
Approach:	They evaluated proprietary AI systems and open-weight models to assess investment risk appetite using carefully curated user profiles.
Outcome:	The proposed models exhibit significant variance when user attributes that should not influence risk computation are changed.

Copied to clipboard

Jue Zhang, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang

DevEval: A Manually-Annotated Code Generation Benchmark Aligned with Real-World Code Repositories (2024.findings-acl)

Challenge:	Large Reasoning Models generate explicit reasoning traces alongside final answers . the extent to which these traces influence answer generation remains unclear .
Approach:	They conduct empirical evaluation of Large Reasoning Models that include explicit reasoning . they also show that answer tokens attend substantially to reasoning tokens .
Outcome:	The results show that including explicit reasoning improves answer quality across domains . they also show that answer tokens attend substantially to reasoning tokens - the authors .

Copied to clipboard

Jia Li, Ge Li, Yunfei Zhao, Yongmin Li, Huanyu Liu, Hao Zhu, Lecheng Wang, Kaibo Liu, Zheng Fang, Lanshen Wang, Jiazheng Ding, Xuanming Zhang, Yuqi Zhu, Yihong Dong, Zhi Jin, Binhua Li, Fei Huang, Yongbin Li, Bin Gu, Mengfei Yang

Beyond Math: Stories as a Testbed for Memorization-Constrained Reasoning in LLMs (2026.eacl-long)

Challenge:	Existing benchmarks are poorly aligned with real-world code repositories and are insufficient to evaluate the coding abilities of Large Language Models (LLMs).
Approach:	They propose a repository-level benchmark named DevEval to evaluate LLMs' coding abilities in real-world code repositories.
Outcome:	The proposed benchmarks show that the LLMs perform better in real-world code repositories than existing benchmarks.

Copied to clipboard

Yuxuan Jiang, Francis Ferraro

Punctuations and Predicates in Language Models (2026.findings-eacl)

Challenge:	Large Language Models' (LLMs) performance is greatly inflated by memorization, a study finds . authors propose a framework to analyze how LLMs reason under different degrees of memory access.
Approach:	They propose a framework that allows large language models to reason under different degrees of memory access.
Outcome:	Evaluating GPT-4o, LLaMA3.3-70B, and DeepSeek V3 on character-centric story understanding benchmarks, they find up to a 45.2% accuracy drop under the Restrictive Setting.

Copied to clipboard

Sonakshi Chauhan, Maheep Chaudhary, Choy Kwan Kiu, Samuel Nellessen, Nandi Schoots

Elections go bananas: A First Large-scale Multilingual Study of Pluralia Tantum using LLMs (2026.eacl-long)

Challenge:	Recent work has shown that LLMs perform tasks in ways that diverge significantly from human reasoning.
Approach:	They examine the computational importance of punctuation tokens in large language models . they use zeroing and layer-swapping techniques to examine their necessity and sufficiency .
Outcome:	The proposed model differs in GPT-2, DeepSeek, and Gemma in that punctuation is necessary and sufficient in multiple layers . the findings offer new insight into the internal mechanisms of punctuations in LLMs and have implications for interpretability and model analysis.

Copied to clipboard

Elena Spaziani, Kamyar Zeinalipour, Pierluigi Cassotti, Nina Tahmasebi

Multilinguality

Challenge:	a large amount of annotated sentences for each feature can be used for in-depth analysis.
Approach:	They propose an annotation framework for lexicalization of pluralia tantum . they use an LLM to annotate each instance from the reference corpus .
Outcome:	The proposed framework provides useful annotators for semantic, syntactic and sense categories with accuracy ranging from 51% to 89% on a hand-annotated testset.

KodCode: A Diverse, Challenging, and Verifiable Synthetic Dataset for Coding (2025.findings-acl)

Copied to clipboard

Zhangchen Xu, Yang Liu, Yueqin Yin, Mingyuan Zhou, Radha Poovendran

NOVER: Incentive Training for Language Models via Verifier-Free Reinforcement Learning (2025.emnlp-main)

Challenge:	Existing code-focused resources typically fail to ensure either the breadth of coverage or verifiable correctness.
Approach:	They propose a synthetic dataset that provides high-quality, verifiable training data for Large Language Models for coding.
Outcome:	The proposed dataset surpasses Qwen2.5-Coder-32B-Instruct and DeepSeek-R1-Distill-Llama-70B in performance on coding benchmarks.

