EMNLP 2024 Papers — Page 10
Conference on Empirical Methods in Natural Language Processing · 1268 papers
Quantifying the Gaps Between Translation and Native Perception in Training for Multimodal, Multilingual Retrieval
Kyle Buettner (University of Pittsburgh), Adriana Kovashka (University of Pittsburgh)
RetrievalTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Quantified the gap between translation and native perception during training through multimodal multilingual retrieval experiments, and evaluated three strategies for enhancing translation.
Quantum Recurrent Architectures for Text Classification
Wenduan Xu (Quantinuum), Konstantinos Meichanetzidis (Quantinuum)
ClassificationRecurrent Neural NetworkText
🎯 What it does: Designed and trained a quantum recurrent neural network (QRNN) based on parameterized quantum circuits for text classification tasks.
QuBE: Question-based Belief Enhancement for Agentic LLM Reasoning
Minsoo Kim (Seoul National University), Seung-won Hwang (Seoul National University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the QuBE method, which constructs belief states through question-answering and generates belief-based reasoning to address the reasoning drift problem of LLM agents in partially observable environments.
QUDSELECT: Selective Decoding for Questions Under Discussion Parsing
Ashima Suvarna (University of California, Los Angeles), Nanyun Peng (University of California, Los Angeles)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposed a joint training framework called QUDSELECT, which uses instruction-tuned LLMs to simultaneously predict anchor sentences and corresponding questions, and employs selective decoding (sampling + evaluation criteria) during inference to generate question-answer relationships that meet QUD theory standards.
QUIK: Towards End-to-end 4-Bit Inference on Generative Large Language Models
Saleh Ashkboos (ETH Zurich), Dan Alistarh (Institute of Science and Technology Austria)
GenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Joint 4-bit quantization of weights and activations in large generative language models using a mixed-precision scheme, retaining only a small number of outliers as high precision to construct the QUIK format and achieve efficient inference on GPUs.
QUITE: Quantifying Uncertainty in Natural Language Text in Bayesian Reasoning Scenarios
Timo Pierre Schrader (Bosch Center for Artificial Intelligence), Annemarie Friedrich (University of Augsburg)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the QUITE dataset, providing natural language descriptions of multi-domain real-world Bayesian networks, with background premises given in the form of probability values or estimated probability terms, followed by evidence-question pairs requiring the model to output the confidence probability; simultaneously compare the performance of large language model (LLM) prompt-based reasoning with neuro-symbolic ProbLog parser.
RA2FD: Distilling Faithfulness into Efficient Dialogue Systems
Zhiyuan Zhu (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyKnowledge DistillationTransformerContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose a RA2FD method based on a teacher-student framework, which trains retrieval-free dialogue models by using multiple knowledge-injected responses generated by the RAG teacher through sequence-level distillation and contrastive learning;
RAFT: Realistic Attacks to Fool Text Detectors
James Liyuan Wang (Columbia University), Chengzhi Mao (Columbia University)
Adversarial AttackLarge Language ModelText
🎯 What it does: Proposes RAFT, a zero-shot black-box attack framework that misleads text detectors by replacing key words in LLM-generated text with grammatically correct substitutions.
RAG-QA Arena: Evaluating Domain Robustness for Long-form Retrieval Augmented Question Answering
Rujun Han (Google), Vittorio Castelli (AWS AI Labs)
Domain AdaptationTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the LFRQA dataset, which includes multi-domain, long-form, and coherent human-annotated answers for cross-domain robustness evaluation of retrieval-augmented generative QA (RAG-QA), and built the RAG-QA ARENA evaluation framework based on this dataset;
Ranking Manipulation for Conversational Search Engines
Samuel Pfrommer (UC Berkeley), Somayeh Sojoudi (UC Berkeley)
RetrievalAdversarial AttackTransformerPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Studied the natural ranking tendencies of retrieval-augmented generation (RAG) models in conversational search engines, and proposed an attack method to manipulate product rankings by injecting adversarial prompts (prompt injection)
RAR: Retrieval-augmented retrieval for code generation in low resource languages
Avik Dutta (Microsoft), Vu Le (Microsoft)
GenerationAI Code AssistantTransformerSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: For low-resource programming languages, a two-step retrieval (RAR) mechanism is proposed, first retrieving the driving context through examples or documents, and then filtering relevant information from another resource (document or example) via affected retrieval for GPT-4 code generation.
RAt: Injecting Implicit Bias for Text-To-Image Prompt Refinement Models
Ziyi Kou (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
Adversarial AttackTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Proposes RAt, an adversarial attack framework that implicitly injects conceptual bias into text-to-image prompt refinement models.
