EMNLP 2024 Papers — Page 3
Conference on Empirical Methods in Natural Language Processing · 1268 papers
CHIQ: Contextual History Enhancement for Improving Query Rewriting in Conversational Search
Fengran Mo (Université de Montréal), Jian-Yun Nie (Université de Montréal)
RetrievalTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the CHIQ method, which first uses an open-source LLM to perform semantic enhancement of the conversation history (disambiguation, answer expansion, pseudo response prediction, topic shift detection, and summary generation), followed by generating queries using the enhanced history;
CItruS: Chunked Instruction-aware State Eviction for Long Sequence Modeling
Yu Bai (Beijing Institute of Technology), Jackie CK Cheung
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the CItruS method, combining block processing with instruction-aware state eviction, to efficiently handle long sequences during inference while retaining critical task information.
CleanGen: Mitigating Backdoor Attacks for Generation Tasks in Large Language Models
Yuetai Li (University of Washington), Radha Poovendran (University of Washington)
GenerationSafty and PrivacyLarge Language ModelText
🎯 What it does: This paper develops a novel inference-time defense strategy called CLEANGEN to mitigate backdoor attacks in large language models (LLMs) during generation tasks.
CliMedBench: A Large-Scale Chinese Benchmark for Evaluating Medical Large Language Models in Clinical Scenarios
Zetian Ouyang (East China Normal University), Liang He (Shanghai Jiao Tong University)
TransformerLarge Language ModelTextBiomedical DataElectronic Health RecordsBenchmarkChain-of-Thought
🎯 What it does: Proposed a Chinese clinical medicine large language model evaluation benchmark called CliMedBench based on real medical cases.
ClimRetrieve: A Benchmarking Dataset for Information Retrieval from Corporate Climate Disclosures
Tobias Schimanski (University of Zurich), Markus Leippold (University of Zurich)
RetrievalTransformerTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed an expert-annotated ClimRetrieve dataset and evaluated the performance of different embedding retrieval strategies in corporate climate disclosure question-answering tasks.
Closing the Loop: Learning to Generate Writing Feedback via Language Model Simulated Student Revisions
Inderjeet Jayakumar Nair (University of Michigan), Lu Wang (University of Michigan)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Design and implement an iterative optimization framework named PROF, which directly enhances the effectiveness of generated writing feedback by simulating the student revision process using language models.
Cluster-Norm for Unsupervised Probing of Knowledge
Walter Laurito (FZI), Kaarel Hänni (Caltech)
Representation LearningText
🎯 What it does: This paper proposes the cluster-normalization method, which clusters activation maps of contrastive pairs and normalizes them separately to reduce interfering features, thereby enhancing unsupervised knowledge probing performance.
Clustering and Ranking: Diversity-preserved Instruction Selection through Expert-aligned Quality Estimation
Yuan Ge (Northeastern University), JingBo Zhu
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: During the supervised fine-tuning phase, this paper addresses the issues of instruction data quality and diversity by proposing a method for selecting instructions based on quality assessment and clustering. It first evaluates the quality of instruction pairs using an expert-aligned scoring model, then retains diversity through clustering, filtering a high-quality subset from a large-scale instruction set for fine-tuning.
CMD: a framework for Context-aware Model self-Detoxification
Zecheng Tang (Soochow University), Min Zhang (Soochow University)
GenerationSafty and PrivacyData-Centric LearningConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextChain-of-Thought
🎯 What it does: Developed a Context-aware Model self-Detoxification (CMD) framework, which first conducts toxicity detection and detoxification on the input context, then generates text under a safe context using a language model, and performs contrastive learning through self-generated synthetic data;
CmdCaliper: A Semantic-Aware Command-Line Embedding Model and Dataset for Security Research
Sian-Yao Huang (CyCraft AI Lab), Chun-Ying Huang (CyCraft AI Lab)
ClassificationData SynthesisRetrievalAnomaly DetectionRepresentation LearningLarge Language ModelContrastive LearningText
🎯 What it does: Designed and released the first command-line similar pair dataset CyPHER and a specialized command-line embedding model CmdCaliper.
CMR Scaling Law: Predicting Critical Mixture Ratios for Continual Pre-training of Language Models
Jiawei Gu (Sun Yat-sen University Independent Researcher), Fei Tan (Nanjing University)
Data-Centric LearningTransformerLarge Language ModelTextFinance Related
🎯 What it does: Studied the impact of the mixture ratio of combining general and domain-specific corpora during Continual Pre-Training (CPT) on the performance of large language models (LLMs), and proposed the concept of 'Critical Mixture Ratio (CMR)' along with its predictive law.
CoBa: Convergence Balancer for Multitask Finetuning of Large Language Models
Zi Gong (Ant Group), Jianguo Li (Ant Group)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the CoBa method to dynamically balance task convergence rates in multi-task fine-tuning of large language models (LLMs).