Copied to clipboard

Wei Liu, Siya Qi, Xinyu Wang, Chen Qian, Yali Du, Yulan He

Breaking the Reasoning Barrier A Survey on LLM Complex Reasoning through the Lens of Self-Evolution (2025.findings-acl)

Challenge:	Recent advances in reinforcement learning, such as DeepSeek R1-Zero, highlight the effectiveness of incentive training, but these methods rely on external verifiers, which limits their applicability to domains like mathematics and coding, where such verifier is readily available.
Approach:	They propose a general reinforcement learning framework that requires only standard supervised fine-tuning data with no need for an external verifier.
Outcome:	The proposed framework outperforms the model of the same size distilled from large reasoning models such as DeepSeek R1 671B by 7.7%.

Copied to clipboard

Tao He, Hao Li, Jingchang Chen, Runxuan Liu, Yixin Cao, Lizi Liao, Zihao Zheng, Zheng Chu, Jiafeng Liang, Ming Liu, Bing Qin

FRAME: Feedback-Refined Agent Methodology for Enhancing Medical Research Insights (2025.findings-acl)

Challenge:	OpenAI's O1 and subsequent projects like DeepSeek R1 have significantly advanced research on complex reasoning in LLMs.
Approach:	They analyze existing reasoning studies from the perspective of self-evolution and summarize O1-like works from open-source projects like DeepSeek R1 and Kimi-k1.5.
Outcome:	The proposed models are based on open-source models and pioneer advanced methodologies like Scaling Reinforcement Learning (RL).

Copied to clipboard

Chengzhang Yu, Yiming Zhang, Zhixin Liu, Zenghui Ding, Yining Sun, Zhanpeng Jin

TripleFact: Defending Data Contamination in the Evaluation of LLM-driven Fake News Detection (2025.acl-long)

Challenge:	Existing approaches to automate scientific research are limited by human cognitive constraints and timeintensive workflows.
Approach:	They propose a framework that enhances medical paper generation through iterative refinement and structured feedback.
Outcome:	The proposed framework achieves significant improvements over conventional methods across multiple models and evaluation dimensions.

Copied to clipboard

Cheng Xu, Nan Yan

Agents Under Siege: Breaking Pragmatic Multi-Agent LLM Systems with Optimized Prompt Attacks (2025.acl-long)

Challenge:	Existing evaluation paradigms for fake news detection are based on static datasets and closed-world assumptions that are inadvertently memorized during pre-training.
Approach:	They propose a framework to mitigate BDC risk while prioritizing real-world applicability by integrating three components to assess robustness against human-crafted misinformation.
Outcome:	The proposed framework mitigates BDC risk while prioritizing real-world applicability.

Copied to clipboard

Rana Shahroz, Zhen Tan, Sukwon Yun, Charles Fleming, Tianlong Chen

FrameNet-Cultures: A Benchmark for Evaluating LLMs via Cross-Cultural Frame Semantics (2026.findings-acl)

Challenge:	Multi-agent LLMs are prone to adversarial attacks because of constraints such as limited token bandwidth and latency between message delivery.
Approach:	They propose a permutation-invariant adversarial attack that optimizes prompt distribution across latency and bandwidth constraints to bypass distributed safety mechanisms within the system.
Outcome:	The proposed method outperforms conventional attacks by up to 7 on multiple models.

Copied to clipboard

Neda Jamshidi, Anders Søgaard, Monica Bianchini

Implicit Reasoning in Transformers is Reasoning through Shortcuts (2025.findings-acl)

Challenge:	Existing evaluation paradigms for large language models lack rigorous methods to evaluate cultural alignment . FRAMENET-CULTURES is an open-ended benchmark for evaluating cultural alignment in LLMs .
Approach:	They propose a benchmark for evaluating cultural alignment in large language models based on Fillmore-style frame semantics.
Outcome:	The proposed benchmark is based on Fillmore-style frame semantics.