RaTEScore: A Metric for Radiology Report Generation
Weike Zhao (Shanghai Jiao Tong University), Weidi Xie (Shanghai Jiao Tong University)
RecognitionGenerationTransformerLarge Language ModelTextBiomedical DataBenchmark
🎯 What it does: Proposed a RaTEScore, an entity-level medical text evaluation metric, for assessing the quality of AI-generated radiology reports.
Rationale-Aware Answer Verification by Pairwise Self-Evaluation
Akira Kawabata (Asahi Shimbun Company), Saku Sugawara (National Institute of Informatics)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Studied the impact of reasoning path quality on the credibility of answer verifiers in large language models (LLMs), and proposed the REPS method, which iteratively filters high-quality reasoning paths for training reliable answer verifiers.
Rationalizing Transformer Predictions via End-To-End Differentiable Self-Training
Marc Brinner (Bielefeld University), Sina Zarrieß (Bielefeld University)
ClassificationExplainability and InterpretabilityTransformerAuto EncoderText
🎯 What it does: Proposed an end-to-end differentiable Transformer interpreter that performs classification and provides class importance scores for each word with a single model.
Re-Evaluating Evaluation for Multilingual Summarization
Jessica Zosa Forde (Brown University), Ellie Pavlick (Eleuther Ai)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Built and released a small-scale trilingual (English, Chinese, Indonesian) multilingual summarization evaluation dataset, employing pairwise comparisons and Elo scores for fine-grained ranking; systematically examined the correlation between existing automatic evaluation metrics (ROUGE, BERTScore, GPT-4) and human assessments.
RE-RAG: Improving Open-Domain QA Performance and Interpretability with Relevance Estimator in Retrieval-Augmented Generation
Kiseung Kim (Seoul National University), Jay-Yoon Lee (Seoul National University)
GenerationRetrievalExplainability and InterpretabilityTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the RE-RAG framework, which integrates a relevance estimator into traditional RAG. This allows for re-ranking retrieved contexts and provides the generator with precise relevance weights and confidence scores.
Re-Reading Improves Reasoning in Large Language Models
Xiaohan Xu, Shuai Ma (Beihang University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Designed and verified a simple re-reading prompt (RE2) method to enhance the performance of decoder-only LLMs in reasoning tasks.
Re-ReST: Reflection-Reinforced Self-Training for Language Agents
Zi-Yi Dou (University of California, Los Angeles), Nanyun Peng (University of California, Los Angeles)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextMultimodality
🎯 What it does: Propose a self-training framework based on reflection reinforcement (Re-ReST), which improves the performance of language agents by leveraging environmental feedback through a reflection model to enhance self-generated samples.
Read Anywhere Pointed: Layout-aware GUI Screen Reading with Tree-of-Lens Grounding
Yue Fan (University of California, Santa Cruz), Xin Eric Wang (University of California, Santa Cruz)
Object DetectionGenerationLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Propose a multi-modal large model framework based on the Tree-of-Lens (ToL) to generate natural language descriptions that include both content and layout information after point annotation at any screen position, addressing the ScreenPR task.
ReadMe++: Benchmarking Multilingual Language Models for Multi-Domain Readability Assessment
Tarek Naous (Georgia Institute Of Technology), Wei Xu (Georgia Institute Of Technology)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the README++ multilingual and multidomain sentence readability assessment dataset, annotated with human readability scores based on the CEFR six-level framework; systematically benchmarked the readability evaluation performance of various language models under supervised, unsupervised, and few-shot prompting settings.
RealVul: Can We Detect Vulnerabilities in Web Applications with LLM?
Di Cao (University of Science and Technology of China), Xiuwei Shang (University of Science and Technology of China)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the RealVul framework, which uses LLM to detect vulnerabilities in PHP web applications.
REAR: A Relevance-Aware Retrieval-Augmented Framework for Open-Domain Question Answering
Yuhao Wang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
RetrievalLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the REAR framework, which introduces an explicit relevance assessment module into retrieval-augmented generation (RAG) systems to improve the LLM's ability to judge and utilize the credibility of retrieved documents;
Reasoning in Token Economies: Budget-Aware Evaluation of LLM Reasoning Strategies
Junlin Wang (Duke University), Ben Athiwaratkun (Together AI)
Computational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose an evaluation framework based on computational budget (number of queries, tokens, cost) to uniformly compare multiple LLM inference strategies, revealing the significant impact of budget on performance.
Reasoning or a Semblance of it? A Diagnostic Study of Transitive Reasoning in LLMs
Houman Mehrafarin (Heriot-Watt University), Ioannis Konstas (Heriot-Watt University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: The study investigates the actual reasoning ability of LLMs in transitive reasoning through diagnostic experiments on the QASC and Bamboogle datasets.
