EMNLP 2024 Papers — Page 11
Conference on Empirical Methods in Natural Language Processing · 1268 papers
SEGMENT+: Long Text Processing with Short-Context Language Models
Wei Shi (Fudan University), Yanghua Xiao (Fudan University)
Explainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the SEGMENT+ framework, which enables short-context language models to efficiently process long texts through structured notes and filtering modules.
Selective Vision is the Challenge for Visual Reasoning: A Benchmark for Visual Argument Understanding
Jiwan Chung (Yonsei University), Youngjae Yu (Yonsei University)
RecognitionObject DetectionLarge Language ModelVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Constructed the VisArgs dataset and proposed three evaluation tasks for visual argumentation (localization, identification, deduction).
Self-AMPLIFY: Improving Small Language Models with Self Post Hoc Explanations
Milan Bhan (Sorbonne Université, CNRS, LIP6), Marie-Jeanne Lesot (Sorbonne Université, CNRS, LIP6)
Explainability and InterpretabilityTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes Self-AMPLIFY, a framework that enhances the performance of small autoregressive language models (SLM) in in-context learning (ICL) by automatically generating reasoning steps using self-model post-hoc explanations.
Self-Bootstrapped Visual-Language Model for Knowledge Selection and Question Answering
Dongze Hao (Institute of Automation, Chinese Academy of Sciences), Jing Liu (Institute of Automation, Chinese Academy of Sciences)
RetrievalLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Utilize large vision-language models to construct a knowledge selector and answer generator, enhancing the synergy between knowledge retrieval and question answering through self-guided cyclic training.
Self-Powered LLM Modality Expansion for Large Speech-Text Models
Tengfei Yu (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
🎯 What it does: Propose a Self-Powered approach to expand LLM into a multimodal speech-text model (LSM) using self-supervised instruction texts containing only ASR data, and address the model's speech anchoring bias through attention analysis to enhance instruction following and speech-text fusion capabilities.
Self-Refine Instruction-Tuning for Aligning Reasoning in Language Models
Leonardo Ranaldi (Idiap Research Institute), Andre Freitas
Knowledge DistillationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes a "Self-Refine Instruction-tuning" method, which first fine-tunes a small model using chain-of-thought (CoT) examples generated by a teacher large model, and then employs direct preference optimization (DPO) to enable the small model to self-improve its reasoning pathways, thereby aligning the small model's reasoning capabilities with those of the large model.
Self-Training for Sample-Efficient Active Learning for Text Classification with Pre-Trained Language Models
Christopher Schröder (Center for Scalable Data Analytics and Artificial Intelligence), Gerhard Heyer (Center for Scalable Data Analytics and Artificial Intelligence)
ClassificationTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: The study combines self-training with active learning to improve sample efficiency in text classification using pre-trained language models.
Self-Training Large Language and Vision Assistant for Medical Question Answering
Guohao Sun (Rochester Institute of Technology), Zhiqiang Tao (Rochester Institute of Technology)
TransformerLarge Language ModelVision Language ModelBiomedical DataBenchmark
🎯 What it does: Propose a two-stage self-training framework called STLLaVA-Med, which first lets the large vision-language model (LVLM) automatically generate medical question-answer pairs, then uses GPT-4o to generate preference labels and performs fine-tuning via Direct Preference Optimization (DPO), thereby improving the model's reasoning and answering quality in medical visual question answering.
Semantic Training Signals Promote Hierarchical Syntactic Generalization in Transformers
Aditya Yedetore (Boston University), Najoung Kim (Boston University)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: This paper investigates how Transformers, after training on both form and semantics, can better generalize to hierarchical syntactic rules;
Semantics and Sentiment: Cross-lingual Variations in Emoji Use
Giulio Zhou (University of Edinburgh), Sumin Zhao (University of Edinburgh)
TransformerLarge Language ModelText
🎯 What it does: This study conducted two experiments to collect the literal meanings of emojis and their literal/metaphorical uses in sentences across three languages: English, Portuguese, and Chinese, and explored their relationship with emotions.
Semformer: Transformer Language Models with Semantic Planning
Yongjing Yin (Zhejiang University), Yue Zhang (School of Engineering, Westlake University)
Representation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Propose the Semformer model, which introduces learnable planning tokens into Transformer language models and enhances the model's prospective reasoning by predicting the latent semantic representations of future text through an autoencoder;
Sequential API Function Calling Using GraphQL Schema
Avirup Saha (IBM Research), Sameep Mehta (IBM Research)
AI Code AssistantLarge Language ModelPrompt EngineeringTextGraphBenchmarkChain-of-Thought
🎯 What it does: Constructed a REST API call sequence dataset called GraphQLRestBench based on GraphQL schema, and proposed new evaluation metrics and frameworks for sequential function calls.
ShadowLLM: Predictor-based Contextual Sparsity for Large Language Models
Yash Akhauri (Cornell University), Mohamed S. Abdelfattah (Cornell University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the ShadowLLM method, utilizing gradient-based pruning criteria and a first-layer predictor to achieve context sparsification in large language models.