Copied to clipboard

Tianhe Lin, Jian Xie, Siyu Yuan, Deqing Yang

HarmRLVR: Weaponizing Verifiable Rewards for Harmful LLM Alignment (2026.acl-long)

Challenge:	Language models can perform step-by-step reasoning and achieve high accuracy in both in-domain and out-of-domain tests via implicit reasoning.
Approach:	They train GPT-2 from scratch on a curated multi-step mathematical reasoning dataset and conduct analytical experiments to investigate how language models perform implicit reasoning in multi- step tasks.
Outcome:	The proposed model performs better on multi-step tasks than the explicit reasoning model.

Copied to clipboard

Yuexiao Liu, Lijun Li, Xingjun Wang, Jing Shao

LLM-SLM Collaborative Framework of Idiomatic Expression Generation (2026.acl-long)

Challenge:	Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have gained significant attention due to their objective and verifiably verifier reward signals.
Approach:	They propose to exploit RLVR for alignment reversibility by using GRPO to reverse alignment with merely 64 harmful prompts without responses.
Outcome:	The proposed method outperforms fine-tuning and RLHF in reasoning and code generation tasks while maintaining general capabilities.

Copied to clipboard

Hui Gao, Changhao Song, Peng Zhang, Jing Zhang, Chang Yang, Liuxian Ge

MMATH: A Multilingual Benchmark for Mathematical Reasoning (2025.findings-emnlp)

Challenge:	Existing methods for idiomatic expression generation lack parallel data and manual annotations.
Approach:	They propose an iterative LLM-SLM collaborative framework that replaces human supervision for idiomatic expression data generation.
Outcome:	The proposed framework outperforms DeepSeek-R1 in Chinese Idiom Polishing with a 25.2% improvement in accuracy.

Copied to clipboard

Wenyang Luo, Xin Zhao, Jing Sha, Shijin Wang, Ji-Rong Wen

KazMMLU: Evaluating Language Models on Kazakh, Russian, and Regional Knowledge of Kazakhstan (2025.acl-long)

Challenge:	a benchmark for multilingual complex reasoning spans 374 high-quality math problems across 10 typologically diverse languages.
Approach:	They propose a benchmark for multilingual complex reasoning across 10 languages . they show reasoning in English and answering in target languages can enhance performance .
Outcome:	The proposed benchmark demonstrates that models with high-quality reasoning can perform in multiple languages.

Copied to clipboard

Mukhammed Togmanov, Nurdaulet Mukhituly, Diana Turmakhan, Jonibek Mansurov, Maiya Goloburda, Akhmed Sakip, Zhuohan Xie, Yuxia Wang, Bekassyl Syzdykov, Nurkhan Laiyk, Alham Fikri Aji, Ekaterina Kochmar, Preslav Nakov, Fajri Koto

DiSCo: Device-Server Collaborative LLM-based Text Streaming Services (2025.findings-acl)

Challenge:	Kazakh language remains underrepresented in the field of natural language processing despite the country's population exceeding twenty million . however, there is a lack of dedicated models and benchmark evaluations specifically tailored to Kazakh languages.
Approach:	They propose to create a dataset specifically designed for Kazakh language with 23,000 questions sourced from authentic educational materials and manually validated by native speakers and educators.
Outcome:	The first MMLU-style dataset specifically designed for Kazakh language.

Copied to clipboard

Ting Sun, Penghan Wang, Fan Lai

Dropping Experts, Recombining Neurons: Retraining-Free Pruning for Sparse Mixture-of-Experts LLMs (2025.findings-emnlp)

Challenge:	Large language models (LLMs) have introduced significant cost and quality of experience (QoE) challenges in serving millions of daily requests.
Approach:	They propose a device-server cooperative scheduler that optimizes users’ QoE by adaptively routing requests and migrating response generation between endpoints while maintaining cost constraints.
Outcome:	Evaluations on real-world workloads show that the proposed scheduler can reduce tail TTFT (11-52%) and mean TTTT (6-78%) while maintaining comparable QoE levels.

