ACL 2024 Papers — Page 3
Annual Meeting of the Association for Computational Linguistics · 940 papers
DeVAn: Dense Video Annotation for Video-Language Models
Tingkai Liu (ByteDance), Hongxia Yang (Chinese Academy of Sciences)
GenerationRetrievalData-Centric LearningTransformerVision Language ModelContrastive LearningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: This paper creates the DeVAn dataset, containing 8.5k YouTube video clips, each annotated by five human annotators with one short caption and 3-10 long summaries, to evaluate the capabilities of vision-language models in video-text generation and text-video retrieval; meanwhile, multiple models are benchmarked, and a new evaluation metric BLEURT is proposed, which better aligns with human preferences; additionally, an end-to-end model based on VideoCoCa is enhanced with ASR encoding to provide retrieval baselines.
DIALECTBENCH: An NLP Benchmark for Dialects, Varieties, and Closely-Related Languages
Fahim Faisal (George Mason University), Antonios Anastasopoulos (George Mason University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed a large-scale NLP benchmark across dialects and variants—DIALECTBENCH, covering 40 language families, 281 dialects/variants, and 10 text tasks.
Dialogue Summarization with Mixture of Experts based on Large Language Models
Yuanhe Tian (University of Science and Technology of China), Yan Song (University of Science and Technology of China)
GenerationTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a Mixture-of-Experts (MoE) framework based on large language models, which selects experts via role-oriented routing (RoR) and synthesizes the final summary through fusion generation (FG);
DiffuCOMET: Contextual Commonsense Knowledge Diffusion
Silin Gao (EPFL), Antoine Bosselut (Sony Group Corporation)
GenerationData SynthesisTransformerDiffusion modelText
🎯 What it does: Developed a contextual common sense knowledge generation model called DIFFUCOMET based on diffusion models, which can automatically generate diverse and relevant common sense facts given a narrative context.
Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines
Michael Toker (Technion Israel Institute Of Technology), Yonatan Belinkov (Technion Israel Institute Of Technology)
GenerationExplainability and InterpretabilityTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes the DIFFUSION LENS method, which generates corresponding images by taking hidden states from each Transformer layer of the text encoder and directly inputting them into a diffusion model, thereby visualizing and analyzing the internal computational process of the text encoder.
DiFiNet: Boundary-Aware Semantic Differentiation and Filtration Network for Nested Named Entity Recognition
Yuxiang Cai (University of Electronic Science and Technology of China), JiayeYang JiayeYang
RecognitionTransformerTextBenchmark
🎯 What it does: Proposed and implemented the Boundary‑Aware Semantic Differentiation and Filtration Network (DiFiNet) framework for nested named entity recognition, aiming to improve boundary detection accuracy and noise suppression capabilities.
Digital Socrates: Evaluating LLMs through Explanation Critiques
Yuling Gu (Allen Institute for AI), Peter Clark (Allen Institute for AI)
Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed and implemented the 'explanation critiquing' task, constructed the DS Critique Bank dataset, and trained a model called Digital Socrates capable of automatically evaluating the quality of LLM explanations.
Direct Large Language Model Alignment Through Self-Rewarding Contrastive Prompt Distillation
Aiwei Liu (Tsinghua University), Lijie Wen (Tsinghua University)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a fully automatic alignment method for LLMs called DLMA, which does not require human-annotated preference data. It generates response pairs using contrastive prompts and aligns them through probabilistic comparison of self-assessment scores.
Direct Metric Optimization for Image Captioning through Reward-Weighted Augmented Data Utilization
Takumi Takada (SoftBank Corp), Kazuya Ueki (Meisei University)
GenerationOptimizationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: Propose a method called Direct Metric Optimization (DMO) that directly optimizes the final evaluation metrics for image caption generation by replacing the exploration process in reinforcement learning with offline text data augmentation and using reward-weighted self-supervised training to directly optimize the metrics.
Disambiguate Words like Composing Them: A Morphology-Informed Approach to Enhance Chinese Word Sense Disambiguation
Yue Wang (Peking University), Yang Liu (Peking University)
ClassificationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed a large-scale publicly available Chinese word sense disambiguation resource (morphological-lexical dictionary MorInv, dictionary WrdInv, corpus MiCLS, and OOV test set), and proposed the MorBERT model that leverages morphological and lexical knowledge for word sense disambiguation.
Discursive Socratic Questioning: Evaluating the Faithfulness of Language Models’ Understanding of Discourse Relations
Yisong Miao (National University of Singapore), Min-Yen Kan (National University of Singapore)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the DISQ (Discursive Socratic Questioning) method, which evaluates the faithfulness of large language models (LLMs) in understanding discourse relations through a series of question-and-answer interactions; and systematically assess the discourse reasoning capabilities of multiple LLMs based on this framework.
Disentangled Learning with Synthetic Parallel Data for Text Style Transfer
Jingxuan Han (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
GenerationData SynthesisTransformerLarge Language ModelContrastive LearningTextChain-of-Thought
🎯 What it does: Propose the DisenTrans framework, achieving text style transfer through attribute/content splitting; generate and filter high-quality parallel data using Chain-of-Thought (CoT), followed by training with contrastive and seq2seq losses.