Reasoning Paths with Reference Objects Elicit Quantitative Spatial Reasoning in Large Vision-Language Models
Yuan-Hong Liao (University of Toronto, Vector Institute), David Acuna (NVIDIA)
TransformerPrompt EngineeringVision Language ModelImageMultimodalityPoint CloudBenchmarkChain-of-Thought
🎯 What it does: Explored the capabilities of large-scale vision-language models in quantitative spatial reasoning, and proposed the human-annotated Q-Spatial Bench benchmark along with the zero-shot prompt SpatialPrompt.
Reasoning Robustness of LLMs to Adversarial Typographical Errors
Esther Gan (National University of Singapore), Michael Shieh (National University of Singapore)
Adversarial AttackTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Studies the robustness of large language models (LLM) in chain-of-thought (CoT) reasoning, proposes an Adversarial Typo Attack (ATA) algorithm targeting typing errors in inputs, and constructs a cross-model, cross-dataset robustness benchmark named R ATA 2.
Rebuilding ROME : Resolving Model Collapse during Sequential Model Editing
Akshat Gupta (UC Berkeley), Gopala Anumanchipalli (UC Berkeley)
TransformerLarge Language ModelText
🎯 What it does: Reimplement Rank-One Model Editing (ROME) and propose r-ROME to address the issue of model crashes (disabling edits) during continuous editing.
ReCaLL: Membership Inference via Relative Conditional Log-Likelihoods
Roy Xie (Duke University), Bhuwan Dhingra (Duke University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a membership inference attack method called RECALL, which leverages the conditional language modeling capabilities of large language models by adding non-member prefixes to target text to detect whether the text belongs to the model's pre-training data.
RECANTFormer: Referring Expression Comprehension with Varying Numbers of Targets
Bhathiya Hemanthage (Heriot-Watt University), Oliver Lemon (Heriot-Watt University)
Object DetectionTransformerVision Language ModelMultimodality
🎯 What it does: Proposed RECANTFormer, a single-stage, decoder-free Transformer model for Generalized Referring Expression Comprehension (GREC) tasks;
Reconsidering Sentence-Level Sign Language Translation
Garrett Tanzer (Google), David Uthus (Google)
VideoText
🎯 What it does: This paper investigates the necessity of shifting tasks from sentence-level to longer discourse-level in sign language machine translation, and verifies the limitations of sentence-level translation through human benchmark experiments on the How2Sign dataset.
Reconstruct Your Previous Conversations! Comprehensively Investigating Privacy Leakage Risks in Conversations with GPT Models
Junjie Chu (CISPA Helmholtz Center for Information Security), Yang Zhang (CISPA Helmholtz Center for Information Security)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Explores and evaluates the risk of malicious users reconstructing and leaking previous user conversation content during multi-round dialogues with GPT models (e.g., GPT-3.5, GPT-4), proposes two advanced attack methods (UNR and PBU), and introduces three defense strategies based on LLM internal mechanisms (Prompt-Based, Few-Shot-Based, Composite Defense).
Recurrent Alignment with Hard Attention for Hierarchical Text Rating
Chenxi Lin (East China Normal University), Xiaomin Zhu (AMS)
ClassificationTransformerLarge Language ModelText
🎯 What it does: Propose a framework named RAHA based on two large language models for hierarchical text scoring tasks.
Red Teaming Language Models for Processing Contradictory Dialogues
Xiaofei Wen (University of California, Davis), Muhao Chen (University of California, Davis)
GenerationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a new task of detecting and modifying dialog contradictions, and constructs a dataset containing 12,387 dialogues (6,130 containing contradictions), accompanied by annotations on contradiction positions and reasons;
Related Work and Citation Text Generation: A Survey
Xiangci Li (University of Texas at Dallas), Jessica Ouyang (University of Texas at Dallas)
GenerationTransformerTextReview/Survey Paper
🎯 What it does: This paper systematically reviews the current state of research in the field of related work generation (RWG), organizing the task definition, method classification, dataset sources, evaluation methods, and challenges faced.
Relevance Is a Guiding Light: Relevance-aware Adaptive Learning for End-to-end Task-oriented Dialogue System
Zhanpeng Chen (Peking University), Yuexian Zou (Peking University)
GenerationRetrievalTransformerContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposes a two-stage Relevance-aware Adaptive Learning (ReAL) framework for end-to-end task-oriented dialogue systems, first eliminating hard negative samples through adaptive pre-training, then aligning retrieval and generation via metric distribution alignment.
Repairs in a Block World: A New Benchmark for Handling User Corrections with Multi-Modal Language Models
Javier Chiyah-Garcia (Heriot-Watt University), Arash Eshghi (Heriot-Watt University)
Robotic IntelligenceSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper investigates and addresses the third-position repair (TPR) problem in multimodal dialogue, constructing the BLOCKWORLD-REPAIRS dataset and evaluating the performance of multimodal vision-language models in instruction following and repair tasks.