Copied to clipboard

Yixiao Zhou, Ziyu Zhao, Dongzhou Cheng, Zhiliang Wu, Jie Gui, Yi Yang, Fei Wu, Yu Cheng, Hehe Fan

Would LLMs be Good Historical Linguists and Chinese Dialect Learners? (2026.acl-long)

Challenge:	Sparse Mixture-of-Experts (SMoE) architectures require loading all expert parameters . previous work focused on expert pruning and merging but focused on neuron-level structure .
Approach:	They propose a task-agnostic framework for expert pruning and reconstruction . it prunes redundant experts using router statistics, then decomposes them into neuron-level expert segments .
Outcome:	The proposed framework reduces the number of experts and memory usage, making it easier to deploy.

Copied to clipboard

Yicheng Liu, Shumin Shi, Youchao Zhou, Xingchen Zhang

Syntactic Text Processing

Challenge:	Large language models struggle with low-resource Chinese dialects due to substantial phonological divergence.
Approach:	They propose to incorporate Middle Chinese, the common historical ancestor of modern Chinese dialects, into LLMs to improve dialectal pronunciation modeling.
Outcome:	The proposed approach improves on standard Chinese but struggles with low-resource Chinese dialects . the proposed model improves over baselines while revealing variation across dialects.

EfficientXLang: Towards Improving Token Efficiency Through Cross-Lingual Reasoning (2025.findings-emnlp)

Copied to clipboard

Sanchit Ahuja, Praneetha Vaddamanu, Barun Patra

Multilinguality

Challenge:	Despite recent advances in Reasoning Language Models, most research focuses solely on English, even though many models are pretrained on multilingual data.
Approach:	They evaluate three open-source RLMs: DeepSeek R1, Qwen 2.5, and Qwend 3 across four math datasets and seven typologically diverse languages.
Outcome:	The proposed model reduces token usage and preserves accuracy even after translation into English.

Dynamics of Cognitive Heterogeneity: Investigating Behavioral Biases in Multi-Stage Supply Chains with LLM-Based Simulation (2026.acl-long)

Copied to clipboard

Jiuyun Jiang, Yuecheng Hong, Bo Yang, Jin Yang, Guangxin Jiang, Xiaomeng Guo, Guang Xiao

IS-CoT: Breaking the Long-form Generation Collapse via Interleaved Structural Thinking (2026.acl-long)

Challenge:	Recent advances in Large Language Models (LLMs) allow repeatable experiments in which individual characteristics can be precisely defined.
Approach:	They propose a scalable experimental paradigm using Large Language Models to simulate multi-stage supply chain dynamics.
Outcome:	The proposed model systematically replicates and validates the results of a behavioral simulation on agents in multi-stage supply chain dynamics.

Copied to clipboard

Zechen Sun, Yuyang Sun, Zecheng Tang, Juntao Li, Wenpeng Hu, Wenliang Chen, Zhunchen Luo, Guotong Geng, Min Zhang

Choosing a Model, Shaping a Future: Comparing LLM Perspectives on Sustainability and its Relationship with AI (2025.findings-emnlp)

Challenge:	Existing models with reasoning capabilities suffer from a severe length collapse in open-ended writing .
Approach:	They propose a framework that embeds a dynamic plan-write-reflect cycle into the generation process and train a model with interleaved reasoning traces.
Outcome:	The proposed framework achieves state-of-the-art performance on long-form benchmarks compared to other models on the same dataset.

Copied to clipboard

Annika Bush, Meltem Aksoy, Markus Pauly, Greta Ontrup

Can LLMs be Literary Companions?: Analysing LLMs on Bengali Figures of Speech Identification (2025.emnlp-main)

Challenge:	Claude, DeepSeek, GPT, LLaMA, and Mistral conceptualize sustainability and its relationship with AI.
Approach:	They investigated how five state-of-the-art LLMs conceptualize sustainability and its relationship with AI.
Outcome:	The proposed models differed significantly between models, revealing significant biases and biase . Claude, DeepSeek, GPT, LLaMA, and Mistral conceptualize sustainability and its relationship with AI.