Shaking Up VLMs: Comparing Transformers and Structured State Space Models for Vision & Language Modeling
Georgios Pantazopoulos (Heriot-Watt University), Arash Eshghi (Heriot-Watt University)
TransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Replace the Transformer with the Mamba structured state space model (SSM) to construct a vision-language model (VLM), and evaluate it on multiple vision-language tasks.
SHIELD: Evaluation and Defense Strategies for Copyright Compliance in LLM Text Generation
Xiaoze Liu (Purdue University), Jing Gao (Purdue University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Proposed the SHIELD framework to assess and safeguard copyright compliance in LLM-generated texts, constructed an evaluation dataset containing copyrighted, non-copyrighted, and partially copyrighted texts, and designed an Agent-based defense mechanism;
Shortcuts Arising from Contrast: Towards Effective and Lightweight Clean-Label Attacks in Prompt-Based Learning
Xiaopeng Xie (Beijing University of Posts and Telecommunications), Joey Tianyi Zhou (A*STAR)
ClassificationAdversarial AttackTransformerPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a lightweight, effective, and stealthy clean-label backdoor attack method based on contrastive shortcut injection (CSI), specifically designed for prompt-based learning scenarios.
Show and Guide: Instructional-Plan Grounded Vision and Language Model
Diogo Glória-Silva (NOVA School of Science and Technology), Joao Magalhaes
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose MM-PlanLLM, a multimodal large language model that supports multimodal input and output, generates steps in plan-guided dialogues, retrieves video clips, and generates the next step based on user-uploaded images.
SignCLIP: Connecting Text and Sign Language by Contrastive Learning
Zifan Jiang (University Of Zurich), Sarah Ebling
RetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: Propose SignCLIP, a contrastive learning model that maps spoken text and sign language videos into a shared embedding space;
SimLLM: Detecting Sentences Generated by Large Language Models Using Similarity between the Generation and its Re-generation
Hoang-Quoc Nguyen-Son (National Institute of Information and Communications Technology), Koji Zettsu (National Institute of Information and Communications Technology)
ClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose a detection method called SimLLM based on the similarity between original sentences and their proofread versions. The method generates multiple proofread sentences using large language models, sorts them by similarity, and then inputs them into RoBERTa for discrimination.
Simul-MuST-C: Simultaneous Multilingual Speech Translation Corpus Using Large Language Model
Mana Makinae (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
Data SynthesisTransformerLarge Language ModelTextMultimodalityBenchmarkAudio
🎯 What it does: Rewrite the MuST-C dataset using the Salami technique with GPT-4o to construct the Simul-MuST-C multilingual simultaneous speech translation corpus, and train it on the end-to-end SiST model.
Simultaneous Interpretation Corpus Construction by Large Language Models in Distant Language Pair
Yusuke Sakai (Nara Institute of Science and Technology), Taro Watanabe (Nara Institute of Science and Technology)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextAudio
🎯 What it does: Use large language models (LLMs) to convert existing speech translation corpora into approximately simultaneous interpretation (SI) style corpora following the Chunk-wise Monotonic Translation (CWMT) guidelines, constructing the LLM-SI-Corpus, and using this corpus to fine-tune a simultaneous machine translation (SiMT) model.
Simultaneous Masking, Not Prompting Optimization: A Paradigm Shift in Fine-tuning LLMs for Simultaneous Translation
Matthew Raffel (Oregon State University), Lizhong Chen (Oregon State University)
GenerationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a new Fine-tuning paradigm called SimulMask, which simulates parallel translation directly within LLMs through attention masks, avoiding issues such as training-inference mismatch, position confusion, and high computational costs caused by traditional prompting optimization methods.
SLANG: New Concept Comprehension of Large Language Models
Lingrui Mei (Institute of Computing Technology, Chinese Academy of Sciences), Xueqi Cheng (Institute of Computing Technology, Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the SLANG evaluation benchmark to measure large language models' understanding of internet slang and memes, and introduce the FOCUS method (based on causal inference) to enhance models' reasoning and interpretation of new concepts.
Small Agent Can Also Rock! Empowering Small Language Models as Hallucination Detector
Xiaoxue Cheng (Renmin University of China), Ji-Rong Wen (Renmin University of China)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextChain-of-Thought
🎯 What it does: This paper proposes and implements HaluAgent—a self-autonomous hallucination detection framework based on a small open-source language model, integrating a multifunctional toolbox, a three-stage fine-grained detection process, and a memory mechanism. The model is fine-tuned using GPT-4-synthesized bilingual detection trajectories, achieving hallucination detection performance comparable to or even better than GPT-4.
Small LLMs Are Weak Tool Learners: A Multi-LLM Agent
Weizhou Shen (Sun Yat Sen University), Fei Huang (Alibaba Group)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
🎯 What it does: Propose the α-UMi multi-LLM framework, decomposing tool learning tasks into three small models as planner, caller, and summarizer, significantly enhancing the tool usage capability of small models.
