Conference on Empirical Methods in Natural Language Processing · 593 papers
Constructions are Revealed in Word Distributions
Joshua Rozner (Stanford University), Cory Shain (University of Texas at Austin)
CodeRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Utilizes the pre-trained language model RoBERTa through masking and intervention to calculate global affinity (probability of a word and its context) and local affinity (distributional differences between word pairs), thereby identifying syntactic constructions and their internal interactions.
Continuous-Time Attention: PDE-Guided Mechanisms for Long-Sequence Transformers
Yukun Zhang (Chinese University of Hong Kong), Xueqing Zhou (Fudan University)
CodeClassificationGenerationTransformerText
🎯 What it does: Propose a Continuous-Time Attention framework, integrating partial differential equations (PDEs) (diffusion, wave, reaction-diffusion) driven continuous-time dynamics into the self-attention mechanism of Transformers, enabling attention weights to evolve over pseudo-time.
Controllable Memorization in LLMs via Weight Pruning
Chenjie Ni (Northwestern University), Yanfu Zhang (William and Mary)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a gradient-based weight pruning framework for controllably adjusting the memory rate of large language models, capable of both suppressing and amplifying the model's memory of training data.
Zihao Zhao (Johns Hopkins University), Anjalie Field (Johns Hopkins University)
CodeGenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBiomedical DataElectronic Health Records
🎯 What it does: Proposed a privacy-preserving synthetic text generation method based on entity-aware control codes, combining in-context learning and prefix tuning variants.
CoPL: Collaborative Preference Learning for Personalizing LLMs
Youngbin Choi (POSTECH), Dongwoo Kim (POSTECH)
CodeRecommendation SystemGraph Neural NetworkLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose CoPL, which combines graph-structured collaborative filtering (GCF) and MoLE expert models to personalize LLMs, achieving user preference learning under sparse annotations.
CopySpec: Accelerating LLMs with Speculative Copy-and-Paste
Razvan-Gabriel Dumitru (University of Arizona), Mihai Surdeanu (University of Arizona)
CodeComputational EfficiencyLarge Language ModelText
🎯 What it does: This paper proposes a method called CopySpec, which can automatically identify and copy repeated token sequences in the context during LLM inference, thereby reducing unnecessary computations.
🎯 What it does: Proposed the COURTREASONER benchmark, evaluating LLMs' ability in complete judicial reasoning using expert-annotated US court opinions;
CoVoGER: A Multilingual Multitask Benchmark for Speech-to-text Generative Error Correction with Large Language Models
Zhengdong Yang (Kyoto University), Chenhui Chu (Kyoto University)
CodeRecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkAudio
🎯 What it does: Constructed a multilingual, multitask speech-to-text generative error correction benchmark named CoVoGER, and conducted a systematic evaluation of large language models on this benchmark.
Creativity in LLM-based Multi-Agent Systems: A Survey
Yi-Cheng Lin (National Taiwan University), Yun-Nung Chen (National Taiwan University)
CodeGenerationTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextReview/Survey Paper
🎯 What it does: This paper provides a systematic review of creativity in large language model (LLM)-driven multi-agent systems, proposing dimensions such as agent proactivity, personality configuration, creative generation techniques, and evaluation methods, and offering a unified framework and future research directions.
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: Proposed the CRDIAL framework, decomposing cognitive restructuring (CR) into two-stage multi-round dialogues, integrating emotional support with multi-channel loop mechanisms, and generating a high-quality bilingual dialogue dataset CRISP based on GPT-4o. Subsequently, trained the Qwen-2.5-7B/14B models CRISPERS to achieve human-computer interactive psychotherapy.
CRITICTOOL: Evaluating Self-Critique Capabilities of Large Language Models in Tool-Calling Error Scenarios
Shiting Huang (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
CodeTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Designed and constructed the CRITICTOOL evaluation benchmark to fine-grainedly assess large language models' self-criticism capabilities in tool call error scenarios (identifying errors, analyzing errors, correcting errors, retrying/skipping/terminating).
CTCC: A Robust and Stealthy Fingerprinting Framework for Large Language Models via Cross-Turn Contextual Correlation Backdoor
Zhenhua Xu (Zhejiang University), Meng Han (Zhejiang University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed and implemented a backdoor fingerprint framework triggered by cross-turn dialogue context relevance for verifying the ownership of large language models in black-box environments.