Copied to clipboard

Sourav Das, Kripabandhu Ghosh

Abstain-R1: Calibrated Abstention and Post-Refusal Clarification via Verifiable RL (2026.findings-acl)

Challenge:	despite Bengali being among the most spoken languages, the NLP efforts on it remain limited.
Approach:	They present a dataset that includes Bengali figures of speech on six poets . they deploy state-of-the-art Large Language Models to the dataset and fine-tune the best models .
Outcome:	The proposed dataset reveals that two open-source LLMs perform better than others in Bengali . the framework can be reproduced for English and other low-resource languages .

Copied to clipboard

Haotian Zhai, Jingcheng Liang, Dongyeop Kang

Challenge:	Existing abstention methods produce generic refusals or encourage follow-up clarifications without verifying whether they identify the key missing information.
Approach:	They propose a clarification-aware RLVR reward that rewards correct answers on unanswerable queries while optimizing explicit abstention and semantically aligned post-refusal clarification on unannounced queries.
Outcome:	The proposed model improves abstention and clarification on unanswerable queries while maintaining strong performance on answerable queries.

Exposing Numeracy Gaps: A Benchmark to Evaluate Fundamental Numerical Abilities in Large Language Models (2025.findings-acl)

Copied to clipboard

Haoyang Li, Xuejia Chen, Zhanchao Xu, Darian Li, Nicole Hu, Fei Teng, Yiming Li, Luyu Qiu, Chen Jason Zhang, Li Qing, Lei Chen

AraReasoner: Evaluating Reasoning-Based LLMs for Arabic NLP (2025.findings-emnlp)

Challenge:	Existing benchmarks focus on linguistic competence or structured mathematical problem-solving, neglecting fundamental numerical reasoning required in real-world scenarios.
Approach:	They propose a benchmark to evaluate numerical capabilities for large language models . they use a dataset to assess number recognition, arithmetic operations, contextual retrieval, comparison, summary, and multi-step reasoning.
Outcome:	The proposed benchmark evaluates six fundamental numerical capabilities: number recognition, arithmetic operations, contextual retrieval, comparison, summary, and multi-step reasoning.

Copied to clipboard

Ahmed Abul Hasanaath, Aisha Alansari, Ahmed Ashraf, Salmane Chafik, Hamzah Luqman, Saad Ezzini

EvoCoT: Overcoming the Exploration Bottleneck in Reinforcement Learning for LLMs (2026.findings-acl)

Challenge:	Large language models have shown remarkable progress in reasoning abilities and general natural language processing tasks, yet their performance on Arabic data remains underexplored.
Approach:	They compare reasoning-focused LLMs with deepSeek models across 15 Arabic NLP tasks . they use zero-shot, few-shot and fine-tuning to evaluate their capacity for linguistic reasoning .
Outcome:	The proposed models outperform strong models on Arabic datasets and are compared with other models.

Copied to clipboard

Huanyu Liu, Jia Li, Yihong Dong, Chang Yu, Taozhi Chen, Lecheng Wang, Yongding Tao, Bin Gu, Ge Li

TENP: Trapezoidal Expert Neuron Pruning For Mixture-of-Experts (2026.findings-acl)

Challenge:	Existing approaches to reinforcement learning with verifiable reward (RLVR) are limited by difficulty or lack of exploration.
Approach:	They propose a self-evolving curriculum learning framework based on chain-of-thought reasoning optimization that constrains exploration space by self-generating and verifying CoT trajectories.
Outcome:	The proposed framework enables LLMs to solve previously unsolved problems without external supervision and is compatible with various RL fine-tuning methods.