Disinformation Capabilities of Large Language Models
Ivan Vykopal (Faculty of Information Technology, Brno University of Technology), Maria Bielikova (Kempelen Institute of Intelligent Technologies)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Evaluate the capability of large language models in generating fake news articles
Dissecting Human and LLM Preferences
Junlong Li (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
Large Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Perform a fine-grained decomposition of dialogue preferences between humans and 32 large language models, quantifying the impact of multiple attributes (such as correctness, length, and harmlessness) on preferences, and exploring the manipulability of different model sizes, alignment training, and evaluation benchmarks.
Distributional Inclusion Hypothesis and Quantifications: Probing for Hypernymy in Functional Distributional Semantics
Chun Hei Lo (Chinese University of Hong Kong), Guy Emerson (University of Cambridge)
Representation LearningAuto EncoderTextBenchmark
🎯 What it does: Investigate whether the Functional Distributional Semantics (FDS) model can learn hierarchical relationships (hypernyms) from corpora and experimentally verify its performance on synthetic and real-world corpora.
DM-BLI: Dynamic Multiple Subspaces Alignment for Unsupervised Bilingual Lexicon Induction
Ling Hu (Beijing Foreign Studies University), Yuemei Xu (Beijing Foreign Studies University)
Domain AdaptationRepresentation LearningContrastive LearningText
🎯 What it does: Studied a dynamic multi-subspace alignment framework DM-BLI for unsupervised bilingual lexicon induction.
Do Large Language Models Discriminate in Hiring Decisions on the Basis of Race, Ethnicity, and Gender?
Haozhe An (University of Maryland, College Park), Rachel Rudinger (University of Maryland, College Park)
TransformerLarge Language ModelText
🎯 What it does: Design 820 hiring decision email templates, perform large-scale reasoning on 300 common first names in the U.S. (categorized by race/ethnicity and gender), and statistically analyze the proportion of acceptance/rejection emails generated by LLMs to evaluate implicit bias in the model along the race/gender dimensions.
Do Large Language Models Latently Perform Multi-Hop Reasoning?
Sohee Yang (Google DeepMind), Sebastian Riedel (Google DeepMind)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper investigates whether large language models can implicitly perform multi-hop reasoning when processing two-step reasoning prompts, providing evidence through analysis of internal representations;
Do Llamas Work in English? On the Latent Language of Multilingual Transformers
Chris Wendler (École Polytechnique Fédérale de Lausanne), Robert West (École Polytechnique Fédérale de Lausanne)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Investigates whether Llama-2 uses English as an implicit 'pivot' language for non-English inputs, and analyzes the evolution of language preferences in its hidden representations across layers.
DocFinQA: A Long-Context Financial Reasoning Dataset
Varshini Reddy (Kensho Technologies), Chris Tanner
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: This paper proposes DocFinQA, a long document financial question-answering dataset based on complete SEC reports;
DocLens: Multi-aspect Fine-grained Evaluation for Medical Text Generation
Yiqing Xie (Carnegie Mellon University), Carolyn Rose
GenerationTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Proposed the DOCLENS evaluation framework, which can fine-grainedly assess the completeness, conciseness, and attribution of medical text generation results, and supports automatic metric calculation using multiple evaluators (NLI, open-source instruction models, GPT-4).
DocLLM: A Layout-Aware Generative Language Model for Multimodal Document Understanding
Dongsheng Wang (JPMorgan AI Research), Xiaomo Liu (JPMorgan AI Research)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Propose DocLLM, which leverages the spatial layout of OCR bounding boxes as a lightweight multimodal input to extend LLMs for document understanding.
DocMath-Eval: Evaluating Math Reasoning Capabilities of LLMs in Understanding Long and Specialized Documents
Yilun Zhao (Yale University), Arman Cohan (Yale University)
TextTabularBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed a benchmark named DOCMATH-EVAL to evaluate the numerical reasoning capabilities of large language models in understanding and analyzing long professional documents with mixed text and tables.
Document-level Claim Extraction and Decontextualisation for Fact-Checking
Zhenyun Deng (University of Cambridge), Andreas Vlachos (University of Cambridge)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a document-level claim extraction and decontextualization framework that can identify central claims from multi-sentence documents and rewrite them into independently understandable sentences;
Document-Level Machine Translation with Large-Scale Public Parallel Corpora
Proyag Pal (University of Edinburgh), Kenneth Heafield (University of Edinburgh)
TransformerText
🎯 What it does: Built a large-scale document-level parallel corpus and trained a context-aware machine translation model
Dodo: Dynamic Contextual Compression for Decoder-only LMs
Guanghui Qin (Johns Hopkins University), Benjamin Van Durme (Johns Hopkins University)
CompressionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes DODO (Dynamic Contextual Compression for Decoder-only LMs), a technique that dynamically compresses long texts into a few 'nugget' vectors, which can be used both as autoregressive language models and as context compressors;
Does DetectGPT Fully Utilize Perturbation? Bridging Selective Perturbation to Fine-tuned Contrastive Learning Detector would be Better
Shengchao Liu (Xi'an Jiaotong University), Chao Shen (Queen Mary University of London)
ClassificationAnomaly DetectionTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed PECOLA, a machine-generated text detection method combining perturbation and contrastive learning, to address the random perturbation and threshold dependency issues of DetectGPT.
Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research
Luca Soldaini, Kyle Lo (Allen Institute for AI)
Representation LearningData-Centric LearningLarge Language ModelText
🎯 What it does: Built and released an open English pre-training corpus named Dolma containing 3 trillion tokens, covering six sources including web pages, academic papers, code, books, social media, and providing high-performance data cleaning and mixing tools.
DolphCoder: Echo-Locating Code Large Language Models with Diverse and Multi-Objective Instruction Tuning
Yejie Wang (Beijing University of Posts and Telecommunications), Xunliang Cai (Meituan)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes DolphCoder, an instruction-tuning model that generates diverse responses through multiple system prompts and combines code evaluation objectives to enhance code generation capabilities.
Domain Adaptation for Subjective Induction Questions Answering on Products by Adversarial Disentangled Learning
Yufeng Zhang (Sun Yat-sen University), Jian Yin (Sun Yat-sen University)
Domain AdaptationTransformerReinforcement LearningContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper addresses the domain adaptation problem of product subjective question answering (SUBJPQA) in low-resource scenarios, proposing an adaptive model based on adversarial decomposition learning. The model can transfer knowledge from the source domain to the target domain and generate comprehensive answers that include objective facts and multi-perspective subjective opinions.
Don’t Buy it! Reassessing the Ad Understanding Abilities of Contrastive Multimodal Models
Anna Bavaresco (University of Amsterdam), Raquel Fernández (University of Amsterdam)
RetrievalRecommendation SystemData-Centric LearningVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: This paper re-examines the advertising understanding task, finding that existing evaluation settings can be exploited by models using anchoring cues between text and images to achieve high scores, and proposes a new adversarial test set called TRADE.
Don’t Go To Extremes: Revealing the Excessive Sensitivity and Calibration Limitations of LLMs in Implicit Hate Speech Detection
Min Zhang (Virginia Tech), Chang-Tien Lu (Texas A & M University-Corpus Christi)
ClassificationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: Evaluate the performance of large language models in implicit hate speech detection and confidence calibration, revealing issues of over-sensitivity to vulnerable groups and extreme confidence distributions.
Don’t Hallucinate, Abstain: Identifying LLM Knowledge Gaps via Multi-LLM Collaboration
Shangbin Feng (University of Washington), Yulia Tsvetkov (University of Washington)
TransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: Proposed and implemented an LLM 'give-up' mechanism to refuse answering when knowledge gaps occur, evaluated on multiple knowledge-intensive QA tasks.
Don’t Rank, Combine! Combining Machine Translation Hypotheses Using Quality Estimation
Giorgos Vernikos (École Polytechnique Fédérale de Lausanne), Andrei Popescu-Belis (École Polytechnique Fédérale de Lausanne)
GenerationTransformerLarge Language ModelText
🎯 What it does: Proposes the QE-fusion method, which combines high-quality segments from multiple candidate translations using quality estimation metrics to produce a better translation.
DoRA: Enhancing Parameter-Efficient Fine-Tuning with Dynamic Rank Distribution
Yulong Mao (Beijing Key Lab of Traffic Data Analysis and Mining), Jinan Xu (Beijing Key Lab of Traffic Data Analysis and Mining)
Computational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: Propose the DoRA method, achieving parameter-efficient fine-tuning based on dynamic low-rank distribution.
Dr.Academy: A Benchmark for Evaluating Questioning Capability in Education for Large Language Models
Yuyan Chen (Fudan University), Yanghua Xiao (Fudan University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the Dr.Academy benchmark to evaluate the ability of large language models (LLMs) to generate high-quality questions when acting as teachers in educational settings.
Draft & Verify: Lossless Large Language Model Acceleration via Self-Speculative Decoding
Jun Zhang, Sharad Mehrotra (University of California, Irvine)
Computational EfficiencyHyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: Proposes a self-speculative decoding scheme, where the original model skips certain layers during the drafting phase to generate a draft, which is then validated by the full model in one go, achieving inference acceleration.
DRAGIN: Dynamic Retrieval Augmented Generation based on the Real-time Information Needs of Large Language Models
Weihang Su (Tsinghua University), Yiqun Liu (Tsinghua University)
GenerationRetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a dynamic retrieval-augmented generation framework called DRAGIN, which can determine when and what to retrieve during text generation by large language models based on real-time information needs;
DUAL-REFLECT: Enhancing Large Language Models for Reflective Translation through Dual Learning Feedback Mechanisms
Andong Chen (Harbin Institute Of Technology), Min Zhang (Harbin Institute Of Technology)
GenerationLarge Language ModelText
🎯 What it does: Proposed and implemented the DUAL-REFLECT framework, leveraging large language models and bidirectional learning (back-translation) loops to achieve self-reflection and improve translation quality.