Social Bias Probing: Fairness Benchmarking for Language Models
Marta Marchiori Manerba (University of Pisa), Isabelle Augenstein (University of Copenhagen)
TransformerTextBenchmark
🎯 What it does: Propose a social bias detection framework that conducts fine-grained fairness evaluation of language models using the newly constructed SOFA benchmark.
SOUL: Unlocking the Power of Second-Order Optimization for LLM Unlearning
Jinghan Jia (Michigan State University), Sijia Liu (MIT-IBM Watson AI Lab, IBM Research)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Proposed a second-order optimization-based LLM forgetting framework called SOUL to enhance the forgetting effectiveness of large language models while maintaining their original utility.
SparseGrad: A Selective Method for Efficient Fine-tuning of MLP Layers
Viktoriia A. Chekalina (Artificial Intelligence Research Institute), Ivan Oseledets (Artificial Intelligence Research Institute)
Computational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes a sparse gradient fine-tuning method called SparseGrad for the MLP block of Transformer models, achieving parameter-efficient fine-tuning by leveraging the sparsity in the gradient space;
Speaking in Wavelet Domain: A Simple and Efficient Approach to Speed up Speech Diffusion Model
Xiangyu Zhang (University of New South Wales), Lina Yao (University of New South Wales)
GenerationData SynthesisComputational EfficiencyDiffusion modelAudio
🎯 What it does: This paper achieves speech data compression by changing the generation target of speech diffusion models from time-domain signals to discrete wavelet domain signals, thereby improving training and inference speeds by nearly two times without altering the model structure; simultaneously, it proposes two frontend modules, low-frequency enhancement and multi-level wavelet acceleration, to further enhance quality and speed;
SpecHub: Provable Acceleration to Multi-Draft Speculative Decoding
Ryan Sun (Lehigh University), Lichao Sun (Lehigh University)
OptimizationComputational EfficiencyTextBenchmark
🎯 What it does: Propose the SpecHub method, which leverages sparse two-draft joint distributions and linear programming to improve the acceptance rate and batch efficiency of multi-draft speculative decoding.
SpeechQE: Estimating the Quality of Direct Speech Translation
HyoJung Han (University of Maryland), Marine Carpuat (University of Maryland)
Convolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkAudio
🎯 What it does: For the speech translation quality estimation (SpeechQE) task, a benchmark dataset was constructed and two systems were implemented: end-to-end (E2E) and cascaded (ASR + text-QE). The systems can provide sentence-level quality scores for direct speech translation outputs and detect error segments.
Speechworthy Instruction-tuned Language Models
Hyundong Justin Cho, Jonathan May (University of Southern California, Information Sciences Institute)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Explored how to enable instruction-tuned language models (ITLM) to generate more suitable responses for voice interaction, proposing two methods: prompt engineering based on broadcast industry experience and preference learning using 20K voice preference data, combining both approaches.
SPEED++: A Multilingual Event Extraction Framework for Epidemic Prediction and Preparedness
Tanmay Parekh (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
TransformerLarge Language ModelText
🎯 What it does: Proposed the SPEED++ multilingual event extraction framework for extracting epidemic-related events from social media and enabling early warning and information aggregation
Split and Merge: Aligning Position Biases in LLM-based Evaluators
Zongjie Li (Hong Kong University of Science and Technology), Yang Liu (Nanyang Technological University)
Large Language ModelTextBenchmark
🎯 What it does: Proposed the PORTIA system, which eliminates position bias in LLM evaluation and improves consistency by segmenting answers, aligning lengths, and aligning semantics.
Sprout: Green Generative AI with Carbon-Efficient LLM Inference
Baolin Li (Northeastern University), Devesh Tiwari (Northeastern University)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextTime Series
🎯 What it does: Developed the SPROUT framework, which dynamically adjusts the generation length of LLMs by leveraging generation directives and carbon intensity information to reduce carbon emissions during the inference phase.
SRF: Enhancing Document-Level Relation Extraction with a Novel Secondary Reasoning Framework
Fu Zhang (Northeastern University), Yongxue Wu (Northeastern University)
TransformerTextBenchmark
🎯 What it does: Proposed a document-level relation extraction framework named Secondary Reasoning Framework (SRF).
Stable Language Model Pre-training by Reducing Embedding Variability
Woojin Chung (KAIST AI), Se-Young Yun (KAIST AI)
OptimizationRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a pre-training stability evaluation method based on Token Embedding Variability (TEV) and designs a multi-head low-rank attention (MLRA) to reduce TEV, enhancing model stability and performance.
StablePrompt : Automatic Prompt Tuning using Reinforcement Learning for Large Language Model
Minchan Kwon (KAIST), Junmo Kim (KAIST)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Propose a reinforcement learning-based automated prompt tuning method called StablePrompt, which trains an agent model on LLMs to search for optimal prompts, using Adaptive Proximal Policy Optimization (APPO) and anchor models to enhance training stability and expand the search space.