CoCoLoFa: A Dataset of News Comments with Common Logical Fallacies Written by LLM-Assisted Crowds
Min-Hsuan Yeh (University of Wisconsin-Madison), Ting-Hao Kenneth Huang (Pennsylvania State University)
ClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the COCOLOFA dataset by collecting news comments containing eight common logical fallacies using an LLM-assisted crowdsourcing approach, and trained models to detect and classify fallacies using this data.
CoCoST: Automatic Complex Code Generation with Online Searching and Correctness Testing
Xinyi He (Xi'an Jiaotong University), Dongmei Zhang (Microsoft Research)
AI Code AssistantTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the CoCoST framework, integrating online search with query planning, automatically generated test cases, and input/output serialization to enhance the quality of large language models (LLMs) in complex code generation.
Code Prompting Elicits Conditional Reasoning Abilities in Text+Code LLMs
Haritz Puerto, Iryna Gurevych (TU Darmstadt)
AI Code AssistantPrompt EngineeringTextChain-of-Thought
🎯 What it does: Improve the performance of text+code large language models (LLMs) on conditional reasoning tasks by converting natural language questions into code prompts (code prompting).
CodeAgent: Autonomous Communicative Agents for Code Review
Xunzhu Tang (University of Luxembourg), Tegawendé F. Bissyandé (University of Luxembourg)
AI Code AssistantTransformerLarge Language ModelAgentic AITextChain-of-Thought
🎯 What it does: Proposed a CodeAgent framework for automated code review based on multi-agent large language models, simulating the collaborative process of real code review teams;
CodeJudge: Evaluating Code Generation with Large Language Models
Weixi Tong (Huazhong University of Science and Technology), Tianyi Zhang (Purdue University)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed and implemented a framework called CODEJUDGE that utilizes large language models (LLMs) to evaluate the semantic correctness of code generation results without test cases;
CoEvol: Constructing Better Responses for Instruction Finetuning through Multi-Agent Cooperation
Renhao Li (University of Macau), Min Yang (Shenzhen Institute of Advanced Technology)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the CoEvol framework, which employs LLM-based multi-agent systems (debaters, advisors, editors, judges) through a debate-suggestion-edit-judgment cycle to iteratively improve responses in instruction fine-tuning data, thereby generating higher-quality training samples.
Coffee-Gym: An Environment for Evaluating and Improving Natural Language Feedback on Erroneous Code
Hyungjoo Chae (Yonsei University), Jinyoung Yeo (Yonsei University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the COFFEE-GYM environment, which includes the COFFEE dataset and the COFFEEEVAL reward function, for training and evaluating natural language feedback models to improve code editing;
CoGen: Learning from Feedback with Coupled Comprehension and Generation
Mustafa Omer Gul (Cornell University), Yoav Artzi (Cornell University)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Designed and implemented a system (COGEN) capable of simultaneously performing language understanding and generation, and continuously learning through human interaction, deployed in a two-player dialogue game.
Collaborative Performance Prediction for Large Language Models
Qiyuan Zhang (City University of Hong Kong), Chen Ma (City University of Hong Kong)
OptimizationExplainability and InterpretabilityData-Centric LearningTabularBenchmark
🎯 What it does: Propose a collaborative performance prediction framework (CPP) that predicts performance on various downstream tasks by leveraging historical LLM performance and design factors.
Collective Critics for Creative Story Generation
Minwook Bae (Ulsan National Institute of Science and Technology), Hyounghun Kim (Ulsan National Institute of Science and Technology)
GenerationTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Proposes the CRITICS framework, which leverages multi-round LLM collective criticism (including the planning phase CRPLAN and the text phase CRTEXT) to enhance the creativity and expressiveness of long-form stories while maintaining narrative coherence.
CommonIT: Commonality-Aware Instruction Tuning for Large Language Models via Data Partitions
Jun Rao (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes CommonIT, an instruction tuning strategy that partitions data based on commonalities (task, embedding, length) and samples within a single group, enhancing LLM's instruction following ability.
Commonsense Knowledge Editing Based on Free-Text in LLMs
Xiusheng Huang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences), Kang Liu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes an editing method for free-text form common sense knowledge in large language models and constructs the corresponding benchmark dataset CKEBench. Through the knowledge localization experiment (KLFT), it reveals that common sense knowledge is distributed in MLP and attention layers and is dispersed. Subsequently, a Dynamics-aware Editing Method (DEM) is designed, which includes a dynamic perception module and a knowledge editing module, achieving precise editing of common sense knowledge. Finally, experiments are conducted on GPT-J(6B) and LLaMA-2(7B), demonstrating that DEM significantly outperforms existing editing methods in metrics such as Score and Commonsense.
Communicating with Speakers and Listeners of Different Pragmatic Levels
Kata Naszadi, Christof Monz (University of Amsterdam)
Vision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Simulate language learning and interaction, investigating how varying reasoning depths between speaker and listener impact communication success
Community-Cross-Instruct: Unsupervised Instruction Generation for Aligning Large Language Models to Online Communities
Zihao He (University of Southern California), Kristina Lerman (University of Southern California)
Recommendation SystemData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes COMMUNITY-CROSS-INSTRUCT, a fully unsupervised framework for instruction fine-tuning of online communities, generating LLM digital twins aligned with community perspectives and automatically generating questionnaires to assess alignment.