Dwell in the Beginning: How Language Models Embed Long Documents for Dense Retrieval
João Coelho (Carnegie Mellon University), Chenyan Xiong (University of Lisbon)
RetrievalTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Investigate the position bias of Transformer models in long-text retrieval, finding that embeddings pay more attention to the beginning of documents.
DynaSemble: Dynamic Ensembling of Textual and Structure-Based Models for Knowledge Graph Completion
Ananjan Nandi (Indian Institute of Technology Delhi), Mausam .
Graph Neural NetworkGraph
🎯 What it does: Propose a dynamic weighted ensemble method called DynaSemble, which utilizes the score distributions of text models and structural models to predict missing links in knowledge graphs.
EasyGen: Easing Multimodal Generation with BiDiffuser and LLMs
Xiangyu Zhao (Hong Kong Polytechnic University), Xiao-Ming Wu (Hong Kong Polytechnic University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelImageTextMultimodality
🎯 What it does: Built EasyGen, an end-to-end multi-modal generative model that integrates the bidirectional conditional diffusion model BiDiffuser with large language models (LLMs), achieving high-quality image-to-text and text-to-image generation, and supporting multi-modal dialogue.
ECBD: Evidence-Centered Benchmark Design for NLP
Yu Lu Liu (Mila - Quebec Artificial Intelligence Institute), Ziang Xiao (Microsoft Research)
TextBenchmark
🎯 What it does: Proposed and implemented the Evidence-Centered Benchmark Design (ECBD) framework for systematically designing and evaluating the effectiveness of NLP benchmarks, with case studies (BoolQ, SuperGLUE, HELM) validating the framework's feasibility.
EconAgent: Large Language Model-Empowered Agents for Simulating Macroeconomic Activities
Nian Li (Tsinghua University), Qingmin Liao (Tsinghua University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTabularFinance Related
🎯 What it does: Designed and implemented EconAgent, an economic agent based on large language models (LLMs), for macroeconomic simulation, modeling labor and consumer markets, and achieving human-like decision-making through perception, memory, and action modules.
Effects of diversity incentives on sample diversity and downstream model performance in LLM-based text augmentation
Jan Cegin (Brno University of Technology), Peter Brusilovsky (Kempelen Institute of Intelligent Technologies)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied the impact of incorporating diversity-promoting methods from crowdsourcing (such as banned words, prompts, and chains) during text augmentation with LLMs on the diversity of generated text and the performance of downstream models.
Efficient OCR for Building a Diverse Digital History
Jacob Carlson (Harvard University), Melissa Dell (Harvard University)
RecognitionRetrievalComputational EfficiencyConvolutional Neural NetworkTransformerContrastive LearningImageText
🎯 What it does: Propose EffOCR, an OCR system centered on character-level retrieval that achieves high-precision recognition for low-resource documents without relying on sequence language models.
EFSA: Towards Event-Level Financial Sentiment Analysis
Tianyu Chen (Key Laboratory of AI Safety, Chinese Academy of Sciences), Xiang Ao (Key Laboratory of AI Safety, Chinese Academy of Sciences)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkFinance RelatedChain-of-Thought
🎯 What it does: This paper proposes an event-level task for financial sentiment analysis (EFSA), converting event extraction into a classification task, and constructs a predictive framework containing five components: company, industry, coarse and fine-grained event categories, and sentiment polarity, while providing a corresponding large-scale Chinese dataset.
EIT: Enhanced Interactive Transformer
Tong Zheng (Northeastern University), JingBo Zhu
GenerationTransformerText
🎯 What it does: Proposed a new Transformer architecture called Enhanced Interactive Transformer (EIT), introducing multi-to-multi mapping (M2M), as well as internal subspace interaction (ISI) and cross-subspace interaction (CSI) within its multi-head self-attention mechanism to achieve collaborative consistency among attention heads.
Eliciting Better Multilingual Structured Reasoning from LLMs through Code
Bryan Li (University of Pennsylvania), Saab Mansour (/aws AI Labs)
AI Code AssistantTransformerSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a multilingual structured reasoning benchmark called xSTREET, extending the original STREET dataset to six languages; meanwhile, it designs two methods to enhance multilingual reasoning capabilities: first, performing LoRA fine-tuning (TCC) on source code comments after multilingual translation during training, and second, reconstructing the reasoning graph by using code-style prompts (SIM) during inference.
EmbSpatial-Bench: Benchmarking Spatial Understanding for Embodied Tasks with Large Vision-Language Models
Mengfei Du (Fudan University), Zhongyu Wei (Fudan University)
Data SynthesisRobotic IntelligenceSupervised Fine-TuningVision Language ModelImageTextMultimodalityPoint CloudMeshBenchmark
🎯 What it does: Constructed EmbSpatial-Bench, a benchmark for evaluating spatial understanding from an egocentric perspective, and generated the EmbSpatial-SFT instruction tuning dataset based on this benchmark to enhance the spatial perception capabilities of LVLM in embodied tasks.