Standardize: Aligning Language Models with Expert-Defined Standards for Content Generation
Joseph Marvin Imperial (University of Bath), Harish Tayyar Madabushi (University of Bath)
GenerationLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes the STANDARDIZE framework, which utilizes retrieval-based contextual learning to align the language model's generation process with expert-established standards (e.g., CEFR, CCS), generating text that conforms to pedagogical guidelines.
STAR: SocioTechnical Approach to Red Teaming Language Models
Laura Weidinger (Google DeepMind), William Isaac (Google DeepMind)
Adversarial AttackPrompt EngineeringText
🎯 What it does: This paper proposes the STAR framework for socially technical red team testing of large language models, combining parameterizable instructions, demographically matched evaluators, and an arbitration mechanism to systematically enhance the steerability and signal quality of red teaming.
Statistical Uncertainty in Word Embeddings: GloVe-V
Andrea Vallebueno (Stanford University), Daniel E. Ho (Stanford University)
Representation LearningText
🎯 What it does: Proposed the GloVe-V method, which assigns reconstruction error variance estimates to GloVe word embeddings to achieve statistical uncertainty quantification in downstream tasks.
Step-by-Step Reasoning to Solve Grid Puzzles: Where do LLMs Falter?
Nemika Tyagi (Arizona State University), Chitta Baral (Arizona State University)
Large Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the GridPuzzle dataset and assess the performance of LLMs in reasoning chains and final answers for grid puzzles
Stepwise Verification and Remediation of Student Reasoning Errors with Large Language Model Tutors
Nico Daheim (Technical University of Darmstadt), Mrinmaya Sachan (ETH Zurich)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Split the generation of teacher responses into two steps: first detecting errors (verification) in the student's problem-solving steps, then generating targeted feedback based on the verification results.
Still Not Quite There! Evaluating Large Language Models for Comorbid Mental Health Diagnosis
Amey Hengle (Indian Institute of Technology Bombay), Rashmi Gupta (Sophia College for Women)
ClassificationTransformerLarge Language ModelTextBenchmark
🎯 What it does: This study proposes ANGST, a new benchmark for the classification of comorbid depression and anxiety based on social media posts. Unlike existing datasets, ANGST supports multi-label classification, allowing each post to be labeled as both/depression and/or anxiety simultaneously.
STOP! Benchmarking Large Language Models with Sensitivity Testing on Offensive Progressions
Robert Morabito (Brock University), Ali Emami (Brock University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Built and evaluated the STOP dataset to detect bias in LLMs through progressively escalating offensive scenarios; conducted idealized and real-world performance assessments of multiple closed-source and open-source models; and improved answer rates in known bias evaluation tasks using Llama 3-70b trained on STOP.
Story Embeddings — Narrative-Focused Representations of Fictional Stories
Hans Ole Hatzel (Universitat Hamburg), Chris Biemann (Universitat Hamburg)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed the StoryEmb model, which trains story embeddings using contrastive learning to make summaries of similar narratives have close vectors.
Story Morals: Surfacing value-driven narrative schemas using large language models
David G Hobson (McGill University), Andrew Piper (McGill University)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes a narrative structure annotation task based on the concept of 'story morality,' utilizing large language models to automatically extract and verify values and lessons across various text types (fairy tales, novels, movies/tv shows, social media personal stories, news);
StorySparkQA: Expert-Annotated QA Pairs with Real-World Knowledge for Children’s Story-Based Learning
Jiaju Chen (East China Normal University), Yuling Sun (East China Normal University)
Data SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Created the StorySparkQA dataset, which contains 5,868 expert-annotated question-answer pairs based on children's stories, integrating real-world knowledge;
STORYSUMM: Evaluating Faithfulness in Story Summarization
Melanie Subbiah (Columbia University), Kathleen McKeown (Columbia University)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the STORYSUMM dataset for evaluating the faithfulness of narrative text summarization, constructed with extended golden labels through multiple human annotation protocols.
Strategic Demonstration Selection for Improved Fairness in LLM In-Context Learning
Jingyu Hu (University of Bristol), Mengnan Du (New Jersey Institute of Technology)
ClassificationTransformerLarge Language ModelTabular
🎯 What it does: This paper investigates the impact of selecting different demonstration samples on the fairness of binary classification tasks with tabular data in the context of large language models (LLM) in-context learning (ICL), and proposes an efficient demonstration selection method based on clustering-genetic algorithm (FCG) to improve the balance between prediction performance and fairness.
Strength Lies in Differences! Improving Strategy Planning for Non-collaborative Dialogues via Diversified User Simulation
Tong Zhang (Sichuan University), Tat-Seng Chua (National University of Singapore)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Designed and validated a customized strategic planning method TRIP for non-cooperative dialogue, which can generate personalized strategies based on users' psychological characteristics and behavioral patterns.