D-CoDe: Scaling Image-Pretrained VLMs to Video via Dynamic Compression and Question Decomposition
Yiyang Huang (Northeastern University), Yun Fu (Northeastern University)
CodeComputational EfficiencyConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose the D-CoDe framework, leveraging dynamic compression and problem decomposition to transfer image-pretrained VLM without training to video tasks;
DAMON: A Dialogue-Aware MCTS Framework for Jailbreaking Large Language Models
Xu Zhang (Wangxuan Institute of Computer Technology Peking University), Xiaojun Wan (Wangxuan Institute of Computer Technology Peking University)
CodeAdversarial AttackTextSequential
🎯 What it does: Propose DAMON, a multi-round dialogic jailbreaking framework based on Monte Carlo Tree Search (MCTS), which automatically generates sub-instruction sequences to bypass the safety alignment of large language models (LLMs);
🎯 What it does: Propose a framework named DAQu that utilizes multi-table metadata from relational databases to generate incremental representations for retrieval queries, addressing the issue of insufficient information in short queries.
DCR: Quantifying Data Contamination in LLMs Evaluation
Cheng Xu (University College Dublin), Tahar Kechadi
CodeExplainability and InterpretabilityLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and validated the Data Corruption Risk (DCR) framework to detect and quantify the degree of data contamination in large language models during benchmark evaluations, and adjusted model accuracy using the DCR Factor.
DDO: Dual-Decision Optimization for LLM-Based Medical Consultation via Multi-Agent Collaboration
Zhihao Jia (Central South University), Jianxin Wang (Central South University)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIBiomedical DataElectronic Health Records
🎯 What it does: Proposed the DDO (Dual-Decision Optimization) framework, which separates and optimizes two decision-making processes in medical consultations: symptom inquiry and disease diagnosis, through multi-agent collaboration.
DEBATE, TRAIN, EVOLVE: Self‐Evolution of Language Model Reasoning
Gaurav Srivastava (Virginia Tech), Xuan Wang (Virginia Tech)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Proposes the DEBATE, TRAIN, EVOLVE (DTE) framework, which employs self-generated reasoning trajectories from multi-agent debates to conduct self-evolution training on a single language model.
DEL-ToM: Inference-Time Scaling for Theory-of-Mind Reasoning via Dynamic Epistemic Logic
Yuheng Wu (Stanford University), Zhaozhuo Xu (Stevens Institute of Technology)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the DEL-ToM framework, which structurally and verifiably reasons about belief updates in Theory-of-Mind (ToM) tasks during inference by scaling the reasoning time of large language models through Dynamic Epistemic Logic (DEL).
Demystifying Domain-adaptive Post-training for Financial LLMs
Zixuan Ke (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
CodeDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkFinance Related
🎯 What it does: Proposed the FINDAP framework for systematic post-training of LLMs in the financial domain, comprising FinCap (core capabilities), FinRec (joint CPT+IT with preference alignment), FinTrain (fine-grained dataset), and FinEval (comprehensive evaluation).
Diagnosing Memorization in Chain-of-Thought Reasoning, One Token at a Time
Huihan Li (University of Southern California), Xiang Ren (University of Southern California)
CodeExplainability and InterpretabilityTransformerLarge Language ModelTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the STIM (Source-aware Token-level Identification of Memorization) framework for fine-grained evaluation of memorization levels at the token level in chain-of-thought (CoT) reasoning, assessing each generated token's memorization degree from three sources (local, mid-range, long-range), and using these memorization scores to predict erroneous tokens.
Dialect-SQL: An Adaptive Framework for Bridging the Dialect Gap in Text-to-SQL
Jie Shi (Fudan University), Wei Wang (Fudan University)
CodeGenerationAI Code AssistantTransformerLarge Language ModelTextTabularRetrieval-Augmented Generation
🎯 What it does: Proposes an adaptive framework called Dialect-SQL that uses ORM code as an intermediate language to bridge text-to-SQL tasks across different SQL dialects.
DICE: Structured Reasoning in LLMs through SLM-Guided Chain-of-Thought Correction
Yiqi Li (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Propose a lightweight framework called DICE, which first lets a large model (LLM) generate natural language answers, then uses a small model (SLM) to analyze and correct the answers through chain-of-thought (CoT), ultimately outputting answers in a structured format.