CommVQA: Situating Visual Question Answering in Communicative Contexts
Nandita Shankar Naik (Stanford University), Elisa Kreiss (University of California, Los Angeles)
TransformerLarge Language ModelVision Language ModelImageTextBenchmark
🎯 What it does: Explored the visual question answering task in a communication context, constructed the CommVQA dataset, and conducted benchmarking.
CompAct: Compressing Retrieved Documents Actively for Question Answering
Chanwoong Yoon (Korea University), Jaewoo Kang (Korea University)
RetrievalCompressionTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Proposed the COMPACT framework, which actively compresses retrieved multiple documents and iteratively generates a concise answer context by integrating the question context during compression until the judgment of 'sufficient to answer the question' is met;
Comparing a BERT Classifier and a GPT classifier for Detecting Connective Language Across Multiple Social Media
Josephine Lukito (University of Texas at Austin), Natalie Jomini Stroud (University of Texas at Austin)
ClassificationTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper develops and evaluates two classifiers, BERT and GPT-3.5 Turbo, for detecting 'linking language'—expressions that promote cross-political stance communication and understanding—in social media posts.
Comparing Neighbors Together Makes it Easy: Jointly Comparing Multiple Candidates for Efficient and Effective Retrieval
Jonghyun Song (Seoul National University), Jay-Yoon Lee (Seoul National University)
RetrievalComputational EfficiencyTransformerText
🎯 What it does: Propose a retrieval and re-ranking framework named Comparing Multiple Candidates (CMC), which improves candidate context representations by parallelly comparing multiple candidate vectors within the self-attention layer;
Computational Meme Understanding: A Survey
Khoi P. N. Nguyen (University of Texas at Dallas), Vincent Ng (University of Texas at Dallas)
ClassificationConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Review the current research on automatic understanding of memes (CMU) using computer vision and language models. Construct a three-dimensional classification framework based on form, function, and theme. Systematically organize main tasks (classification, explanation, interpretation) with corresponding datasets, models, and evaluation metrics. Analyze technical bottlenecks and ethical risks, and provide future research directions.
Concept Space Alignment in Multilingual LLMs
Qiwei Peng (University of Copenhagen), Anders Søgaard (University of Copenhagen)
RetrievalRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: By treating parallel concepts in multilingual WordNet as dictionaries, conducting linear alignment experiments on multilingual large language models to evaluate the quality of their concept space alignment.
Concept-skill Transferability-based Data Selection for Large Vision-Language Models
Jaewoo Lee (KAIST), Sung Ju Hwang (KAIST)
Data-Centric LearningVision Language ModelMultimodality
🎯 What it does: Propose the COINCIDE method, which clusters visual instruction tuning data using internal activations of a small vision-language model and selects training samples based on the transitivity and density of the clusters;
Conditional and Modal Reasoning in Large Language Models
Wesley H. Holliday (University of California, Berkeley), Cedegao E. Zhang (Massachusetts Institute of Technology)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Evaluate the performance of 29 large language models on logical reasoning involving conditional and modal sentences, and construct a benchmark test covering various fundamental and controversial reasoning patterns;
Connecting the Dots: Evaluating Abstract Reasoning Capabilities of LLMs Using the New York Times Connections Word Game
Prisha Samdarshi, Smaranda Muresan (Columbia University)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Evaluate the abstract reasoning abilities of large language models (LLMs) in the New York Times (NYT) Connections puzzle game, and compare them with novice and expert human players.
Consecutive Batch Model Editing with HooK Layers
Shuaiyi Li (Chinese University of Hong Kong), Wai Lam (Tencent)
TransformerLarge Language ModelText
🎯 What it does: Propose the CoachHook method, utilizing a hook layer and a continuous update mechanism to achieve continuous batch model editing.
Consistent Autoformalization for Constructing Mathematical Libraries
Lan Zhang (University of Manchester), Andre Freitas (University of Manchester)
AI Code AssistantTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed an automated formalization framework based on large language models, integrating three mechanisms: retrieval-augmented generation, denoising, and syntax error feedback to build consistent and scalable mathematical libraries.
Consistent Bidirectional Language Modelling: Expressive Power and Representational Conciseness
Georgi Shopov (Bulgarian Academy of Sciences), Stefan Gerdjikov (Bulgarian Academy of Sciences)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed a class of bidirectional language models named latent language models, proving their consistency for generation and scoring; simultaneously providing systematic characterization of its subclass, rational language models.
Consolidating Ranking and Relevance Predictions of Large Language Models through Post-Processing
Le Yan (Google Research), Harrie Oosterhuis (Google Research)
RetrievalTransformerLarge Language ModelText
🎯 What it does: This paper proposes an unsupervised post-processing method that aligns relevance labels generated by large language models (LLMs) in pseudo-evaluation mode with preference orders generated in contrastive ranking mode through constrained regression, thereby simultaneously improving ranking quality and the accuracy of relevance prediction.