RepEval: Effective Text Evaluation with LLM Representation
Shuqian Sheng (Shanghai Jiao Tong University), Chenghu Zhou (Chinese Academy Of Sciences)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes RepEval, a metric that evaluates text quality by leveraging the projection of hidden representations from large language models, supporting both absolute evaluation and comparative evaluation;
RepMatch: Quantifying Cross-Instance Similarities in Representation Space
Mohammad Reza Modarres (Tehran Institute for Advanced Studies, Khatam University), Mohammad Taher Pilehvar (Cardiff University)
Representation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the RepMatch method, which quantifies the similarity of training instance subsets (even individual instances) in the model's representation space by comparing the Grassmann similarity of LoRA adaptation matrices fine-tuned on subsets.
Representational Analysis of Binding in Language Models
Qin Dai (Tohoku University), Kentaro Inui (MBZUAI)
Explainability and InterpretabilityRepresentation LearningTransformerText
🎯 What it does: Study the internal binding mechanisms in language models, identify the low-rank subspace (OI subspace), and demonstrate its causal impact on entity-attribute binding behaviors.
Resampled Datasets Are Not Enough: Mitigating Societal Bias Beyond Single Attributes
Yusuke Hirota (Osaka University), Alice Xiang (Sony AI)
ClassificationObject DetectionGenerationData SynthesisVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Developed a synthetic data generation process based on text-guided inpainting to eliminate social bias in image-text datasets.
Rethinking Pragmatics in Large Language Models: Towards Open-Ended Evaluation and Preference Tuning
Shengguang Wu (Peking University), Qi Su (Peking University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodality
🎯 What it does: This paper proposes changing traditional multiple-choice evaluation to open-ended free-response evaluation, and improves the social pragmatic reasoning ability of large language models through preference optimization (PO) rather than supervised fine-tuning (SFT).
Rethinking Pruning Large Language Models: Benefits and Pitfalls of Reconstruction Error Minimization
Sungbin Shin (POSTECH), Namhoon Lee (POSTECH)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Focusing on sparsification pruning for large language models (LLMs), the authors re-examine the current 'blocking-stepwise reconstruction' method, propose multiple reconstruction techniques to significantly reduce reconstruction error, and investigate the trade-off between reconstruction error and model generalization.
Rethinking the Evaluation of In-Context Learning for LLMs
Guoxin Yu (Institute of Computing Technology, Chinese Academy of Sciences), Xiang Ao (Institute of Computing Technology, Chinese Academy of Sciences)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper explores the impact of example configuration cost in in-context learning (ICL) of large language models (LLMs) on performance, and proposes a two-dimensional evaluation framework;
Rethinking the Reversal Curse of LLMs: a Prescription from Human Knowledge Reversal
Zhicong Lu (Chinese Academy of Sciences), Xunliang Cai (Meituan)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Study and address the 'reversal curse' issue that large language models encounter when facing 'reversed relationships'
Rethinking the Role of Proxy Rewards in Language Model Alignment
Sungdong Kim (NAVER Cloud), Minjoon Seo (KAIST AI)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Construct an interpretable white-box reward function through reverse reward engineering, using features such as length, repetition penalty, query relevance, and query type branches to align large language models, and verify its ability to replicate the monotonic relationship of gold standard rewards.
Rethinking Token Reduction for State Space Models
Zheng Zhan (Northeastern University), Yanzhi Wang (Northeastern University)
Computational EfficiencyTextBenchmark
🎯 What it does: A unified post-training token compression method for state space models (SSM), particularly Mamba, is proposed, combining token importance and similarity to achieve refined trimming and merging;
Retrieval-enriched zero-shot image classification in low-resource domains
Nicola Dall’Asen, Elisa Ricci (University of Trento)
ClassificationRetrievalLarge Language ModelVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes a zero-shot image classification method called CORE, which enhances the representations of query images and class prototypes by retrieving relevant texts from a large-scale text-image database, achieving better classification performance in low-resource visual domains.
Retrieve-Plan-Generation: An Iterative Planning and Answering Framework for Knowledge-Intensive LLM Generation
Yuanjie Lyu, Enhong Chen (Anhui Conch Information Technology Engineering Co Ltd)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes the Retrieve-Plan-Generation (RPG) framework, integrating iterative planning and retrieval into LLM generation to enhance relevance and accuracy for knowledge-intensive generation tasks.
Retrieved In-Context Principles from Previous Mistakes
Hao Sun (Peking University), Fei Huang (Alibaba Group)
Computational EfficiencyKnowledge DistillationLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built a teacher-student framework called RICP, where the teacher model analyzes student errors and generates high-level reasons and fine-grained prompts. Subsequently, errors are hierarchically clustered to obtain task-level and problem-level learning principles, which are then concatenated into existing prompts to enhance the reasoning performance of large language models.