DiNaM: Disinformation Narrative Mining with Large Language Models
Witold Sosnowski (Polish-Japanese Academy of Information Technology), Adam Wierzbicki (Polish-Japanese Academy of Information Technology)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed the DiNaM algorithm, which detects, verifies, and refines misinformation in fact-checking articles using LLMs, then generates misinformation narratives via an embedding + clustering approach.
PeiFeng Wang, Shafiq Joty (Salesforce AI Research)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Trained large-scale base evaluation models (8B, 12B, 70B) through direct preference optimization (DPO) to learn three types of evaluation tasks: chain-of-thought (CoT) criticism, standard judgment, and response reasoning.
Direct Value Optimization: Improving Chain-of-Thought Reasoning in LLMs with Refined Values
Hongbo Zhang (Zhejiang University), Yue Zhang (Southern University of Science and Technology)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose Direct Value Optimization (DVO), which directly optimizes the reasoning path of large language models by using step-level value signals during the Chain-of-Thought (CoT) process;
Discrepancy Detection at the Data Level: Toward Consistent Multilingual Question Answering
Lorena Calvo-Bartolomé (Universidad Carlos III de Madrid), Jordan Lee Boyd-Graber
CodeAnomaly DetectionData-Centric LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a four-stage LLM-assisted pipeline named MIND for detecting factual and cultural differences in multilingual knowledge bases, first aligning documents via a multilingual topic model, then generating questions, retrieving evidence, generating answers, and determining differences.
DivScore: Zero-Shot Detection of LLM-Generated Text in Specialized Domains
Zhihui Chen (National University of Singapore), Mengling Feng (National University of Singapore)
CodeDomain AdaptationAnomaly DetectionKnowledge DistillationTransformerLarge Language ModelTextBiomedical Data
🎯 What it does: Proposes a zero-shot detection framework called DivScore, leveraging normalized entropy and cross-entropy scores, and constructing domain-adapted language models through unsupervised domain knowledge distillation, specifically for detecting LLM-generated text in medical and legal domains.
DIWALI - Diversity and Inclusivity aWare cuLture specific Items for India: Dataset and Assessment of LLMs for Cultural Text Adaptation in Indian Context
Pramit Sahoo (Indian Institute of Technology Hyderabad), Maunendra Sankar Desarkar (Indian Institute of Technology Hyderabad)
CodeDomain AdaptationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the DIWALI project, which is specific to multi-regional Indian culture, and used it to evaluate the performance of LLMs in cultural text adaptation tasks.
DMDTEval: An Evaluation and Analysis of LLMs on Disambiguation in Multi-domain Translation
Zhibo Man (Beijing Jiaotong University), Jinan Xu (Beijing Jiaotong University)
CodeDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a test set for ambiguous words in multi-domain translation and designed multiple prompt-based strategies to systematically evaluate the ambiguity resolution capabilities of large language models (LLMs) in multi-domain translation; subsequently evaluated five open-source LLMs across four language pairs and thirteen domains, investigating the impact of prompt strategies and domain knowledge on translation performance and ambiguity resolution.
Do LLMs Encode Frame Semantics? Evidence from Frame Identification
Jayanth Krishna Chundru (University of Cincinnati), Tianyu Jiang (University of Cincinnati)
CodeRecognitionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Investigate whether large language models (LLMs) have internalized frame semantic knowledge, and perform frame identification tasks on FrameNet through prompting and fine-tuning.
Do You Know About My Nation? Investigating Multilingual Language Models’ Cultural Literacy Through Factual Knowledge
Eshaan Tanwar (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed a cross-lingual, cross-cultural literacy evaluation dataset called XNationQA, and evaluated multiple multilingual large language models (LLMs) on it
DocAgent: An Agentic Framework for Multi-Modal Long-Context Document Understanding
Li Sun (Boston University), Chenyu You (Boston University)
CodeRetrievalTransformerLarge Language ModelAgentic AIMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Built a multi-modal long document understanding framework called DocAgent, which uses hierarchical outlines to guide LLM agents for efficient retrieval, followed by a review agent to verify answers and a memory module to enable cross-task knowledge transfer.
Dovetail: A CPU/GPU Heterogeneous Speculative Decoding for LLM inference
Libo Zhang (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
CodeComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Implementing the lossless inference acceleration method Dovetail in a CPU/GPU heterogeneous environment, combined with a draft-verification mechanism for LLM inference acceleration.