Emergent Word Order Universals from Cognitively-Motivated Language Models
Tatsuki Kuribayashi (Mohamed bin Zayed University of Artificial Intelligence), Timothy Baldwin (Mohamed bin Zayed University of Artificial Intelligence)
Explainability and InterpretabilityData-Centric LearningRecurrent Neural NetworkTransformerLarge Language ModelText
🎯 What it does: By training various traditional and cognitively motivated language models (LMs), computing perplexity (a measure of difficulty) on 64 artificially generated word order configurations, and comparing them with real language word order frequencies (from WALS), to explain and simulate the universality of word order in world languages.
EmoBench: Evaluating the Emotional Intelligence of Large Language Models
Sahand Sabour (Tsinghua University), Minlie Huang (Tsinghua University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper constructs the EmoBench benchmark, manually designing 400 bilingual (Chinese-English) multiple-choice questions to evaluate the emotional intelligence performance of large language models in two aspects: emotional understanding and emotional application.
Empowering Character-level Text Infilling by Eliminating Sub-Tokens
Houxing Ren (Shanghai Jiao Tong University), Hongsheng Li (Shanghai Artificial Intelligence Laboratory)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a new text filling method called FIM-SE to eliminate confusion caused by subwords and support character-level filling;
Emulated Disalignment: Safety Alignment for Large Language Models May Backfire!
Zhanhui Zhou (Shanghai Artificial Intelligence Laboratory), Yu Qiao (Shanghai Artificial Intelligence Laboratory)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Construct a training-free attack method that generates more harmful responses by comparing the output token distributions of a safety-aligned model with its pre-trained version during inference.
End-to-end Learning of Logical Rules for Enhancing Document-level Relation Extraction
Kunxun Qi (Sun Yat-sen University), Hai Wan (Sun Yat-sen University)
Explainability and InterpretabilityRepresentation LearningTransformerSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed an end-to-end framework called JMRL for learning logical rules, jointly training a document-level relation extraction model and a rule reasoning module, enabling rule learning and extraction tasks to be optimized simultaneously within the same training process.
Enhancing Contrastive Learning with Noise-Guided Attack: Towards Continual Relation Extraction in the Wild
Ting Wu (Fudan University), Xuanjing Huang (Fudan University)
Adversarial AttackTransformerContrastive LearningText
🎯 What it does: This paper proposes a noise-robust contrastive learning framework called NaCL, designed for the noisy-CRE (noisy continuous relation extraction) scenario, enabling the learning of new relations while preventing catastrophic forgetting.
Enhancing Dialogue State Tracking Models through LLM-backed User-Agents Simulation
Xingguang Wang (University of Illinois Urbana-Champaign), Cheng Niu (University of Illinois Urbana-Champaign)
Data SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringText
🎯 What it does: This paper uses GPT-4 to simulate user-agent dialogues, generating synthetic multi-turn dialogues with dialogue state tags, and performs two-stage fine-tuning on LLaMA-2 using this data to improve dialogue state tracking performance.
Enhancing EEG-to-Text Decoding through Transferable Representations from Pre-trained Contrastive EEG-Text Masked Autoencoder
Jiaqi Wang (Harbin Institute of Technology), Zhiguo Zhang (Fudan University)
Representation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningTextBiomedical Data
🎯 What it does: Proposed a pretraining model CET-MAE based on EEG and text alignment, and an EEG-to-Text decoding framework E2T-PTR that utilizes this model.
Enhancing Explainable Rating Prediction through Annotated Macro Concepts
Huachi Zhou (Hong Kong Polytechnic University), Xiao Huang (Hong Kong Polytechnic University)
Recommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: By using macro concepts as mediators, the information bottleneck is employed to extract micro features from user reviews, and macro concepts are annotated using LLM to enhance user/item embeddings and generate interpretable rating predictions.
Enhancing In-Context Learning via Implicit Demonstration Augmentation
Xiaoling Zhou (Peking University), Shikun Zhang (Peking University)
ClassificationMeta LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a method called IDAICL for implicit augmentation of examples in in-context learning, achieving demonstration enhancement by performing semantic direction sampling on demonstrations in the depth feature space, thereby improving the accuracy and stability of large pre-trained language models in few-shot reasoning.
Enhancing Large Language Models in Coding Through Multi-Perspective Self-Consistency
Baizhou Huang (Wangxuan Institute of Computer Technology, Peking University), Nan Duan (Microsoft Research Asia)
AI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: By letting LLM generate outputs from three perspectives—solutions, specifications, and test cases—and constructing a tripartite graph, utilizing multi-perspective consistency for self-consistency re-ranking to improve code generation quality.
Enhancing Multilingual Capabilities of Large Language Models through Self-Distillation from Resource-Rich Languages
Yuanchi Zhang (Tsinghua University), Yang Liu (Tsinghua University)
Knowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Enhancing the multilingual capabilities of large language models in low-resource languages through self-distillation from resource-rich languages; generating answers in the original resource-rich language (e.g., English) by the model itself, then combining these answers with machine translation, code-switching, and a small amount of parallel corpora to construct a diverse multilingual training set, further fine-tuning models such as LLaMA‑2 and SeaLLM.
Enhancing Noise Robustness of Retrieval-Augmented Language Models with Adaptive Adversarial Training
Feiteng Fang (University of Science and Technology of China), Ruifeng Xu (Shenzhen University)
RetrievalAdversarial AttackTransformerTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Studied the robustness of retrieval-augmented language models against retrieval noise and proposed an adaptive adversarial training method called RAAT.