Retrieved Sequence Augmentation for Protein Representation Learning
Chang Ma (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
Representation LearningProtein Structure PredictionTransformerSupervised Fine-TuningBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval sequence augmentation (RSA) method for incremental learning, which obtains homologous or structurally similar sequences via dense retrieval, directly concatenates them to the original sequence, and encodes them using a Transformer to enhance protein representation learning.
Retrospex: Language Agent Meets Offline Reinforcement Learning Critic
Yufei Xiang (Nanjing University), Cam-Tu Nguyen (Nanjing University)
Recurrent Neural NetworkTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Construct an LLM agent framework called Retrospex, which rescores past experiences using an offline RL critic to improve decision quality without expanding the LLM context.
Reuse Your Rewards: Reward Model Transfer for Zero-Shot Cross-Lingual Alignment
Zhaofeng Wu (Massachusetts Institute of Technology), Ahmad Beirami
Representation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper studies the use of source language reward models (RMs) in a multilingual setting for zero-shot cross-lingual alignment, i.e., directly using an RM trained in another language for reinforcement learning or best-of-n re-ranking in the target language, to improve the language model's alignment with human preferences.
Reusing Transferable Weight Increments for Low-resource Style Generation
Chunzhen Jin (Northeastern University), Osmar Zaiane (University of Alberta)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose the TWIST framework, achieving low-resource text style transfer by reusing transferable weight increments
Revealing Personality Traits: A New Benchmark Dataset for Explainable Personality Recognition on Dialogues
Lei Sun (Renmin University of China), Qin Jin (Renmin University of China)
RecognitionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Propose the explainable personality recognition task, constructing a dialogue-based explainable personality dataset named PersonalityEvd, and defining two subtasks: EPR-S (state recognition) and EPR-T (trait recognition);
Revealing the Parallel Multilingual Learning within Large Language Models
Yongyu Mu (Northeastern University), JingBo Zhu
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: The paper proposes and verifies the use of parallel multilingual input (PMI) for cross-lingual context learning in large language models (LLMs), exploring its impact on model performance and neural activation patterns.
Reverse-Engineering the Reader
Samuel Kiegeland (ETH Zürich), Ryan Cotterell (ETH Zürich)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a new alignment method: by fine-tuning a language model to implicitly optimize the parameters of a linear regressor, enabling the model's self-information (surprisal) to better predict psychometric data such as human reading time.
Revisiting Automated Evaluation for Long-form Table Question Answering
Yuqi Wang (Independent Researcher), Yilun Zhao (Yale University)
TransformerLarge Language ModelTabularBenchmark
🎯 What it does: Constructed a meta-evaluation dataset called LFTQA-Eval containing 2988 human-annotated examples, and systematically evaluated the reliability of existing automatic evaluation metrics for long-table question answering.
Revisiting Supertagging for faster HPSG parsing
Olga Zamaraeva (Universidade da Coruña), Carlos Gómez-Rodríguez (Universidade da Coruña)
Computational EfficiencyRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Multiple HPSG super taggers are constructed using SVM, LSTM+CRF, and BERT, with the best model integrated into the ACE parser to enhance parsing speed and accuracy.
Revisiting Supervised Contrastive Learning for Microblog Classification
Junbo Huang (University of Hamburg), Ricardo Usbeck (Leuphana University)
ClassificationTransformerSupervised Fine-TuningText
🎯 What it does: In the Weibo text classification task, the authors fine-tune RoBERTa-base and incorporate a linear combination of supervised contrastive learning (SCL) and cross-entropy loss in the loss function.
Revisiting the Robustness of Watermarking to Paraphrasing Attacks
Saksham Rastogi (Indian Institute of Science), Danish Pruthi (Indian Institute of Science)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper investigates the robustness of text watermarks against rewriting attacks and successfully extracts a green list through reverse engineering methods, further enhancing the effectiveness of rewriting attacks.
Revisiting Who’s Harry Potter: Towards Targeted Unlearning from a Causal Intervention Perspective
Yujian Liu (University of California Santa Barbara), Shiyu Chang (University of California Santa Barbara)
Explainability and InterpretabilityKnowledge DistillationTransformerPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a targeted 'forgetting' task for large language models, constructs a new benchmark WPU, and improves and generalizes the Who's Harry Potter (WHP) method based on causal intervention theory, proposing a more complete targeted forgetting algorithm.
RevMUX: Data Multiplexing with Reversible Adapters for Efficient LLM Batch Inference
Yige Xu (Joint NTU-UBC Research Centre of Excellence in Active Living for the Elderly), Chunyan Miao (Joint NTU-UBC Research Centre of Excellence in Active Living for the Elderly)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose a reversible adapter framework called RevMUX, which achieves batch inference acceleration through multiplexing without fine-tuning large language models.