Copied to clipboard

Jiangyang He, Shaolin Zhu, Deyi Xiong

Green & Sustainable NLP

Challenge:	Existing compression approaches remove entire experts, disrupting routing topology and harming performance, or rely on unstructured weight pruning with limited practical efficiency.
Approach:	They propose a structured Trapezoidal Expert Neuron Prunning framework that uses a trapezoidal pattern to identify and retain important experts while applying expert neuron pruning (ENP) to less important experts.
Outcome:	The proposed framework outperforms the full-parameter model by 10% on code generation tasks under a sparse activation of experts and a 40% routing sparsity.

When Punctuation Matters: A Large-Scale Comparison of Prompt Robustness Methods for LLMs (2025.findings-emnlp)

Copied to clipboard

Mikhail Seleznyov, Mikhail Chaichuk, Gleb Ershov, Alexander Panchenko, Elena Tutubalina, Oleg Somov

Model-Dependent Moderation: Inconsistencies in Hate Speech Detection Across LLM-based Systems (2025.findings-acl)

Challenge:	Large Language Models (LLMs) are sensitive to subtle, non-semantic variations in prompt phrasing and formatting.
Approach:	They propose to evaluate 4 methods for improving prompt robustness within a unified experimental framework.
Outcome:	The proposed methods are compared to 8 models from Llama, Qwen and Gemma families and are generalized against multiple types of distribution shifts.

Copied to clipboard

Neil Fasching, Yphtach Lelkes

Information Retrieval

Challenge:	Content moderation systems powered by large language models are increasingly deployed to detect hate speech . if two systems produce different outcomes for the same content, it undermines consistency and predictability .
Approach:	They analyze 1.3+ million sentences from a factorial design to determine hate speech classification . they find identical content receives markedly different classification values across systems .
Outcome:	The proposed model finds that identical content receives markedly different classification values across systems.

Un-considering Contextual Information: Assessing LLMs’ Understanding of Indexical Elements (2025.findings-acl)

Copied to clipboard

Metehan Oğuz, Yavuz Faruk Bakman, Duygu Nur Yaldiz

Information Extraction & Text Mining

Challenge:	Large Language Models (LLMs) excel in coreference resolution tasks, but previous studies only assessed performance with nouns and third person pronouns.
Approach:	They evaluate LLMs' performance on coreference resolution with indexicals like I, you, here and tomorrow which come with unique challenges due to their linguistic properties.
Outcome:	The proposed models perform well with some indexicals while struggling with others.

Temporal Consistency for LLM Reasoning Process Error Identification (2025.findings-emnlp)

Copied to clipboard

Jiacheng Guo, Yue Wu, Jiahao Qiu, Kaixuan Huang, Xinzhe Juan, Ling Yang, Mengdi Wang

HelpSteer3: Human-Annotated Feedback and Edit Data to Empower Inference-Time Scaling in Open-Ended General-Domain Tasks (2025.acl-long)

Challenge:	Empirical evaluations show consistent performance improvements over baseline methods . 7B/8B distilled models outperform all 70B/72B models and GPT-4o on ProcessBench .
Approach:	They propose a temporal consistency method that leverages consistency in a sequence of self-reflection actions to improve verification accuracy.
Outcome:	The proposed method outperforms existing methods on three benchmarks . it leverages consistency in a sequence of self-reflection actions to improve accuracy .

Copied to clipboard

Zhilin Wang, Jiaqi Zeng, Olivier Delalleau, Daniel Egert, Ellie Evans, Hoo-Chang Shin, Felipe Soares, Yi Dong, Oleksii Kuchaiev

Challenge:	Inference-Time Scaling is critical to the success of recent models such as OpenAI o1 and DeepSeek R1 . however, many techniques require tasks to have answers that can be verified .
Approach:	They use data to train dedicated Feedback and Edit Models capable of inference-time scaling for open-ended tasks.
Outcome:	The proposed model can reach SoTA performance on Arena Hard at 92.7 as of 5 Mar 2025.