DPED: Multi-Layer Noise Distillation for Privacy-Preserving Text Embeddings
Shuya Feng (University of Alabama at Birmingham), Yuan Hong (University of Connecticut)
CodeSafty and PrivacyKnowledge DistillationRepresentation LearningText
🎯 What it does: Proposed the DPED framework, which achieves differential privacy text embedding training by utilizing teacher-student distillation and multi-layer noise injection
DSCD: Large Language Model Detoxification with Self-Constrained Decoding
Ming Dong (Central China Normal University), Tingting He (Central China Normal University)
CodeSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposed a LLM detoxification method called DSCD based on self-constrained decoding, which utilizes token-level toxicity layers to locate and dynamically adjust the distribution of the next token.
🎯 What it does: Propose the DyePack framework, which embeds multiple backdoor samples into public benchmark test sets to detect whether LLMs used the test set during training without requiring access to internal model information.
Dynamic Collaboration of Multi-Language Models based on Minimal Complete Semantic Units
Chao Hao (Great Bay University), Zitong Yu (Great Bay University)
CodeComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper studies a token-level collaborative inference method for multiple models based on the dynamic selection strategy (DDS) using minimum complete semantic units (MCSU) and distribution distance.
E2LLM: Encoder Elongated Large Language Models for Long-Context Understanding and Reasoning
Zihan Liao (East China Normal University), Wei Zhang (East China Normal University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Chunk long texts, compress each chunk into soft prompts using a pre-trained text encoder, align them to the decoder's input space via an adapter, and complete understanding and reasoning tasks through the decoder.
EasyRec: Simple yet Effective Language Models for Recommendation
Xubin Ren (University of Hong Kong), Chao Huang (University of Hong Kong)
CodeRecommendation SystemTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes EasyRec, a recommendation framework that combines language models with collaborative filtering, achieving excellent performance in both text zero-shot recommendation and text-enhanced collaborative filtering scenarios.
EduAdapt: A Question Answer Benchmark Dataset for Evaluating Grade-Level Adaptability in LLMs
Numaan Naeem (Mohamed bin Zayed University of Artificial Intelligence), Muhammad Abdul-Mageed (University of British Columbia)
CodeLarge Language ModelTextBenchmark
🎯 What it does: Proposed the EDUADAPT benchmark, covering all K-12 education stages, nine subjects, and nearly 50,000 QA pairs, to evaluate the grade-level adaptability of large language models.
Efficient Beam Search for Large Language Models Using Trie-Based Decoding
Brian J Chan (National Chengchi University), Hen-Hsen Huang (Academia Sinica)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a parallel beam search decoding method based on a prefix tree, significantly reducing memory usage while maintaining generation quality through shared KV cache.
Efficient Model Development through Fine-tuning Transfer
Pin-Jie Lin (Virginia Tech), Tu Vu (Virginia Tech)
CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method to transfer fine-tuning updates (diff vector) from old model versions to new model versions, achieving significant performance improvements without additional training; meanwhile, it explores the feasibility of this method in scenarios such as multilingual model development and iterative upgrades.
EnAnchored-X2X: English-Anchored Optimization for Many-to-Many Translation
Sen Yang (Nanjing University), Shanbo Cheng (ByteDance Research)
CodeGenerationData SynthesisOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: By leveraging English parallel data and the en2x capabilities of LLMs, this paper proposes an English-anchored x2x translation generation and evaluation framework, generating and filtering synthetic data while performing preference optimization to enhance the performance of LLMs in multilingual x2x translation.
Enhancing Chinese Offensive Language Detection with Homophonic Perturbation
Junqi Wu (South China University Of Technology), Wu Wei (South China University Of Technology)
CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Constructed a large-scale homophonic Chinese hate speech detection dataset HED-COLD, and proposed a homophone-aware pre-training and fine-tuning framework to enhance the model's robustness against homophonic attacks.
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Designed and implemented a complete pipeline from preprocessing to post-processing, generating high-quality, ultra-detailed image captions, and leveraging these data to significantly enhance the performance of large vision-language models.
Enhancing LLM Language Adaption through Cross-lingual In-Context Pre-training
Linjuan Wu (Zhejiang University), Weiming Lu (Zhejiang University)
CodeDomain AdaptationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented Cross-lingual In-context Pre-training (CrossIC-PT), which enhances the cross-lingual transfer capability of large language models (LLMs) by interleaving semantically related bilingual texts into a single context window through continuous pre-training.