Right for Right Reasons: Large Language Models for Verifiable Commonsense Knowledge Graph Question Answering
Armin Toroghi (University of Toronto), Scott Sanner (University of Toronto)
Explainability and InterpretabilityTransformerLarge Language ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a verifiable common-sense knowledge graph question-answering framework, R³, which leverages the inherent common-sense knowledge of large language models (LLMs) and strictly grounds reasoning steps on knowledge graph triplets, forming a tree-structured search process;
RLHF Can Speak Many Languages: Unlocking Multilingual Preference Optimization for LLMs
John Dang (Cohere For AI), Sara Hooker (Cohere For AI)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Preference optimization for large language models in multilingual environments, constructing high-quality synthetic comparison data across 23 languages, training and aligning the Aya-23 8B model, exploring differences between online and offline RLHF methods, and demonstrating cross-lingual transfer effects.
RoCEL: Advancing Table Entity Linking through Distinctive Row and Column Contexts
Yuanzheng Wang (Chinese Academy Of Sciences Key Laboratory Of Network Data Science And Technology), Xueqi Cheng (Chinese Academy Of Sciences Key Laboratory Of Network Data Science And Technology)
RetrievalTransformerLarge Language ModelTabularBenchmark
🎯 What it does: This paper proposes a new table entity linking method called RoCEL, which explicitly distinguishes and separately models row context and column context to enhance the performance of table entity disambiguation tasks.
Roleplay-doh: Enabling Domain-Experts to Create LLM-simulated Patients via Eliciting and Adhering to Principles
Ryan Louie (Stanford University), Diyi Yang (Stanford University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Developed the Roleplay-doh tool to assist mental health professionals in collaboratively creating AI patients through human-AI collaboration, supporting the automatic transformation of expert qualitative feedback into natural language principles that guide LLMs in role-playing.
RoseLoRA: Row and Column-wise Sparse Low-rank Adaptation of Pre-trained Language Model for Knowledge Editing and Fine-tuning
Haoyu Wang (SUNY Albany), Jing Gao (Purdue University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes Row and Column-wise Sparse Low-Rank Adaptation (RoseLoRA), which incorporates row and column-wise sparse constraints within the LoRA framework, enabling the update of only the most critical parameters for specific tasks, thereby achieving efficient knowledge editing and fine-tuning of pre-trained language models.
RoTBench: A Multi-Level Benchmark for Evaluating the Robustness of Large Language Models in Tool Learning
Junjie Ye (Fudan University), Xuanjing Huang (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed the RoTBench multi-level noise benchmark for evaluating the robustness of LLMs in tool learning, and developed the RoTTuning training strategy based on this, significantly enhancing the adaptability of LLMs to different noise environments.
RSA-Control: A Pragmatics-Grounded Lightweight Controllable Text Generation Framework
Yifan Wang (Saarland University), Vera Demberg (Saarland University)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the RSA-Control framework, which utilizes speaker-listener recursive reasoning to achieve zero-training controllable text generation, and evaluates it on tasks such as toxicity reduction, bias mitigation, and readability summarization.
RuBLiMP: Russian Benchmark of Linguistic Minimal Pairs
Ekaterina Taktasheva (University of Edinburgh), Vladislav Mikhailov (University of Oslo)
TextBenchmark
🎯 What it does: Constructed and released the Russian minimal contrastive sentence pairs (RuBLiMP) benchmark, covering 45 syntactic/morphological/semantic phenomena, with 45k pairs, providing automated generation, decontamination, and manual verification processes.
RULE: Reliable Multimodal RAG for Factuality in Medical Vision Language Models
Peng Xia, Huaxiu Yao (Unc Chapel Hill)
Large Language ModelSupervised Fine-TuningVision Language ModelBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose the RULE method, combining retrieval-augmented generation, provable fact risk control, and preference tuning, significantly improving the factual accuracy of Med-LVLMs.
RWKV-CLIP: A Robust Vision-Language Representation Learner
Tiancheng Gu (University of Sydney), Jiankang Deng (Imperial College)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a diversity description generation framework based on LLM, and designed the first visual-language pre-training model using the RWKV architecture, named RWKV-CLIP.
Safely Learning with Private Data: A Federated Learning Framework for Large Language Model
Jia-Ying Zheng, Zhi-Ming Zheng
Federated LearningSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Designed the FL-GLM framework, which splits the LLM into input/output blocks on the client side and large blocks on the server side, achieving secure and efficient federated learning through encrypted communication;
Safety Arithmetic: A Framework for Test-time Safety Alignment of Language Models by Steering Parameters and Activations
Rima Hazra (Singapore University of Technology and Design), Soujanya Poria (Singapore University of Technology and Design)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose the SAFETY ARITHMETIC framework, utilizing two steps—Harm Direction Removal and Safety Alignment—to achieve safety alignment during testing across three scenarios: base model, SFT (Supervised Fine-Tuning), and editing model;
Satyrn: A Platform for Analytics Augmented Generation
Marko Sterbentz (Northwestern University), Kristian J Hammond (Northwestern University)
GenerationTransformerLarge Language ModelTextTabularElectronic Health Records
🎯 What it does: Propose the SATYRN platform to achieve analysis-enhanced generation (AAG), generating fact sets by analyzing structured databases and guiding LLMs to produce accurate and coherent reports.