CONTESTS: a Framework for Consistency Testing of Span Probabilities in Language Models
Eitan Wagner (Hebrew University of Jerusalem), Omri Abend (Hebrew University of Jerusalem)
Explainability and InterpretabilityTransformerText
🎯 What it does: This paper investigates the probability consistency of language models when computing the joint probability over word spans, proposing the CONTESTS framework that evaluates probability consistency under different inference orders through statistical testing.
Context-Aware Adapter Tuning for Few-Shot Relation Learning in Knowledge Graphs
Liu Ran (Singapore Management University), Yuan Fang (Singapore Management University)
Representation LearningMeta LearningGraph
🎯 What it does: Designed and implemented RelAdapter, a context-aware adapter for few-shot relation learning in knowledge graphs.
Context-Aware Assistant Selection for Improved Inference Acceleration with Large Language Models
Jerry Huang (Mila Quebec AI Institute), Sarath Chandar (Mila Quebec AI Institute)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Studies how to utilize offline reinforcement learning trained contextual multi-armed bandit strategies to dynamically select the most suitable draft models among multiple black-box draft models, thereby improving the reasoning speed of large language models.
Context-aware Watermark with Semantic Balanced Green-red Lists for Large Language Models
Yuxuan Guo (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Researched and implemented a semantic-aware text watermarking method for large language models, aiming to enhance robustness against rewrite attacks while maintaining text quality.
Contextualized Sequence Likelihood: Enhanced Confidence Scores for Natural Language Generation
Zhen Lin (University of Illinois at Urbana-Champaign), Jimeng Sun (University of Illinois at Urbana-Champaign)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a sequence likelihood confidence estimation method based on reweighted self-attention weights—Contextualized Sequence Likelihood (CSL)—for confidence assessment in natural language generation.
Continual Test-time Adaptation for End-to-end Speech Recognition on Noisy Speech
Guan-Ting Lin (National Taiwan University), Hung-yi Lee (National Taiwan University)
RecognitionDomain AdaptationTransformerAudio
🎯 What it does: Propose the Fast-slow TTA framework and the DSUTA method to address performance degradation of Continual Test-time Adaptation (CTTA) on multi-domain noisy speech;
Contrastive Entity Coreference and Disambiguation for Historical Texts
Abhishek Arora (Harvard University), Leander Heldring (National Bureau of Economic Research)
RetrievalTransformerContrastive LearningTextBenchmark
🎯 What it does: The study addresses the challenges of entity coreference and disambiguation in historical texts, proposing a bi-encoder model based on contrastive learning, along with providing large-scale training data and historical benchmark evaluations.
Contrastive Policy Gradient: Aligning LLMs on sequence-level scores in a supervised-friendly fashion
Yannis Flet-Berliac (Cohere), Matthieu Geist (Cohere)
Reinforcement LearningContrastive LearningText
🎯 What it does: Proposed a new reinforcement learning algorithm called Contrastive Policy Gradient (CoPG), which optimizes any reward function directly in offline, non-importance-sampling settings by contrasting rewards and using specific baselines, thereby achieving alignment of large language models (LLMs);
Contribution of Linguistic Typology to Universal Dependency Parsing: An Empirical Investigation
Ali Basirat (University of Copenhagen), Navid Baradaran Hemmati (Certified Translation Agency No. 1141)
Data-Centric LearningText
🎯 What it does: Applied Croft's semantic transformation rules to the Universal Dependencies treebank to construct Typologically-informed Universal Dependencies (TUD), and evaluated its impact on dependency parsing accuracy.
Control Large Language Models via Divide and Conquer
Bingxuan Li (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
GenerationTransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: This paper investigates the performance of large language models (LLMs) in lexically constrained generation under prompt-based control, and proposes a Divide and Conquer (DnC) generation strategy to significantly improve the satisfaction rate.
Controllable Preference Optimization: Toward Controllable Multi-Objective Alignment
Yiju Guo (Renmin University of China), Maosong Sun (Tsinghua University)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a controllable preference optimization (CPO) method, using preference tokens to control the output of large models in multiple objectives such as helpfulness, honesty, and harmlessness, achieving controllable multi-objective alignment.
ControlMath: Controllable Data Generation Promotes Math Generalist Models
Nuo Chen (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringText
🎯 What it does: Propose a controllable mathematical data augmentation method called ControlMath, which constructs a high-quality mathematical general-purpose model training set named ControlMathQA through three steps: controllable equation generation, problem generation, and efficient filtering.
CopyBench: Measuring Literal and Non-Literal Reproduction of Copyright-Protected Text in Language Model Generation
Tong Chen (University of Washington), Pang Wei Koh (University of Washington)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: This study proposes the COPYBENCH benchmark for automatically evaluating literal and non-literal copying of language models on copyrighted texts, while simultaneously measuring practical metrics such as factual recall and fluency.
CorrSynth - A Correlated Sampling Method for Diverse Dataset Generation from LLMs
Suhas S Kowshik (Amazon), Vijit Malik (Amazon)
Data SynthesisTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a correlated sampling method called CORRSYNTH, which generates diverse and label-loyal synthetic datasets using LLMs, followed by training student models.