Strengthening Structural Inductive Biases by Pre-training to Perform Syntactic Transformations
Matthias Lindemann (University of Edinburgh), Ivan Titov (University of Edinburgh)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose an intermediate pre-training method on Transformer, enabling the model to learn syntactic transformations based on dependency trees, thereby reinforcing structural inductive bias;
Structure Guided Prompt: Instructing Large Language Model in Multi-Step Reasoning by Exploring Graph Structure of the Text
Kewei Cheng (Amazon), Yizhou Sun (Intel Labs)
TransformerLarge Language ModelPrompt EngineeringTextGraphChain-of-Thought
🎯 What it does: Propose a zero-shot, task-agnostic three-stage prompting framework (Structure Guided Prompt), which enhances the multi-step reasoning capabilities of large language models (LLMs) by converting natural language text into graph structures, planning navigation paths, and executing reasoning.
Structured Optimal Brain Pruning for Large Language Models
Jiateng Wei (Zhejiang University), Yong Liu (Zhejiang University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose a structured pruning method called SoBP that does not require retraining. For large language models, it first selects pruning structures based on global importance, then refines them locally with a greedy approach, and finally performs module-level reconstruction to maintain performance.
Studying and Mitigating Biases in Sign Language Understanding Models
Katherine Atwell (Northeastern University), Malihe Alikhani (Northeastern University)
RecognitionConvolutional Neural NetworkGraph Neural NetworkVideoTabular
🎯 What it does: Analyze biases in gender, skin color, age, etc., in the ASL Citizen dataset, release participant demographic information, and experiment with various bias mitigation techniques based on this.
Style-Shifting Behaviour of the Manosphere on Reddit
Jai Aggarwal (University of Toronto), Suzanne Stevenson (University of Toronto)
ClassificationRecognitionTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Systematically studied the language style characteristics of Men's Rights Activists (Manosphere) on Reddit, and evaluated their style transfer and infiltration in 14 non-Men's Rights subreddits.
Style-Specific Neurons for Steering LLMs in Text Style Transfer
Wen Lai (Technical University of Munich), Alexander Fraser (Technical University of Munich)
GenerationDomain AdaptationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose the sNeuron-TST framework, which first identifies and removes source-style-specific neurons, then combines contrastive decoding (Dola) to improve the generation performance of LLMs in text style transfer.
StyleRemix: Interpretable Authorship Obfuscation via Distillation and Perturbation of Style Elements
Jillian Fisher (University of Washington), Yejin Choi (University of Washington)
Safty and PrivacyExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose an interpretable author anonymization method called STYLEREMIX, which uses LoRA modules to perturb the fine-grained style axes of text, enabling targeted alterations to the original author's writing style; simultaneously released two new datasets: AUTHORMIX (30K+ paragraphs, 14 authors, 4 domains) and DISC (1.5K texts, 16 style directions).
Subjective Topic meets LLMs: Unleashing Comprehensive, Reflective and Creative Thinking through the Negation of Negation
Fangrui Lv (Tsinghua University), Changshui Zhang (Kuaishou Technology)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Create the Subjective Topic Benchmark (SJTP) and propose the NeoN framework based on the philosophical principle of 'negation of negation' to enhance large language models' (LLM) performance in synthesis, reflection, and creative thinking.
Subword Segmentation in LLMs: Looking at Inflection and Consistency
Marion Di Marco (Technische Universität München), Alexander Fraser (Technische Universität München)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Analyze the subword segmentation strategies of large language models (e.g., GPT-4o), and evaluate the segmentation quality through morphological analysis across ten languages, investigating the impact of segmentation quality on semantic capture and morphological generation.
Successfully Guiding Humans with Imperfect Instructions by Highlighting Potential Errors and Suggesting Corrections
Lingjun Zhao (University of Maryland), Hal Daumé III (University of Maryland)
Autonomous DrivingExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelContrastive LearningTextMultimodality
🎯 What it does: Developed the HEAR system to automatically detect and annotate hallucinations (errors) in language instructions during visual navigation tasks, providing correction suggestions to assist humans in completing navigation.
Summary of a Haystack: A Challenge to Long-Context LLMs and RAG Systems
Philippe Laban (Salesforce AI Research), Chien-Sheng Wu (Salesforce AI Research)
GenerationData SynthesisRetrievalLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes and implements a long-text retrieval and generation evaluation benchmark called SummHay. It first synthesizes a document collection (Haystack) containing approximately 1 million tokens using LLMs, with controlled predefined insights (insight) embedded in each document. Subsequently, a precise evaluation protocol is designed, requiring models to generate summaries that include corresponding insights and accurately cite source documents based on given queries. Finally, the results are automatically evaluated using three metrics: Coverage, Citation, and Joint.
SUPER: Evaluating Agents on Setting Up and Executing Tasks from Research Repositories
Ben Bogin (Allen Institute for AI), Tushar Khot (Allen Institute for AI)
AI Code AssistantTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposes the SUPER benchmark to evaluate the ability of large language models to set up, configure, and execute complete experiments in low-attention research repositories.