SaySelf: Teaching LLMs to Express Confidence with Self-Reflective Rationales
Tianyang Xu (Purdue University), Jing Gao (Purdue University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Train LLM to provide fine-grained confidence and self-reflective reasoning justifications.
Scalable Data Ablation Approximations for Language Models through Modular Training and Merging
Clara Na (Allen Institute for AI), Pradeep Dasigi (Allen Institute for AI)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Propose a method to approximate large-scale language model data ablation through modular training and parameter averaging, significantly reducing experimental costs.
Scalable Efficient Training of Large Language Models with Low-dimensional Projected Attention
Xingtai Lv, Bowen Zhou (Tsinghua University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed a low-dimensional projection attention (LPA) structure that applies low-rank modules only to the attention layer, significantly reducing the number of parameters and computational overhead while maintaining or even improving model performance.
Scaling Laws Across Model Architectures: A Comparative Analysis of Dense and MoE Models in Large Language Models
Siqi Wang (Meituan Inc), Jingang Wang (Meituan Inc)
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Compare the scaling laws of Dense and Mixture of Experts (MoE) models, verifying and unifying the power-law relationships among training loss, batch size, and learning rate.
Scaling Laws for Linear Complexity Language Models
Xuyang Shen (OpenNLPLab), Yiran Zhong (OpenNLPLab)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Conduct large-scale pretraining of three linear-complexity language models (TNL, HGRN2, cosFormer2) across scales from 70M to 7B, establishing power-law scaling patterns between training loss, computational budget, and model/data size, and evaluating them on multiple downstream tasks (validation perplexity, common sense reasoning, retrieval generation) with traditional Transformers (LLaMA).
Scaling Properties of Speech Language Models
Santiago Cuervo (Université de Toulon), Ricard Marxer (Université de Toulon)
Data SynthesisComputational EfficiencyData-Centric LearningTransformerLarge Language ModelTextAudio
🎯 What it does: Trained over 50 Speech Language Models (SLMs) under different scales and data budgets, evaluating their upstream loss and downstream syntactic and semantic performance, and proposed a new synthetic corpus STINYSTORIES to assess the impact of coarse-grained speech segmentation on performance.
Scaling Synthetic Logical Reasoning Datasets with Context-Sensitive Declarative Grammars
Damien Sileo (University of Lille)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a multilingual (simplified English and TPTP) first-order logic (FOL) reasoning dataset generation framework called Unigram, based on declarative context-sensitive grammar, and trained NLI models using the generated FOL-NLI dataset to enhance logical reasoning capabilities.
ScalingFilter: Assessing Data Quality through Inverse Utilization of Scaling Laws
Ruihang Li (University of Science and Technology of China), Houwen Peng (Microsoft Research Asia)
Data-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Designed and validated a no-reference quality filtering method called ScalingFilter, which evaluates and filters high-quality text by leveraging the perplexity difference between language models of different scales, avoiding reference data bias and enhancing data diversity.
SciAgent: Tool-augmented Language Models for Scientific Reasoning
Yubo Ma (Nanyang Technological University), Aixin Sun (Nanyang Technological University)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed a tool-based scientific reasoning framework SCIAGENT, constructed a tool-enhanced training set MATHFUNC and evaluation benchmark SCITOOLBENCH, aiming to enhance LLM's scientific reasoning capabilities through tool usage.
SciDQA: A Deep Reading Comprehension Dataset over Scientific Papers
Shruti Singh (Iit Gandhinagar), Arman Cohan
Large Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the SCIDQA dataset, consisting of 2,937 question-answer pairs extracted from OpenReview peer review discussions. The dataset underwent multi-stage refinement, including LLM extraction, human screening, and decontextualization, aiming to test models' deep understanding of scientific papers.
SciER: An Entity and Relation Extraction Dataset for Datasets, Methods, and Tasks in Scientific Documents
Qi Zhang (Temple University), Eduard Dragut (Temple University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed a full-text scientific literature entity and relation extraction dataset called SciER, focusing on three categories of entities (datasets, methods, and tasks) and their fine-grained relationships.
SciEx: Benchmarking Large Language Models on Scientific Exams with Human Expert Grading and Automatic Grading
Tu Anh Dinh (Karlsruhe Institute of Technology), Jan Niehues (Karlsruhe Institute of Technology)
TransformerLarge Language ModelMultimodalityBenchmark
🎯 What it does: Introduce the SciEx benchmark, which evaluates the capabilities of large language models on scientific tasks using university-level computer science exam questions, and provides both expert scoring and automatic scoring.