Enhancing Numerical Reasoning with the Guidance of Reliable Reasoning Processes
Dingzirui Wang (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
Explainability and InterpretabilityTransformerSupervised Fine-TuningTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the ENCORE method, which decomposes the answer formula into operators and operands to generate a reliable reasoning process for training small-scale models;
Enhancing Reinforcement Learning with Label-Sensitive Reward for Natural Language Understanding
Kuo Liao (Tencent), Chengguo Yin (Tencent)
Reinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes an improved reinforcement learning framework, RLLR, which enhances the performance of large language models in natural language understanding tasks by utilizing label-sensitive rewards;
ERA-CoT: Improving Chain-of-Thought through Entity Relationship Analysis
Yanming Liu (Zhejiang University), Xuhong Zhang (Zhejiang University)
Large Language ModelTextChain-of-Thought
🎯 What it does: Improve LLM chain-of-thought (CoT) in multi-entity complex scenarios, proposing the ERA-CoT framework to enhance reasoning capabilities through entity relationship analysis.
Error-preserving Automatic Speech Recognition of Young English Learners’ Language
Janick Michot (Zurich University of Applied Sciences), Mark Cieliebak (Zurich University of Applied Sciences)
RecognitionData-Centric LearningTransformerSupervised Fine-TuningContrastive LearningAudio
🎯 What it does: Built an error-preserving automatic speech recognition system for elementary school English learners, collecting and manually annotating 85 hours of child spoken language data containing 45,004 sentences with error annotations.
ESCoT: Towards Interpretable Emotional Support Dialogue Systems
Tenggan Zhang (Renmin University of China), Qin Jin (Renmin University of China)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextSequentialChain-of-Thought
🎯 What it does: Proposed an explainable generation framework named ESCoT for emotion support dialogue systems, and constructed the first emotion support dialogue dataset with chain-of-thought reasoning, ESD-CoT (1,708 dialogues), based on this framework.
Estimating Agreement by Chance for Sequence Annotation
Diya Li (Freenome Holdings, Inc), Chunxiao Zhou (National Institutes of Health)
RecognitionTextBenchmark
🎯 What it does: This paper proposes a chance agreement estimation method for sequence annotation tasks based on a random annotation model.
Estimating the Level of Dialectness Predicts Inter-annotator Agreement in Multi-dialect Arabic Datasets
Amr Keleg (University of Edinburgh), Sharon Goldwater (University of Edinburgh)
ClassificationData-Centric LearningText
🎯 What it does: The study uses the Arabic Dialect Measure (ALDi) to evaluate annotator consistency in a multilingual Arabic dataset and proposes routing high ALDi samples to native speakers of the corresponding dialect to improve annotation quality.
EUROPA: A Legal Multilingual Keyphrase Generation Dataset
Olivier Salaün (Universite De Montreal), Philippe Langlais (Universite De Montreal)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed and publicly released the multilingual (24 languages) keyphrase generation dataset EUROPA for EU court judgments, covering cases from 1957 to 2023 with approximately 280,000 instances;
Evaluating Dynamic Topic Models
Charu Karakkaparambil James (Rheinland-Pfalz Technische Universität Kaiserslautern-Landau), Sophie Fellenz (Rheinland-Pfalz Technische Universität Kaiserslautern-Landau)
Explainability and InterpretabilityRepresentation LearningTextTime Series
🎯 What it does: Proposed and validated temporal dimension evaluation metrics TTQ (Temporal Topic Quality) and DTQ (Dynamic Topic Quality) for dynamic topic models, and verified their effectiveness through manual evaluation.
Evaluating Intention Detection Capability of Large Language Models in Persuasive Dialogues
Hiromasa Sakurai (University of Tokyo), Yusuke Miyao (University of Tokyo)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Investigated the ability of large language models (LLMs) to detect speaker intent in multi-turn persuasive dialogues, constructed a multiple-choice evaluation dataset based on face theory, and assessed the performance of models such as GPT-4, ChatGPT, Llama-2-Chat, and Vicuna.
Evaluating Very Long-Term Conversational Memory of LLM Agents
Adyasha Maharana (University of North Carolina Chapel Hill), Yuwei Fang (Snap Inc)
GenerationData SynthesisData-Centric LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodalityTime SeriesSequentialRetrieval-Augmented Generation
🎯 What it does: Constructed a human-AI hybrid process to generate and validate the LOCOMO dataset, comprising 10 dialogues with approximately 600 turns each and 32 sessions, covering multi-modal (image) and temporal event graphs;
Event-Radar: Event-driven Multi-View Learning for Multimodal Fake News Detection
Zihan Ma (Xi'an Jiaotong University), Xiang Zhao (Nation University of Defence Technology)
ClassificationGraph Neural NetworkTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose an event-driven multi-perspective learning framework called Event-Radar for multi-modal fake news detection;
Every Answer Matters: Evaluating Commonsense with Probabilistic Measures
Qi Cheng (University of Pittsburgh), Xiang Lorraine Li (University of Pittsburgh)
Large Language ModelTextBenchmark
🎯 What it does: Constructed the Commonsense Frame Completion (CFC) task to evaluate models' ability to generate implicit common sense completions.