SURf: Teaching Large Vision-Language Models to Selectively Utilize Retrieved Information
Jiashuo Sun (Xiamen University), Yu Cheng (Chinese University of Hong Kong)
ClassificationRecognitionGenerationRetrievalSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes the self-improving SURf framework, which trains large vision-language models (LVLMs) to selectively utilize retrieved information and suppress irrelevant or misleading retrieval content, enhancing the robustness and performance of RAG tasks.
Surprise! Uniform Information Density Isn’t the Whole Story: Predicting Surprisal Contours in Long-form Discourse
Eleftheria Tsipidi (ETH Zürich), Alex Warstadt (ETH Zürich)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Propose the Structured Context Hypothesis, predicting surprisal contours in text through hierarchical discourse structures (RST and ordinary prose structures), and quantitatively evaluate its effectiveness using Bayesian linear regression.
Surveying the Dead Minds: Historical-Psychological Text Analysis with Contextualized Construct Representation (CCR) for Classical Chinese
Yuqi Chen (Peking University), Mohammad Atari (University of Massachusetts Amherst)
Representation LearningTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Developed and validated the Contextualized Construct Representation (CCR) pipeline for extracting psychological constructs from classical Chinese texts, and constructed the first Chinese historical psychology corpus (C-HI-PSY).
Susu Box or Piggy Bank: Assessing Cultural Commonsense Knowledge between Ghana and the US
Christabel Acquaye (University of Maryland), Rachel Rudinger (University of Maryland)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed and evaluated a cross-cultural common sense question-answering dataset, AMAMMER E, containing 525 multiple-choice questions focusing on cultural differences between Ghana and the United States.
Symbolic Working Memory Enhances Language Models for Complex Rule Application
Siyuan Wang, Xiang Ren (University Of Southern California)
Explainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposes a neuro-symbolic framework based on external working memory for multi-step rule application, enhancing the reasoning ability of large language models.
Synchronous Faithfulness Monitoring for Trustworthy Retrieval-Augmented Generation
Di Wu (University of California, Los Angeles), Kai-Wei Chang (University of California, Los Angeles)
GenerationExplainability and InterpretabilityTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the SYNCHECK monitoring mechanism and the FOD decoding method for real-time detection and enhancing the authenticity of retrieval-augmented generation models.
Synergizing In-context Learning with Hints for End-to-end Task-oriented Dialog Systems
Vishal Vivek Saley (Indian Institute of Technology), Mausam .
GenerationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: By combining the in-context learning of large language models with task prompts (entity type, response length, dialogue closure), an end-to-end task-oriented dialogue system named SyncTOD was constructed to enhance response alignment and performance in low-data scenarios.
SYNFAC-EDIT: Synthetic Imitation Edit Feedback for Factual Alignment in Clinical Summarization
Prakamya Mishra (University of Massachusetts Amherst), Hong Yu (University of Massachusetts Amherst)
TransformerLarge Language ModelSupervised Fine-TuningBiomedical DataElectronic Health Records
🎯 What it does: By using GPT-3.5 and GPT-4 with over 100B parameters as synthetic experts, incrementally generating edit feedback (edit instructions and corresponding summaries), and subsequently utilizing these edit feedback to perform DPO or SALT alignment training on smaller LLMs such as GPT-2 (1.5B) and Llama-2 (7B), aiming to enhance factual consistency in clinical summaries.
SynthesizRR: Generating Diverse Datasets with Retrieval Augmentation
Abhishek Divekar (Amazon), Greg Durrett (University of Texas at Austin)
Data SynthesisRetrievalKnowledge DistillationTransformerPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes SYNTHESIZRR, a text dataset synthesis method based on retrieval enhancement. It uses a teacher LLM to perform task inversion on retrieved documents, generating diverse synthetic samples, which are then used for distillation training of the student model.
Synthetic Knowledge Ingestion: Towards Knowledge Refinement and Injection for Enhancing Large Language Models
Jiaxin Zhang (Intuit AI Research), Sricharan Kumar (Intuit AI Research)
RetrievalData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Proposed a Synthetic Knowledge Ingestion (Ski) method that transforms raw knowledge text into high-quality QA or QA+Context formats directly usable by large language models, thereby enhancing knowledge injection effectiveness.
Systematic Biases in LLM Simulations of Debates
Amir Taubenfeld (Hebrew University of Jerusalem), Ariel Goldstein (Google)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Construct LLM agents and conduct political debate simulations, observe the agents' attitude changes during the debates, and study how implicit biases in LLMs affect the simulation outcomes.
T-FREE: Subword Tokenizer-Free Generative LLMs via Sparse Representations for Memory-Efficient Embeddings
Björn Deiseroth (Aleph Alpha), Samuel Weinbach (Aleph Alpha)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposed T-FREE, a tokenizer-free sparse representation method that directly activates word embeddings using hashed character triplets, enabling text encoding and decoding without requiring a reference corpus.