ESC-Judge: A Framework for Comparing Emotional Support Conversational Agents
Navid Madani (University at Buffalo), Rohini Srihari
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the ESC-Judge framework, which conducts automated, theory-driven evaluations using Hill's E-I-A emotional support model dialogue system.
ESGenius: Benchmarking LLMs on Environmental, Social, and Governance (ESG) and Sustainability Knowledge
Chaoyue He (Alibaba-NTU Global e-Sustainability CorpLab), Chunyan Miao (Alibaba-NTU Global e-Sustainability CorpLab)
CodeTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the ESGenius benchmark, comprising 1,136 expert-validated multiple-choice questions (ESGenius-QA) and a corpus of 231 authoritative ESG documents (ESGenius-Corpus), and evaluated the zero-shot and retrieval-augmented (RAG) performance of 50 LLMs based on this benchmark.
Evaluating Language Translation Models by Playing Telephone
Syeda Jannatus Saba (Stony Brook University), Steven Skiena (Stony Brook University)
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose an unsupervised "telephone game" method that generates training data with semantic drift by translating multiple times between the source and target languages, used to train translation evaluation models.
Chumeng Liang (University of Illinois Urbana-Champaign), Jiaxuan You (University of Illinois Urbana-Champaign)
CodeTransformerLarge Language ModelGraph
🎯 What it does: Proposed DiagramEval, a fine-grained and explainable evaluation framework specifically designed to assess scientific diagrams generated by large language models (LLMs);
CodeAnomaly DetectionSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose EverTracer, which achieves stealthy and robust model fingerprint verification by leveraging gray-box access to LLMs through probabilistic variation signals derived from modified membership inference attacks.
Evolving Chinese Spelling Correction with Corrector-Verifier Collaboration
Linfeng Liu (Shanghai Jiao Tong University), Hai Zhao (Shanghai Jiao Tong University)
CodeComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the Automated Error Correction Iteration (ACI) framework, which generates adaptive training data through collaboration between a BERT error correction model and an LLM validator, achieving Chinese spelling error correction improvement without human annotation.
Examining False Positives under Inference Scaling for Mathematical Reasoning
Yu Wang (University of Science and Technology of China), Fuli Feng (University of Science and Technology of China)
CodeLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Systematically evaluate the occurrence of 'false positives' in large language models during mathematical reasoning tasks, covering various models, datasets, decoding strategies, and reasoning scaling methods;
Expanding before Inferring: Enhancing Factuality in Large Language Models through Premature Layers Interpolation
Dingwei Chen (Sun Yat-Sen University), Chengming Li (Shenzhen MSU-BIT University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposes a training-free, plug-and-play method called 'Premature Layers Interpolation (PLI)', which inserts temporary layers generated via spherical linear interpolation between existing layers in large language models to expand model depth and improve factual accuracy.
ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance
Sijia Yao (Northeastern University), Ge Yu (Tsinghua University)
CodeRetrievalTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Enhance retrieval effectiveness by jointly training large language models (LLM) with dense retrievers, leveraging semantically rich query expansions generated by the LLM.
Exploring morphology-aware tokenization: A case study on Spanish language modeling
Alba Táboas García (Universitat Pompeu Fabra), Leo Wanner (Barcelona Supercomputing Center)
CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes combining morphology segmentation with subword tokenization by using MorphAGram to perform semi-supervised morphology segmentation on Spanish. The segmentation results are then used to train a BPE tokenizer, which is applied in the RoBERTa pretraining model to verify its improvements in language modeling and downstream tasks.
Exploring Quality and Diversity in Synthetic Data Generation for Argument Mining
Jianzhu Bao (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
CodeData SynthesisLarge Language ModelText
🎯 What it does: Propose two synthetic data generation schemes based on LLM: quality-oriented (QOS) and diversity-oriented (DOS), to enhance the training of argument mining models.
Exploring Response Uncertainty in MLLMs: An Empirical Evaluation under Misleading Scenarios
Yunkai Dang (Hong Kong University of Science and Technology), Xuming Hu (Hong Kong University of Science and Technology)
CodeExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningPrompt EngineeringMultimodalityBenchmark
🎯 What it does: Evaluate the response uncertainty of multimodal large language models when encountering misleading prompts and construct the multimodal uncertainty benchmark MUB.