CulturalBench: A Robust, Diverse and Challenging Benchmark for Measuring LMs’ Cultural Knowledge Through Human-AI Red-Teaming (2025.acl-long)

Copied to clipboard

Yu Ying Chiu, Liwei Jiang, Bill Yuchen Lin, Chan Young Park, Shuyue Stella Li, Sahithya Ravi, Mehar Bhatia, Maria Antoniak, Yulia Tsvetkov, Vered Shwartz, Yejin Choi

Robustness in NLP

Challenge:	CulturalBench is a set of 1,696 human-written and human-verified questions to assess LMs’ cultural knowledge covering 45 global regions including underrepresented ones like Bangladesh, Zimbabwe, and Peru.
Approach:	They construct a set of 1,696 human-written and human-verified questions to assess LMs' cultural knowledge, covering 45 global regions including underrepresented ones like Bangladesh, Zimbabwe, and Peru.
Outcome:	The proposed model outperforms other models across cultures, while underperforming on questions related to North Africa, South America and Middle East.

Understanding the Dark Side of LLMs’ Intrinsic Self-Correction (2025.acl-long)

Copied to clipboard

Qingjie Zhang, Di Wang, Haoting Qian, Yiming Li, Tianwei Zhang, Minlie Huang, Ke Xu, Hewu Li, Liu Yan, Han Qiu

Challenge:	Recent studies show that LLMs’ intrinsic self-correction fails without oracle labels as feedback.
Approach:	They propose to use one simple task and three complex tasks with state-of-the-art LLMs like ChatGPT, Llama, and DeepSeek to interpret LLM's intrinsic self-correction.
Outcome:	The proposed methods reveal the dark side of LLMs’ intrinsic self-correction for different tasks, especially for those failure cases.

POW: Political Overton Windows of Large Language Models (2025.findings-emnlp)

Copied to clipboard

Leif Azzopardi, Yashar Moshfeghi

Improving the Language Understanding Capabilities of Large Language Models Using Reinforcement Learning (2025.findings-emnlp)

Challenge:	Political bias in Large Language Models (LLMs) presents a growing concern for the responsible deployment of AI systems.
Approach:	They propose to use the Overton Window as a framework to map the ideological boundaries that a given LLM will espouse, remain neutral on, or refuse to endorse.
Outcome:	The proposed methodology reveals political bias in large language models by examining the political stances of models from eight providers.

Copied to clipboard

Sai Ashish Somayajula, Bokai Hu, Qi Cao, Xin Pan, Pengtao Xie

CARO: Chain-of-Analogy Reasoning Optimization for Robust Content Moderation (2026.findings-acl)

Challenge:	Instruction-fine-tuned large language models (LLMs) under 14B parameters underperform on NLU tasks . we explore a framework to improve the NLU capabilities of LLMs .
Approach:	They propose to use Proximal Policy Optimization to improve NLU capabilities . they frame NLU as a reinforcement learning environment and optimize for reward signals .
Outcome:	The proposed framework outperforms supervised fine-tuning on GLUE and superGLUE tasks.

Copied to clipboard

Bingzhe Wu, Haotian Lu, Yuchen Mou

Do All Autoregressive Transformers Remember Facts the Same Way? A Cross-Architecture Analysis of Recall Mechanisms (2025.emnlp-main)

Challenge:	Current large language models struggle with ambiguous content moderation cases due to misleading "decision shortcuts" . authors propose a two-stage training framework to induce robust analogical reasoning in LLMs .
Approach:	They propose a two-stage training framework to induce robust analogical reasoning in LLMs . they bootstrap analogy reasoning chains via retrieval-augmented generation and SFT .
Outcome:	The proposed framework outperforms state-of-the-art reasoning models and specialized moderation models on ambiguous moderation benchmarks.

Copied to clipboard

Minyeong Choe, Haehyun Cho, Changho Seo, Hyunil Kim

Large Language Models Require Curated Context for Reliable Political Fact-Checking—Even with Reasoning and Web Search (2026.findings-acl)

Challenge:	Prior work has identified MLP modules in early layers as key contributors to factual recall.
Approach:	They conduct a comprehensive evaluation of factual recall across several models . they find that Qwen-based models behave differently from previous patterns .
Outcome:	The results show that attention modules in the earliest layers contribute more to factual recall than MLP modules.