CoSafe: Evaluating Large Language Model Safety in Multi-Turn Dialogue Coreference
Erxin Yu (Hong Kong Polytechnic University), Lanqing Hong (Huawei)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the CoSafe dataset to evaluate the safety of large language models in multi-turn dialogue core citation scenarios
CoTKR: Chain-of-Thought Enhanced Knowledge Rewriting for Complex Knowledge Graph Question Answering
Yike Wu (Southeast University), Jeff Z. Pan (University of Edinburgh)
TransformerLarge Language ModelGraphBenchmarkChain-of-Thought
🎯 What it does: Proposed the CoTKR method, which enhances knowledge rewriting with chain-of-thought (CoT) reasoning to generate more useful natural language knowledge representations, thereby improving the performance of knowledge graph question answering (KGQA).
CoverICL: Selective Annotation for In-Context Learning via Active Graph Coverage
Costas Mavromatis (Amazon Web Services), George Karypis (Amazon Web Services)
Computational EfficiencyData-Centric LearningLarge Language ModelText
🎯 What it does: Select the most informative examples for context learning (ICL) in large language models under a low budget through active graph coverage techniques, thereby reducing annotation and inference costs.
Crafting Personalized Agents through Retrieval-Augmented Generation on Editable Memory Graphs
Zheng Wang (Huawei Technologies Co Ltd), Wei Shi (Huawei Technologies Co Ltd)
AI Code AssistantReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningAgentic AIImageTextRetrieval-Augmented Generation
🎯 What it does: Proposed a personalized intelligent assistant framework EMG-RAG based on large language models (LLMs), which extracts 'memories' from smartphone daily conversations and screenshots to construct an editable memory graph (EMG), and employs reinforcement learning to adaptively select memories on this graph to support downstream tasks such as question-answering, form auto-filling, and user service.
Cross-Domain Audio Deepfake Detection: Dataset and Analysis
Yuang Li (Huawei Translation Services Center), Hao Yang (Huawei Translation Services Center)
Domain AdaptationAnomaly DetectionTransformerSupervised Fine-TuningAudio
🎯 What it does: This study constructs a dataset (CD-ADD) containing over 300 hours of cross-domain zero-shot TTS synthesized speech, and evaluates and enhances the deepfake audio detection performance of Wav2Vec2 and Whisper under various attacks (noise, echo, compression, etc.) through attack augmentation training and few-shot fine-tuning methods.
Cross-domain NER with Generated Task-Oriented Knowledge: An Empirical Study from Information Density Perspective
Zhihao Zhang (Soochow University), Guodong Zhou (Soochow University)
RecognitionDomain AdaptationTransformerLarge Language ModelText
🎯 What it does: Generate task-oriented knowledge (GTOK) automatically using large language models (LLMs) and employ it for masked span language modeling pre-training to construct a cross-domain named entity recognition (CDNER) framework called TOPT.
Cross-lingual Back-Parsing: Utterance Synthesis from Meaning Representation for Zero-Resource Semantic Parsing
Deokhyung Kang (POSTECH), Gary Lee
GenerationData SynthesisTransformerLarge Language ModelText
🎯 What it does: In the zero-resource cross-lingual semantic parsing scenario, a data augmentation method called Cross-lingual Back-Parsing (CBP) based on a multilingual pre-trained model is proposed, which can synthesize target language dialog sentences from the semantic representations of the source language.
Cross-lingual Transfer for Automatic Question Generation by Learning Interrogative Structures in Target Languages
Seonjeong Hwang (POSTECH), Gary Lee (POSTECH)
GenerationDomain AdaptationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the QuIST method, which learns question structure during the reasoning phase using a few example questions in the target language, achieving cross-lingual automatic question generation with only English data training.
CryptoTrade: A Reflective LLM-based Agent to Guide Zero-shot Cryptocurrency Trading
Yuan Li (National University of Singapore), Bingsheng He (National University of Singapore)
TransformerLarge Language ModelAgentic AITextTabularTime SeriesBenchmarkFinance Related
🎯 What it does: Developed CryptoTrade, a cryptocurrency trading agent based on large language models, integrating on-chain transaction data with off-chain news information, and achieving zero-shot daily trading decisions through a self-reflective mechanism.
CSSL: Contrastive Self-Supervised Learning for Dependency Parsing on Relatively Free Word Ordered and Morphologically Rich Low Resource Languages
Pretam Ray (IIT Kharagpur), Pawan Goyal (IIT Kharagpur)
Representation LearningData-Centric LearningTransformerContrastive LearningText
🎯 What it does: This paper proposes a self-supervised contrastive learning module (CSSL) for dependency parsing of low-resource, morphologically rich languages with relatively flexible word order. The module enhances the model's robustness to word order variations by encouraging sentences to be close to their word-order-permuted positive samples and far from negative samples.
Cultural Conditioning or Placebo? On the Effectiveness of Socio-Demographic Prompting
Sagnik Mukherjee (Mohamed bin Zayed University of Artificial Intelligence), Monojit Choudhury (Mohamed bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study explores the effectiveness of socio-demographic prompts in detecting model cultural bias by comparing the response differences between culturally sensitive and non-cultural prompts across four large language models.