Extending Automatic Machine Translation Evaluation to Book-Length Documents
Kuang-Da Wang (National Yang Ming Chiao Tung University), Boris Ginsburg (National Yang Ming Chiao Tung University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a SEGALE evaluation framework that extends existing sentence-level automatic MT evaluation to long documents and even book-level assessments;
Extracting and Combining Abilities For Building Multi-lingual Ability-enhanced Large Language Models
Zhipeng Chen (Renmin University of China), Ji-Rong Wen (Renmin University of China)
CodeTransformerLarge Language ModelText
🎯 What it does: This paper proposes a multilingual capability extraction and combination method called MAEC, which can extract weights related to specific high-level capabilities (such as mathematical reasoning, scientific reasoning) using capability-related corpora in a single language (English), and then transfer these weights to multilingual LLMs through simple addition and subtraction operations, thereby enhancing their cross-lingual high-level reasoning capabilities.
Extracting Linguistic Information from Large Language Models: Syntactic Relations and Derivational Knowledge
Tsedeniya Kinfe Temesgen (Technische Universität München), Alexander Fraser (Technische Universität München)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Conducted cross-lingual linguistic knowledge extraction experiments on GPT-4o and LLaMA, designing three diagnostic tasks: syntax role labeling at the sentence level, derivational morphology decomposition at the word level, and in-depth morphological analysis for German and Amharic.
F2TEval: Human-Aligned Multi-Dimensional Evaluation for Figure-to-Text Task
Tan Yue (Peking University), Dongyan Zhao (Beijing University of Posts and Telecommunications)
CodeGenerationComputational EfficiencyTransformerMixture of ExpertsVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed F2TEval, a no-reference chart-text generation quality assessment method based on five-dimensional human evaluation criteria, and constructed the F2TBenchmark dataset.
FacLens: Transferable Probe for Foreseeing Non-Factuality in Fact-Seeking Question Answering of Large Language Models
Yanling Wang (Zhongguancun Laboratory), Ke Xu (Zhongguancun Laboratory)
CodeClassificationDomain AdaptationLarge Language ModelText
🎯 What it does: Studied how to predict the generation of non-factual answers before large language models (LLMs) respond to factual questions, and proposed a lightweight and transferable model called FacLens.
Fair or Framed? Political Bias in News Articles Generated by LLMs
Junho Yoo (Incheon National University), Youhyun Shin (Incheon National University)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper reveals that models still tend to be left-leaning and may invert the stance of cited content when generating text, by evaluating the political bias of seven LLMs in news generation tasks.
🎯 What it does: Propose a reasoning-time mechanism called FairGen, which is based on adaptive latent guidance and memory modules, enabling text-to-image diffusion models to generate images according to specified sensitive attribute distributions without compromising image quality; simultaneously create Holistic Bias Evaluation Benchmark (HBE), which covers more domains and complex prompts; and evaluate its performance on multiple diffusion models.
Fann or Flop: A Multigenre, Multiera Benchmark for Arabic Poetry Understanding in LLMs
Wafa Al Ghallabi (Mohamed bin Zayed University of AI), Rao Muhammad Anwer (Mohamed bin Zayed University of AI)
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed Fann or Flop—the first benchmark targeting 12 historical eras, 14 poetry genres, encompassing classical and modern free verse Arabic poetry understanding;
FaST: Feature-aware Sampling and Tuning for Personalized Preference Alignment with Limited Data
Thibaut Thonet (NAVER Labs Europe), Marc Dymetman (Independent Researcher)
CodeRecommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a PPALLI framework for personalizing large language models under scenarios where users provide only a small number of preference annotations (less than 100), and release two new datasets, DnD and ELIP, within this framework.
Faster In-Context Learning for LLMs via N-Gram Trie Speculative Decoding
Jinglin Chen (Wuhan University), Ping Wang (Wuhan University)
CodeComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a speculative decoding method based on n-gram Trie, utilizing the overlap between context and model output to accelerate inference in large language models
FedMABench: Benchmarking Mobile GUI Agents on Decentralized Heterogeneous User Data
WenHao Wang (Zhejiang University), Yanfeng Wang (Shanghai AI Laboratory)
CodeFederated LearningBenchmark
🎯 What it does: Developed and released FedMABench, the first unified benchmark for distributed mobile GUI agents, to evaluate the effectiveness of federated learning training.