EWEK-QA : Enhanced Web and Efficient Knowledge Graph Retrieval for Citation-based Question Answering Systems
Mohammad Dehghan (University of Waterloo), Mehdi Rezagholizadeh (Huawei Noah's Ark Lab)
RetrievalComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a reference-based QA system EWEK-QA that simultaneously leverages web text and knowledge graphs, enhancing answer quality and efficiency through adaptive web retrieval and efficient KG triplet extraction.
Ex3: Automatic Novel Writing by Extracting, Excelsior and Expanding
Huang Lei (Institute of Computing Technology, Chinese Academy of Sciences), Tianshi Chen (Cambricon Technologies)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed and implemented the Ex 3 framework for automatically generating long novels: first extract hierarchical structural information from original novels, construct an instruction-based dataset and fine-tune large language models (LLMs), then generate variable-length, coherent novels using tree-based depth-first expansion and entity extraction.
Examining the robustness of LLM evaluation to the distributional assumptions of benchmarks
Charlotte Siska (Microsoft), James Bono (Microsoft)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Evaluate the robustness of large language models (LLMs) across different benchmarks, investigate the impact of benchmark distribution assumptions on model rankings, and propose an evaluation method based on clustering and random weighting
EXAMS-V: A Multi-Discipline Multilingual Multimodal Exam Benchmark for Evaluating Vision Language Models
Rocktim Das (Mohamed bin Zayed University of Artificial Intelligence), Preslav Nakov (Mohamed bin Zayed University of Artificial Intelligence)
RecognitionTransformerVision Language ModelImageTextMultimodalityTabularBenchmark
🎯 What it does: Proposed and constructed EXAMS-V, a college-level multi-subject examination benchmark containing 20,932 questions, spanning 20 subjects, 11 languages, and incorporating multimodal information such as images, tables, and charts, designed to evaluate the comprehensive reasoning capabilities of vision-language models.
Expedited Training of Visual Conditioned Language Generation via Redundancy Reduction
Yiren Jian (Dartmouth College), Hongxia Yang (ByteDance Inc)
GenerationTransformerLarge Language ModelVision Language ModelImageVideoText
🎯 What it does: Proposed the EVLGen framework, which employs a frozen ViT and LLM, and achieves single-stage pre-training through vision-language alignment using TomeFormer;
Experiential Co-Learning of Software-Developing Agents
Chen Qian (Tsinghua University), Maosong Sun (Tsinghua University)
AI Code AssistantTransformerLarge Language ModelAgentic AITextGraphRetrieval-Augmented Generation
🎯 What it does: Proposes an Experiential Co-Learning framework enabling software development LLM agents to collect and leverage experiences from past tasks through a two-phase collaboration (instructor + assistant), achieving higher quality and efficiency in code generation when facing new tasks.
Explainability and Hate Speech: Structured Explanations Make Social Media Moderators Faster
Agostina Calabrese (University of Edinburgh), Francesco Barbieri (Snap Inc)
Explainability and InterpretabilityText
🎯 What it does: This study evaluates the impact of different types of explanations (no explanation, generic explanation, structured explanation) on the decision speed of social media content moderators, with a comparison conducted among 25 professional moderators under three experimental settings.
Explanation-aware Soft Ensemble Empowers Large Language Model In-context Learning
Yue Yu (Georgia Institute of Technology), Michael Bendersky (Google)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Propose a framework called EASE that leverages free-text explanations generated by LLMs to enhance context learning under few-shot scenarios
Explicating the Implicit: Argument Detection Beyond Sentence Boundaries
Paul Roit (Bar Ilan University), Ido Dagan (Bar Ilan University)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Reformulate the cross-sentence semantic argument detection problem as a text entailment task: first extract local arguments using a sentence-level QA-SRL parser, construct concise hypothesis sentences, and then use an NLI model to determine whether the hypothesis is entailed by the document, thereby identifying implicit arguments of predicates that appear across sentences in the document.
Exploiting Intrinsic Multilateral Logical Rules for Weakly Supervised Natural Language Video Localization
Zhe Xu (Xidian University), Cheng Deng (Xidian University)
Object DetectionGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningVideoText
🎯 What it does: Propose a framework for weakly supervised natural language video localization, which utilizes relation-guided prompts to generate an intrinsic temporal relation graph (ITRG), and designs multi-party temporal logic rules based on this to constrain model training, thereby improving the logical consistency and accuracy of localization.
Explore Spurious Correlations at the Concept Level in Language Models for Text Classification
Yuhang Zhou (University of Maryland), Furong Huang (University of Maryland)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper uses ChatGPT to annotate concepts in text classification data, quantifies and evaluates spurious correlations at the concept level in language models during fine-tuning and in-context learning (ICL), and proposes to alleviate this issue by leveraging ChatGPT-generated adversarial counterfactual samples for data rebalancing.