CURE: Context- and Uncertainty-Aware Mental Disorder Detection
Migyeong Kang (Sungkyunkwan University), Jinyoung Han (Sungkyunkwan University)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the CURE framework, which detects mental disorders by leveraging symptoms, contextual information, and uncertainty fusion, and constructed the KoMOS Korean mental health dataset.
Curriculum Consistency Learning for Conditional Sentence Generation
Liangxin Liu (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
GenerationImageTextMultimodalityAudio
🎯 What it does: Proposed and implemented the Curriculum Consistency Learning (CCL) framework to dynamically adjust the difficulty and weight of consistency learning during the training of conditional sentence generation models based on the model's current capabilities.
CUTE: Measuring LLMs’ Understanding of Their Tokens
Lukas Edman (LMU Munich), Alexander Fraser (TU Munich)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Designed the CUTE benchmark to evaluate large language models' morphological understanding and character-level manipulation capabilities regarding subword tokenization.
D2R: Dual-Branch Dynamic Routing Network for Multimodal Sentiment Detection
Yifan Chen (South China Normal University), Fenghuan Li (Guangdong University of Technology)
ClassificationTransformerMultimodality
🎯 What it does: Designed the DR2 dynamic routing network for adaptive interaction path selection in multimodal sentiment detection.
D3CODE: Disentangling Disagreements in Data across Cultures on Offensiveness Detection and Evaluation
Aida Davani (Google Research), Vinodkumar Prabhakaran (Google Research)
Data-Centric LearningText
🎯 What it does: Constructed a cross-cultural dataset D3CODE and analyzed differences in judgments of aggressive language based on moral foundations across regions and individuals.
DA-Code: Agent Data Science Code Generation Benchmark for Large Language Models
Yiming Huang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences), Kang Liu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences)
AI Code AssistantTransformerLarge Language ModelAgentic AITextTabularBenchmark
🎯 What it does: Proposed the DA-Code benchmark to evaluate the capabilities of large language models in agent-based data science code generation tasks; simultaneously constructed the DA-Agent framework as an execution and interaction code agent;
DA^3: A Distribution-Aware Adversarial Attack against Language Models
Yibo Wang (University of Illinois Chicago), Philip S. Yu (University of Illinois Chicago)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a distribution-aware adversarial attack method DA³, which utilizes LoRA fine-tuning of models and incorporates distribution alignment loss to generate adversarial texts that are difficult to be detected by OD.
DAMRO: Dive into the Attention Mechanism of LVLM to Reduce Object Hallucination
Xuan Gong (Tongji University), Zhihua Wei (Tongji University)
Object DetectionExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Conduct an in-depth analysis of the attention distribution between the visual encoder and language decoder in large vision-language models (LVLMs), revealing the presence of high-norm outlier tokens in the visual encoder that align strongly with the decoder's attention, leading to object hallucinations in text generation. Based on this, we propose the training-free DAMRO method: first, filter out outlier background tokens in the visual encoder using ViT's [CLS] attention weights, then suppress the influence of these outliers on LLM decoding through contrastive decoding, significantly reducing object hallucinations.
Dancing in Chains: Reconciling Instruction Following and Faithfulness in Language Models
Zhengxuan Wu (Stanford University), Zhiheng Huang (Denser.ai)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study investigates the trade-off between instruction following and maintaining answer consistency with the context (credibility) and proposes a self-instruction fine-tuning method based on rejection sampling (RESET) to balance both aspects.
Data Advisor: Dynamic Data Curation for Safety Alignment of Large Language Models
Fei Wang (University of Southern California), Aram Galstyan (Amazon AGI Foundations)
Data SynthesisSafty and PrivacyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes DATA ADVISOR, a framework that dynamically guides data generation through LLM to enhance the quality and coverage of safety-aligned data.
Data Contamination Can Cross Language Barriers
Feng Yao (University of California, San Diego), Jingbo Shang (University of California, San Diego)
Data-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated the impact of cross-lingual data contamination on large language model (LLM) benchmarks and proposed detection methods.
Data, Data Everywhere: A Guide for Pretraining Dataset Construction
Jupinder Parmar (NVIDIA), Bryan Catanzaro (NVIDIA)
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Systematically study the entire process of constructing pre-training datasets, including data cleaning, deduplication, quality filtering, data selection, sampling weight allocation, and attribute analysis on 90+ Common Crawl snapshots, exploring how attribute information enhances pre-training set quality.
DataNarrative: Automated Data-Driven Storytelling with Visualizations and Texts
Mohammed Saidul Islam (York University), Shafiq Joty (Salesforce AI)
GenerationTransformerLarge Language ModelAgentic AIPrompt EngineeringImageTextMultimodalityTabularBenchmark
🎯 What it does: This study proposes an automated data storytelling framework based on multi-agent LLMs and constructs a multimodal data storytelling benchmark dataset, DATANARRATIVE, containing 1,449 stories;
DataTales: A Benchmark for Real-World Intelligent Data Narration
Yajing Yang (National University of Singapore), Min-Yen Kan (Sea AI Lab)
GenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextTabularBenchmarkFinance Related
🎯 What it does: Created and released the DATATALES benchmark dataset, aiming to evaluate language models' ability to convert complex financial table data into professional narrative text, and assess factual accuracy, insightfulness, and style through automated methods.