SciPrompt: Knowledge-augmented Prompting for Fine-grained Categorization of Scientific Topics
Zhiwen You (University Of Illinois Urbana Champaign), Jana Diesner (Technical University Of Munich)
ClassificationTransformerSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes SCIPROMPT, a prompt-based fine-tuning framework that automatically retrieves and expands scientific terminology for fine-grained scientific text classification in low-resource scenarios.
SCOI: Syntax-augmented Coverage-based In-context Example Selection for Machine Translation
Chenming Tang (Peking University), Yunfang Wu (Peking University)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a context example method SCOI based on alternating selection of syntactic and lexical coverage to enhance the ICL performance of large language models in machine translation.
Scope-enhanced Compositional Semantic Parsing for DRT
Xiulin Yang (Georgetown University), Johan Bos (University of Groningen)
Representation LearningText
🎯 What it does: This paper proposes the AMS parser, a neural symbolic compositional semantic parsing framework based on the AM parser, specifically designed for parsing in Discourse Representation Theory (DRT), with the addition of a new quantifier scope prediction mechanism.
SEACrowd: A Multilingual Multimodal Data Hub and Benchmark Suite for Southeast Asian Languages
Holy Lovenia (AI Singapore), Samuel Cahyawijaya (Cohere)
Data-Centric LearningTransformerPrompt EngineeringImageTextMultimodalityBenchmarkAudio
🎯 What it does: Built the SEACrowd resource center, integrating approximately 500 datasets, 399 standardized data loaders, covering nearly 1000 Southeast Asian languages, and created a multimodal benchmark based on 13 tasks (covering 36 languages); simultaneously conducted zero-shot evaluation of existing LLM, VLM, and ASR models.
Searching for Best Practices in Retrieval-Augmented Generation
Xiaohua Wang (Fudan University), Xuanjing Huang (Fudan University)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper explores the optimal implementation methods for each module in the Retrieval-Augmented Generation (RAG) framework through systematic experiments, and proposes two practical best practices (maximum performance and balanced efficiency) along with multimodal retrieval expansion;
SecCoder: Towards Generalizable and Robust Secure Code Generation
Boyu Zhang (Zhejiang University), Jianwei Yin (Zhejiang University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a secure code generation framework named SecCoder, which utilizes dense retrieval to select the most relevant security code demonstrations and enhances the security of large models' code generation by incorporating these demonstrations into the input to leverage contextual learning.
Seeing the Forest through the Trees: Data Leakage from Partial Transformer Gradients
Weijun Li (Macquarie University), Mark Dras (Macquarie University)
Safty and PrivacyAdversarial AttackTransformerText
🎯 What it does: This paper investigates the problem of training data leakage in Transformer models during distributed training, where using only partial gradients can lead to data leakage.
SEEKR: Selective Attention-Guided Knowledge Retention for Continual Learning of Large Language Models
Jinghan He (Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences), Jinqiao Wang (Foundation Model Research Center, Institute of Automation, Chinese Academy of Sciences)
Safty and PrivacyKnowledge DistillationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes the SEEKR method, achieving continuous learning in LLMs through selective attention head knowledge retention;
Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?
Neeladri Bhuiya (National University of Singapore), Stefan Winkler (ASUS Intelligent Cloud Services (AICS))
Adversarial AttackTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: The study investigates how large language models are easily misled by 'plausible but incorrect' distractor paths in multi-hop reasoning tasks, and proposes an evaluation method based on generable feasible distractor paragraphs.
SEER: Self-Aligned Evidence Extraction for Retrieval-Augmented Generation
Xinping Zhao (Harbin Institute of Technology (Shenzhen)), Min Zhang (Harbin Institute of Technology (Shenzhen))
GenerationRetrievalComputational EfficiencyTransformerReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed and implemented the SEER framework for self-aligned evidence extraction in retrieval-augmented generation, reducing computational costs and improving answer quality.
Seg2Act: Global Context-aware Action Generation for Document Logical Structuring
Zichao Li (Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Reformulate the document structuring task as one-pass action generation, where a generative language model outputs three types of actions (add title, add paragraph, connect text) in one go to construct the document's hierarchical tree.
Segment Any Text: A Universal Approach for Robust, Efficient and Adaptable Sentence Segmentation
Markus Frohmann (Johannes Kepler University Linz), Markus Schedl (Johannes Kepler University Linz)
SegmentationDomain AdaptationComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Proposed a sentence segmentation model named SAT, capable of achieving robust, scalable, and efficient sentence segmentation in texts lacking punctuation, with varying capitalization, cross-lingual contexts, and multi-domain scenarios. Rapid adaptation to target domains is achieved through additional supervised fine-tuning (SAT+SM) and low-rank adaptation (SAT+LORA).