Table Question Answering for Low-resourced Indic Languages
Vaishali Pal (University of Amsterdam), Maarten de Rijke (University of Amsterdam)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTabularBenchmark
🎯 What it does: Propose a fully automatic, low-cost, large-scale table question answering data generation process, constructing two datasets, BanglaTabQA and HindiTabQA, for low-resource languages (Bangla and Hindi), and training and evaluating multiple models on these datasets.
Tag-grounded Visual Instruction Tuning with Retrieval Augmentation
Daiqing Qi (University of Virginia), Sheng Li (University of Virginia)
RecognitionObject DetectionRetrievalTransformerSupervised Fine-TuningVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval-enhanced label-oriented visual instruction tuning framework (TUNA), which improves the performance of multimodal models in recognizing novel objects, entities, and detailed descriptions by retrieving labels from a large-scale image-text database and combining them with an image-aware label encoder.
Take Off the Training Wheels! Progressive In-Context Learning for Effective Alignment
Zhenyu Liu (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The study investigates the impact of examples on token representations in alignment tasks and proposes a two-stage progressive context learning method (PICA) to reduce reliance on examples and improve alignment performance.
Target-Aware Language Modeling via Granular Data Sampling
Ernie Chang (AI at Meta), Vikas Chandra (AI at Meta)
Safty and PrivacyComputational EfficiencyRepresentation LearningData-Centric LearningText
🎯 What it does: Propose a target-aware data sampling method based on multi-granularity n-gram features, utilizing importance sampling to finely select large-scale pre-trained corpora, thereby achieving language model performance comparable to the complete RefinedWeb data using only about 1% of the data.
Task Arithmetic can Mitigate Synthetic-to-Real Gap in Automatic Speech Recognition
Hsuan Su (National Taiwan University), Hung-yi Lee (National Taiwan University)
RecognitionDomain AdaptationTransformerAudio
🎯 What it does: This paper proposes SYN2REAL, a task vector that encodes the differences between synthetic and real speech into model parameters, enabling ASR adaptation to unseen domains.
Task Oriented In-Domain Data Augmentation
Xiao Liang (Tsinghua University), Jian Jiao (Tsinghua University)
Data SynthesisDomain AdaptationData-Centric LearningLarge Language ModelText
🎯 What it does: Proposes the TRAIT framework, which achieves continuous pre-training through domain data selection and task-oriented synthetic corpus, thereby enhancing the performance of LLMs on downstream tasks in the advertising and mathematics domains.
Taxonomy-guided Semantic Indexing for Academic Paper Search
SeongKu Kang (University of Illinois at Urbana Champaign), Hwanjo Yu (Pohang University of Science and Technology)
RetrievalGraph Neural NetworkTransformerMixture of ExpertsContrastive LearningTextGraph
🎯 What it does: Proposed the TaxoIndex framework, which leverages academic taxonomy trees to guide semantic indexing and enhance academic paper retrieval performance.
Taylor Unswift: Secured Weight Release for Large Language Models via Taylor Expansion
Guanchu Wang (Rice University), Xia Hu (Rice University)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose TaylorMLP, which converts LLM weights into Taylor series parameters, preserving model ownership while increasing computational load to reduce generation speed and prevent abuse.
TCSinger: Zero-Shot Singing Voice Synthesis with Style Transfer and Multi-Level Style Control
Yu Zhang (Zhejiang University, Shanghai AI Laboratory), Zhou Zhao (Zhejiang University, Shanghai AI Laboratory)
GenerationTransformerLarge Language ModelDiffusion modelAudio
🎯 What it does: Propose TCSinger, a zero-shot singing voice synthesis model that supports cross-lingual speech and singing style transfer and multi-level style control;
Teaching Embodied Reinforcement Learning Agents: Informativeness and Diversity of Language Use
Jiajun Xi (University of Michigan), Joyce Chai (University of Michigan)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper combines language feedback with offline reinforcement learning to investigate how the 'informativeness' and 'diversity' of language influence the learning and adaptability of embodied agents;
Teaching LLMs to Abstain across Languages via Multilingual Feedback
Shangbin Feng (University of Washington), Yulia Tsvetkov (University of Washington)
Representation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes using multilingual feedback to enable large language models to self-reflect in multi-lingual question answering, thereby enhancing their ability to refuse inappropriate queries.
Teaching Small Language Models Reasoning through Counterfactual Distillation
Tao Feng (Zhejiang University), Yin Zhang (Zhejiang University)
Knowledge DistillationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes a method based on adversarial data augmentation and multi-perspective chain-of-thought (CoT) distillation, transferring the reasoning capabilities of large language models to small language models, and significantly improving their reasoning accuracy and generalization performance on out-of-distribution samples.