DC-Instruct: An Effective Framework for Generative Multi-intent Spoken Language Understanding
Bowen Xing, Ivor Tsang (Agency for Science, Technology and Research)
RecognitionTransformerLarge Language ModelPrompt EngineeringContrastive LearningBenchmarkAudio
🎯 What it does: Propose the DC-Instruct framework, which employs a generative Prompt learning method using Dual-task Inter-dependent Instructions (DII) and Supervised Contrastive Instructions (SCI) to achieve multi-intent speech language understanding.
De-Identification of Sensitive Personal Data in Datasets Derived from IIT-CDIP
Stefan Larson (Vanderbilt University), Kevin Leach (Vanderbilt University)
ClassificationData SynthesisSafty and PrivacyTransformerContrastive LearningImage
🎯 What it does: Manually annotate five public datasets derived from IIT-CDIP (RVL-CDIP, Tobacco3482, Tobacco800, FUNSD, DocVQA), quantify the various sensitive PII (such as SSN, birth date/place, address, marital status, etc.) contained within them, and build a modular de-identification pipeline based on these annotations. The pipeline replaces sensitive information with synthetic and realistic data, followed by the public release of the de-identified version.
Deciphering Cognitive Distortions in Patient-Doctor Mental Health Conversations: A Multimodal LLM-Based Detection and Reasoning Framework
Gopendra Vikram Singh (IIT Patna), Asif Ekbal (IIT Jodhpur)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningVideoTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: This paper proposes a zero-shot cognitive dissonance detection and reasoning framework ZS-CoDR based on a multimodal LLM, which can simultaneously identify cognitive dissonance and generate reasoning explanations in patient-doctor dialogues.
Deciphering Rumors: A Multi-Task Learning Approach with Intent-aware Hierarchical Contrastive Learning
Chang Yang (Tianjin University), Jing Zhang (Tianjin University)
ClassificationRecurrent Neural NetworkTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed IRDNet, which integrates multi-task learning to jointly perform rumor detection and latent intent mining, thereby enhancing detection accuracy.
Deciphering the Interplay of Parametric and Non-parametric Memory in Retrieval-augmented Language Models
Mehrdad Farahani (Chalmers University of Technology), Richard Johansson (Chalmers University of Technology)
Explainability and InterpretabilityTextRetrieval-Augmented Generation
🎯 What it does: Studied the interaction between parameter memory and retrieval memory in the retrieval-augmented generation model ATLAS, using causal mediation analysis and control experiments to quantify the model's behavior during copying (non-parametric) and recall (parametric).
Decoding Matters: Addressing Amplification Bias and Homogeneity Issue in Recommendations for Large Language Models
Keqin Bao (University of Science and Technology of China), Fuli Feng (National University of Singapore)
Recommendation SystemTransformerLarge Language ModelText
🎯 What it does: This paper proposes a new decoding strategy called D³ to address the issues of amplified bias and homogenization in LLM recommendation systems.
Decoding Susceptibility: Modeling Misbelief to Misinformation Through a Computational Approach
Yanchen Liu (Harvard University), Diyi Yang (Stanford University)
ClassificationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a computational model based on social media forwarding behavior to estimate users' potential susceptibility to COVID-19 misinformation, validated and analyzed through large-scale Twitter data.
Decoding the Echoes of Vision from fMRI: Memory Disentangling for Past Semantic Information
Runze Xia (Nanjing University of Aeronautics and Astronautics), Piji Li (Nanjing University of Aeronautics and Astronautics)
Representation LearningVision Language ModelContrastive LearningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Under continuous visual stimuli, this study investigates the retention of past visual information in working memory using fMRI data, and proposes the Memory Disentangling task, aiming to extract and decouple current and past semantic information from brain signals at a single time point.
Decoding with Limited Teacher Supervision Requires Understanding When to Trust the Teacher
Hyunjong Ok (Pohang University of Science and Technology), Jaeho Lee (Pohang University of Science and Technology)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelImageTextBenchmarkAudio
🎯 What it does: Propose that in scenarios with limited teacher supervision, small-scale LLMs can improve generation quality by using adaptive α mixing and entropy threshold strategies with minimal supervision from large models.
Decompose and Compare Consistency: Measuring VLMs’ Answer Reliability via Task-Decomposition Consistency Comparison
Qian Yang (Mila Quebec AI Institute), Aishwarya Agrawal (University of California Santa Barbara)
Large Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposes a visual language model (VLM) answer reliability assessment method called DeCC based on task decomposition and consistency comparison.