Exploring Alignment in Shared Cross-lingual Spaces
Basel Mousi (Qatar Computing Research Institute, HBKU Research Complex), Ahmed Abdelali (Qatar Computing Research Institute, HBKU Research Complex)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Cluster analysis of the potential concept space in multilingual Transformer models, proposing two metrics, CALIGN and COLAP, to quantify cross-lingual concept alignment and overlap.
Exploring Chain-of-Thought for Multi-modal Metaphor Detection
Yanzhi Xu (Soochow University), Zhongqing Wang (Soochow University)
ClassificationTransformerMultimodalityChain-of-Thought
🎯 What it does: This paper proposes a framework named C4MMD, which leverages a multi-modal large language model (MLLM) through the Chain-of-Thought (CoT) prompting approach to progressively acquire image, text, and fused information. By employing a modal fusion architecture and auxiliary tasks, the framework compresses this knowledge into features, ultimately enhancing the performance of small models in multi-modal metaphor detection.
Exploring Collaboration Mechanisms for LLM Agents: A Social Psychology View
Jintian Zhang (Zhejiang University), Shumin Deng (National University Of Singapore)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Experimentally and theoretically analyze the cooperative mechanisms in multi-agent LLM societies, constructing multi-agent societies with diverse personalities and thinking patterns and assessing their impact on task performance.
Exploring Conditional Variational Mechanism to Pinyin Input Method for Addressing One-to-Many Mappings in Low-Resource Scenarios
Bin Sun (Beijing Institute of Technology), Jie Zhou (Tencent Inc)
GenerationData-Centric LearningTransformerAuto EncoderTextBenchmark
🎯 What it does: Propose a conditional variational mechanism-based pinyin input method model (CV-IME), which enhances candidate diversity and accuracy in low-resource environments by combining continuous and discrete latent variables.
Exploring Hybrid Question Answering via Program-based Prompting
Qi Shi (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
Large Language ModelPrompt EngineeringMultimodality
🎯 What it does: Proposes the HPROPRO framework, which achieves hybrid question answering without retrieval or modal conversion by utilizing programmatic prompts;
Exploring Memorization in Fine-tuned Language Models
Shenglai Zeng (Michigan State University), Dawei Yin (Michigan State University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper conducts large-scale experiments on fine-tuned language models for different downstream tasks, utilizing an automatic plagiarism detection pipeline to evaluate their memory phenomena during the fine-tuning stage, and reveals the relationship between inter-task memory differences and model scale.
Exploring Precision and Recall to assess the quality and diversity of LLMs
Florian Le Bronnec (Université Paris-Dauphine), Alexandre Allauzen (Université Paris-Dauphine)
GenerationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes a distributed evaluation framework based on Precision and Recall to measure the quality and diversity of text generated by large language models (LLMs), enabling comprehensive assessment of LLMs without relying on aligned corpora.
Exploring the Potential of Large Language Models in Computational Argumentation
Guizhen Chen (DAMO Academy, Alibaba Group), Lidong Bing (DAMO Academy, Alibaba Group)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper systematically evaluates the zero-shot and few-shot performance of large language models in computational argumentation tasks (argument mining and argument generation), and proposes a new conversational counter-argument generation benchmark;
Extreme Miscalibration and the Illusion of Adversarial Robustness
Vyas Raina (University of Cambridge), George Karypis (Amazon)
ClassificationAdversarial AttackTransformerText
🎯 What it does: This paper investigates the 'Illusion of Robustness (IOR)' phenomenon in NLP model adversarial robustness, revealing that extreme imbalance calibration can obscure gradients, causing adversarial attack searches to fail and leading to the misconception that models are more robust.
EZ-STANCE: A Large Dataset for English Zero-Shot Stance Detection
Chenye Zhao (University of Illinois Chicago), Cornelia Caragea (University of Illinois Chicago)
ClassificationTransformerSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a large-scale English zero-shot stance detection dataset called EZ-STANCE, and converted the stance detection task into a natural language inference task.
F-Eval: Asssessing Fundamental Abilities with Refined Evaluation Methods
Yu Sun (Shanghai AI Laboratory), Dahua Lin (Shanghai AI Laboratory)
Large Language ModelTextBenchmark
🎯 What it does: Propose the F-Eval benchmark, a bilingual evaluation of LLMs' fundamental capabilities in three dimensions: expression, common sense, and logic, containing 15 sub-datasets and multiple task formats.
FactPICO: Factuality Evaluation for Plain Language Summarization of Medical Evidence
Sebastian Joseph (University of Texas at Austin), Junyi Jessy Li (University of Texas at Austin)
TransformerLarge Language ModelTextBiomedical DataBenchmark
🎯 What it does: Proposes the FACTPICO benchmark, using expert evaluation to measure the factualness of medical RCT abstracts generated by large models, with fine-grained attention to PICO elements, evidence reasoning, and new explanations.
Factual Confidence of LLMs: on Reliability and Robustness of Current Estimators
Matéo Mahaut (Universitat Pompeu Fabra), Lluis Marquez
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes and systematically evaluates multiple methods for estimating the factual confidence of large language models, constructing a unified experimental framework and conducting comparisons across various models and datasets.