🎯 What it does: The study utilizes a small amount of parallel corpus and monolingual corpus in low-resource languages to achieve zero/few-shot machine translation through training high-quality word vectors.
Finding your MUSE: Mining Unexpected Solutions Engine
Nir Sweed (Hebrew University of Jerusalem), Dafna Shahaf (Hebrew University of Jerusalem)
CodeGenerationTransformerLarge Language ModelTextGraphRetrieval-Augmented Generation
🎯 What it does: Constructed a large-scale Functional Concept Graph (FCG) and proposed the MUSE algorithm to generate creative inspirations on this graph, helping users break free from design rigidity.
Yixuan Tang (Hong Kong University of Science and Technology), Yi Yang (Hong Kong University of Science and Technology)
CodeClassificationData SynthesisRetrievalTransformerLarge Language ModelContrastive LearningTextBenchmarkFinance Related
🎯 What it does: Proposed a novel text embedding benchmark for the financial domain, FinMTEB, and developed FinE5, a financial-adapted embedding model fine-tuned using persona-based synthetic data.
FinRAGBench-V: A Benchmark for Multimodal RAG with Visual Citation in the Financial Domain
Suifeng Zhao (Peking University), Jun Gao (Peking University)
CodeRetrievalTransformerVision Language ModelTextMultimodalityBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Constructed FinRAGBench-V, which includes a multimodal retrieval corpus and a QA dataset, and implemented the RGenCite baseline to support visual citations.
FinTrust: A Comprehensive Benchmark of Trustworthiness Evaluation in Finance Domain
Tiansheng Hu (NYU Shanghai), Chen Zhao (NYU Shanghai)
CodeAdversarial AttackTransformerLarge Language ModelTextMultimodalityTabularTime SeriesBenchmarkFinance Related
🎯 What it does: Proposed the FINTRUST benchmark to systematically evaluate the trustworthiness and alignment of large language models in the financial domain.
FlightGPT: Towards Generalizable and Interpretable UAV Vision-and-Language Navigation with Vision-Language Models
Hengxing Cai (Sun Yat-Sen University), Renxin Zhong (Sun Yat-Sen University)
CodeAutonomous DrivingExplainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Proposed the FlightGPT framework, which utilizes vision-language models to achieve UAV visual and language navigation.
FLUID QA: A Multilingual Benchmark for Figurative Language Usage in Dialogue across English, Chinese, and Korean
Seoyoon Park (Yonsei University), Hansaem Kim (Yonsei University)
CodeLarge Language ModelTextBenchmark
🎯 What it does: Researchers proposed the FLUID QA multilingual dialogic metaphor evaluation benchmark, based on FLUTE data with culturally adapted translations, assessing LLMs' ability to use metaphors in multi-turn dialogues.
Follow the Flow: Fine-grained Flowchart Attribution with Neurosymbolic Agents
Manan Suri (University of Maryland), Dinesh Manocha
CodeExplainability and InterpretabilityAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelAgentic AIVision Language ModelImageTextBenchmark
🎯 What it does: Studied the fine-grained flowchart attribution task, proposing FlowPathAgent which performs post-hoc attribution through graph reasoning after flowchart parsing, and constructed the FlowExplainBench benchmark dataset.
Zixuan Weng (University Of Notre Dame), Xiangyu Zhang (Pennsylvania State University)
CodeSafty and PrivacyAdversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a multi-round jailbreak method called FITD based on the psychological 'foot-in-the-door' principle, which induces LLMs to generate inappropriate outputs through progressively escalating prompts.
CodeLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigate the overfitting phenomenon of large language models on public benchmarks, and propose a meta-evaluation framework called C-BOD based on parameterized paraphrase transformation to detect whether models excessively rely on surface-level prompts;
From A and B to A+B: Can Large Language Models Solve Compositional Math Problems?
Xisheng Xiao (South China Agricultural University), Hanlin Zhao (South China Agricultural University)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Studying how to combine two existing mathematical problems logically into a new 'composite question,' and using this method to evaluate the generalization ability of large language models (LLMs) in compositional reasoning.