TEMA: Token Embeddings Mapping for Enriching Low-Resource Language Models
Rodolfo Zevallos (Universitat Pompeu Fabra), Mireia Farrús (Universitat de Barcelona)
Knowledge DistillationRepresentation LearningTransformerTextBenchmark
🎯 What it does: This paper proposes a Token Embedding Mapping Algorithm (TEMA), which significantly improves the quality of word vectors in low-resource language models by mapping the word embeddings of a richly pre-trained model L1 to corresponding positions in a low-resource model L2, thereby reducing perplexity and enhancing downstream task performance.
TempoFormer: A Transformer for Temporally-aware Representations in Change Detection
Talia Tseriotou (Queen Mary University of London), Maria Liakata (Queen Mary University of London)
ClassificationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextTime Series
🎯 What it does: Propose TempoFormer, a Transformer model for time-aware change detection, which can simultaneously model word-level, post-level, and time-series-level representations of text units without using recursion.
Temporally Consistent Factuality Probing for Large Language Models
Ashutosh Bajpai (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: Proposed the TeCFaP task and the TEMP-COFAC dataset to evaluate the temporal factual consistency of large language models, and developed the CoTSeLF framework (multi-task instruction tuning + consistency time-sensitive reinforcement learning) to enhance the model's temporal factual consistency.
Text Fluoroscopy: Detecting LLM-Generated Text through Intrinsic Features
Xiao Yu (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
Anomaly DetectionTransformerLarge Language ModelText
🎯 What it does: Proposed a black-box method called Text Fluoroscopy that extracts the intrinsic features of text by leveraging differences in the lexical space distribution of intermediate layers, thereby detecting text generated by large language models (LLMs).
Text Grafting: Near-Distribution Weak Supervision for Minority Classes in Text Classification
Letian Peng (University of California San Diego), Jingbo Shang (University of California San Diego)
ClassificationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes the Text Grafting framework to generate minority class samples close to the original corpus distribution in extremely weakly supervised text classification, combining the advantages of text mining and generative large models to improve the quality and distribution consistency of minority class samples.
Text-Tuple-Table: Towards Information Integration in Text-to-Table Generation via Global Tuple Extraction
Zheye Deng (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextTabularChain-of-Thought
🎯 What it does: Studied the text-to-table generation task, proposed the LIVESUM dataset to evaluate information integration capability, and designed a three-step Text-Tuple-Table (T3) prompting process.
Text2Chart31: Instruction Tuning for Chart Generation with Automatic Feedback
Fatemeh Pesaran Zadeh (Seoul National University), Gunhee Kim (Seoul National University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageTextMultimodality
🎯 What it does: Constructed a Text2Chart31 dataset covering 31 Matplotlib chart types, including descriptions, code, data tables, and images, and proposed an RL-based instruction tuning method (Preference Reward + Alignment Reward) to enhance LLM performance in chart generation tasks.
The Accuracy Paradox in RLHF: When Better Reward Models Don’t Yield Better Language Models
Yanjun Chen (Hong Kong Polytechnic University), Xiaoyu Shen (Eastern Institute of Technology)
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningText
🎯 What it does: Under the RLHF framework, the impact of reward model accuracy on language model performance was studied, revealing that moderate-accuracy reward models often yield better results;
The Best Defense is Attack: Repairing Semantics in Textual Adversarial Examples
Heng Yang (University of Exeter), Ke Li (University of Exeter)
ClassificationAdversarial AttackTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes the RAPID framework, which repairs the semantics of text adversarial samples by integrating adversarial detectors and perturbation focusing technology within a pre-trained language model.
The Computational Anatomy of Humility: Modeling Intellectual Humility in Online Public Discourse
Xiaobo Guo (Dartmouth College), Soroush Vosoughi (Dartmouth College)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper constructs a 'intellectual humility/intellectual arrogance' codebook tailored for online religious discussions, and manually annotated 350 posts to train and evaluate automatic detection methods based on large language models (LLMs).
The Death and Life of Great Prompts: Analyzing the Evolution of LLM Prompts from the Structural Perspective
Yihan Ma (CISPA Helmholtz Center for Information Security), Yang Zhang (CISPA Helmholtz Center for Information Security)
Large Language ModelPrompt EngineeringText
🎯 What it does: Structurally decompose and analyze 10,538 real-world LLM prompts, and construct an eight-component framework along with 1,168 annotated samples.
The effects of distance on NPI illusive effects in BERT
So Young Lee (Miami University), Mai Ha Vu (University of Toronto)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This study uses the BERT model to investigate the negative polarity item (NPI) illusion effect and experimentally examines the impact of distance, further distinguishing the roles of hierarchical distance and linear distance;
The Emergence of Compositional Languages in Multi-entity Referential Games: from Image to Graph Representations
Daniel Akkerman, Raquel G. Alhama
GenerationRepresentation LearningRecurrent Neural NetworkGraph Neural NetworkReinforcement LearningImageGraph
🎯 What it does: Propose a multi-entity reference game and investigate the impact of graph structures and image inputs on language emergence