DECOR: Improving Coherence in L2 English Writing with a Novel Benchmark for Incoherence Detection, Reasoning, and Rewriting
Xuanming Zhang (Columbia University), Zhou Yu (Columbia University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed the DECOR benchmark, which includes three tasks: incoherence detection, causal reasoning, and sentence rewriting. Generated 1,352 context-sentence pairs from TOEFL-11 articles, and manually annotated 446 incoherent instances and 213 expert rewrites.
DecorateLM: Data Engineering through Corpus Rating, Tagging, and Editing with Language Models
Ranchi Zhao (Modelbest Inc), Maosong Sun (Tsinghua University)
Knowledge DistillationData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed and implemented DecorateLM, a complete workflow for sample-level data engineering on large-scale pre-trained corpora through text scoring, hierarchical labeling, and editing;
Defending Against Social Engineering Attacks in the Age of LLMs
Lin Ai (Columbia University), Julia Hirschberg (Columbia University)
Anomaly DetectionSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper explores the dual role of large language models (LLMs) in chat-based social engineering (CSE)—both generating attack dialogues and detecting/defending against them; constructs the first LLM-generated CSE dialogue dataset, SEConvo, and proposes a modular detection pipeline, ConvoSentinel;
Defending Jailbreak Prompts via In-Context Adversarial Game
Yujun Zhou (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
Adversarial AttackLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose an adversarial game framework (ICAG) that does not require fine-tuning, enhancing LLMs' defense against jailbreak attacks by continuously iterating attack and defense agents in the context.
Defining Knowledge: Bridging Epistemology and Large Language Models
Constanza Fierro (University of Copenhagen), Anders Søgaard (University of Copenhagen)
Explainability and InterpretabilityTransformerLarge Language ModelTextReview/Survey PaperChain-of-Thought
🎯 What it does: This paper reviews five classical definitions of knowledge in philosophy and formalizes them into corresponding evaluation criteria for large language models (LLMs). It then analyzes the alignment between existing NLP research practices in knowledge assessment and these definitions, revealing numerous flaws. Through a questionnaire survey of 105 philosophers and computer scientists, it highlights differences in attitudes between the two groups regarding definitions of knowledge and whether LLMs can 'know.' Finally, it proposes operational assessment protocols for each definition.
DEFT-UCS: Data Efficient Fine-Tuning for Pre-Trained Language Models via Unsupervised Core-Set Selection for Text-Editing
Devleena Das (Georgia Institute of Technology), Vivek Khetan (Accenture Labs)
Representation LearningData-Centric LearningTransformerSupervised Fine-TuningText
🎯 What it does: Proposed the DEFT-UCS framework, which employs unsupervised core-set selection to perform data-efficient fine-tuning of pre-trained language models (PLMs) for text editing tasks.
Delving into Qualitative Implications of Synthetic Data for Hate Speech Detection
Camilla Casula (Fondazione Bruno Kessler), Sara Tonelli (Fondazione Bruno Kessler)
ClassificationData SynthesisTransformerLarge Language ModelText
🎯 What it does: Evaluated and qualitatively analyzed the effectiveness of synthetic data generated by large language models (LLMs) in English hate speech detection, comparing model performance using original data, pure synthetic data, and synthetic-original hybrid training. Manual annotations were conducted on the authenticity of synthetic text, preservation of hate labels, and distribution of identity targets.
DEM: Distribution Edited Model for Training with Mixed Data Distributions
Dhananjay Ram (AGI Foundations, Amazon AWS AI Labs), Sheng Zha (AGI Foundations, Amazon AWS AI Labs)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the Distribution Edited Model (DEM), which first fine-tunes each data source separately and extracts parameter differences between the data source and the base model (distribution vectors), then combines them with weighted sums to obtain a multi-task instruction model.
Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning
Zhaoxuan Tan (University of Notre Dame), Meng Jiang (University of Notre Dame)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Provide each user with a pluggable parameter-efficient fine-tuning (PEFT) module, enabling the embedding of individual behavioral patterns into large language models (LLMs) to achieve model ownership and adaptation to behavioral transfer.
DeMPT: Decoding-enhanced Multi-phase Prompt Tuning for Making LLMs Be Better Context-aware Translators
Xinglin Lyu (Soochow University), Min Zhang (Soochow University)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Designed a Decoding-enhanced Multi-phase Prompt Tuning (DeMPT) method, dividing context-aware NMT into three stages (cross-sentence encoding, intra-sentence encoding, decoding), and introducing heuristic strategies during the decoding phase to enhance LLMs' ability to distinguish and utilize cross-sentence and intra-sentence contexts.
Demystifying Verbatim Memorization in Large Language Models
Jing Huang (Stanford University), Christopher Potts (Stanford University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigated the character-by-character memorization behavior of large language models by injecting sequences into Pythia checkpoints and continuing pre-training, observing memory frequency, quality, and triggering mechanisms.
Dense X Retrieval: What Retrieval Granularity Should We Use?
Tong Chen (University of Washington), Dong Yu (Tencent AI Lab)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Studied the impact of retrieval unit granularity in dense retrieval, proposing the use of propositions as retrieval units and creating FACTOIDWIKI.