From Automation to Autonomy: A Survey on Large Language Models in Scientific Discovery
Tianshi Zheng (HKUST), Yangqiu Song (HKUST)
CodeTransformerLarge Language ModelAgentic AITextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper systematically reviews the application of large language models (LLMs) in scientific discovery, tracing their evolution by mapping three levels of autonomy—tool, analyst, and scientist—to the six stages of the scientific method, and identifying future development challenges;
CodeExplainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: This paper systematically evaluates the geographic-economic bias of six vision-language models when generating chart summaries and attempts prompt-based debiasing methods.
From General Reward to Targeted Reward: Improving Open-ended Long-context Generation Models
Zhihan Guo (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
CodeGenerationData SynthesisLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposed a ProxyReward framework based on reinforcement learning to enhance open-ended long text generation (Open-LTG), utilizing automatically constructed meta-questions and Boolean proxy QA pairs to provide goal-oriented reward signals, eliminating reliance on human annotations and fixed answers.
From Input Perception to Predictive Insight: Modeling Model Blind Spots Before They Become Errors
Maggie Mi (University of Sheffield), Nafise Sadat Moosavi (University of Sheffield)
CodeExplainability and InterpretabilityTransformerText
🎯 What it does: By analyzing the token-level probability distribution of input sequences in language models, an error prediction method based solely on input-side likelihood is proposed.
From Problem-Solving to Teaching Problem-Solving: Aligning LLMs with Pedagogy using Reinforcement Learning
David Dinucu-Jianu (ETH Zurich), Mrinmaya Sachan (ETH Zurich)
CodeReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose an online reinforcement learning framework that simulates student-teacher dialogues, enabling large language models to provide adaptive, step-by-step tutoring in educational scenarios, while balancing teaching quality and problem-solving accuracy through dialogue-level rewards.
🎯 What it does: Propose a zero-shot, schema-only dialogue state tracking (DST) framework that generates diverse synthetic dialogues using dynamic complexity prompting and transfers large language models' reasoning capabilities to small models through a two-stage chain-of-thought (CoT) distillation process.
From Surveys to Narratives: Rethinking Cultural Value Adaptation in LLMs
M. Farid Adilazuarda (Mohamed bin Zayed University of Artificial Intelligence), Alham Fikri Aji (Technical University of Darmstadt)
CodeDomain AdaptationRepresentation LearningData-Centric LearningLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper systematically evaluates methods for adapting large language models (LLMs) to cultural values using the World Values Survey (WVS), supplemented by encyclopedia (Wikipedia) and contextualized norms (NormAd) data, and explores the impact of different data sources on model cultural diversity, task performance, and factual knowledge retention.
🎯 What it does: Train TTS models using limited real transcribed speech, generate hundreds of thousands of hours of synthetic speech via text-to-speech inverse translation (Speech Back-Translation) to expand ASR training data for low-resource languages.
From Unaligned to Aligned: Scaling Multilingual LLMs with Multi-Way Parallel Corpora
Yingli Shen (Tsinghua University), Maosong Sun (Tsinghua University)
CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the TED2025 multilingual parallel corpus (113 languages, up to 50 languages in parallel) and used it for continued pre-training and instruction tuning of large multilingual LLMs, systematically evaluating the impact of factors such as parallelism, language combinations, and instruction targets on model performance.
From Word to World: Evaluate and Mitigate Culture Bias in LLMs via Word Association Test
Xunlian Dai, Haizhou Li (Shenzhen Research Institute Of Big Data)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose a cross-cultural evaluation framework based on the Word Association Test (WAT), design an LLM-adaptive free association task, and further introduce the CultureSteer model to explicitly guide culturally specific associations within the internal semantic space, thereby enhancing the cross-cultural cognitive capabilities of LLMs.
Good Intentions Beyond ACL: Who Does NLP for Social Good, and Where?
Grace LeFevre (Northwestern University), Rob Voigt (Northwestern University)
CodeTransformerLarge Language ModelText
🎯 What it does: This paper conducts a large-scale scientometric analysis of authors and publication contexts across 309,208 NLP papers, systematically mapping the global research ecosystem of NLP-for-Social-Good (NLP4SG);
Chunyang Jiang (Hong Kong University of Science and Technology), Yike Guo
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed an elegant forgetting framework for generative language models (Learning With Forgetting, LWF), which mines pre-trained knowledge through self-generated text, calculates forgetting confidence, and performs periodic gradient ascent-based forgetting during fine-tuning to enhance learning plasticity for downstream tasks.