Copied to clipboard

Matthew R. DeVerna, Kai-Cheng Yang, Harry Yaojun Yan, Filippo Menczer

It’s Not a Walk in the Park! Challenges of Idiom Translation in Speech-to-text Systems (2025.acl-long)

Challenge:	Large language models (LLMs) have raised hopes for automated end-to-end fact-checking, but prior studies report mixed results.
Approach:	They evaluate 15 large language models on 6,000 claims fact-checked by PolitiFact . standard models perform poorly, reasoning offers minimal benefits, and web search provides only moderate gains .
Outcome:	The models predict claim veracity and a curated RAG system improved macro F1 by 233% on average across model variants.

Copied to clipboard

Iuliia Zaitova, Badr M. Abdullah, Wei Xue, Dietrich Klakow, Bernd Möbius, Tania Avgustinova

Text Generation

Challenge:	idioms are defined as words with a figurative meaning not deducible from their individual components.
Approach:	They compare idiom translation as compared to conventional news translation in two languages . they compare MT and SLT systems with MT, Large Language Models and cascaded alternatives .
Outcome:	The proposed systems show better handling of idioms than standard news translation systems.

RSMeM: Knowledge-Enhanced Memory Evolution for Remote Sensing Agents with Systematic Evaluation (2026.acl-long)

Copied to clipboard

Bingxian Wu, Yu Zhang, Zonghao Guo, Tang Liu, Chen Qian, Yuxiang Lu, Xingbo Du, Yanghao Li, Yidan Zhang, Chi Chen, Ling Yao, Chenghu Zhou, Maosong Sun

RePanda: Pandas-powered Tabular Verification and Reasoning (2025.acl-long)

Challenge:	Existing RS agents built on general-purpose LLMs are domain-agnostic, resulting in brittle and error-prone workflows.
Approach:	They propose a knowledge-enhanced memory evolution mechanism that bootstraps RS agents with pre-distilled domain knowledge and iteratively integrates online experience for robust multi-step tool execution.
Outcome:	Experiments show that the new model improves tool-use performance and accuracy . iteratively, iteration of the model integrates online experience for robust multi-step tool execution .

Copied to clipboard

Atoosa Chegini, Keivan Rezaei, Hamid Eghbalzadeh, Soheil Feizi

Towards Economical Inference: Enabling DeepSeek’s Multi-Head Latent Attention in Any Transformer-based LLMs (2025.acl-long)

Challenge:	Existing methods for fact-checking tabular data rely on black-box models with opaque reasoning.
Approach:	They propose a structured fact verification approach that translates claims into executable pandas queries.
Outcome:	The proposed method outperforms existing methods and demonstrates strong OOD robustness.

Copied to clipboard

Tao Ji, Bin Guo, Yuanbin Wu, Qipeng Guo, Shenlixing Shenlixing, Chenzhan Chenzhan, Xipeng Qiu, Qi Zhang, Tao Gui

Analyzing and Modeling LLM Response Lengths with Extreme Value Theory: Anchoring Effects and Hybrid Distributions (2025.emnlp-main)

Challenge:	Multi-head Latent Attention (MLA) is an innovative architecture designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector.
Approach:	They propose a data-efficient fine-tuning method for transitioning from MHA to MLA using a latent vector cache.
Outcome:	The proposed architecture reduces the KV cache size of Llama2-7B by 92.19%, with only 1% drop in LongBench performance.

Copied to clipboard

Liuxuan Jiao, Chen Gao, Yiqian Yang, Chenliang Zhou, YiXian Huang, Xinlei Chen, Yong Li

Interpretable Traces, Unexpected Outcomes: Investigating the Disconnect in Trace-Based Knowledge Distillation (2026.acl-long)

Challenge:	Existing approaches treat length as an incidental output property rather than a statistically regular phenomenon worthy of rigorous modeling.
Approach:	They propose a statistical framework for modeling and controlling large language model response lengths using extreme value theory and cross-validation on Qwen and DeepSeek architectures.
Outcome:	The proposed model improves tail fit and generalizability while maintaining generalizzability.

Copied to clipboard

Siddhant Bhambri, Upasana Biswas, Subbarao Kambhampati