ACL 2024 Papers — Page 7
Annual Meeting of the Association for Computational Linguistics · 940 papers
OceanGPT: A Large Language Model for Ocean Science Tasks
Zhen Bi (Zhejiang University), Huajun Chen (Zhejiang University)
Robotic IntelligenceTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: We constructed the first pre-trained language model for ocean science, OceanGPT, and generated ocean domain instruction data through the multi-agent collaborative DOINSTRUCT framework, followed by creating the OCEANBENCH benchmark evaluation.
OLIVE: Object Level In-Context Visual Embeddings
Timothy Ossowski (University of Wisconsin), Junjie Hu (University of Wisconsin)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose the OLIVE method, which directly injects object-level visual embeddings into large language models, supporting object-level reasoning, zero-shot transfer, and multi-image context prompting.
OLMo: Accelerating the Science of Language Models
Dirk Groeneveld (Allen Institute for Artificial Intelligence), Hannaneh Hajishirzi (Allen Institute for Artificial Intelligence)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Released OLMo, a fully open-source language model with 1B and 7B parameter scales, providing a complete reproducible workflow from data construction, training, evaluation to intermediate checkpoints;
OlympiadBench: A Challenging Benchmark for Promoting AGI with Olympiad-Level Bilingual Multimodal Scientific Problems
Chaoqun He (Tsinghua University), Maosong Sun (Tsinghua University)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Proposed the OlympiadBench benchmark, collecting and organizing 8,476 bilingual Chinese-English competition-level math and physics problems with images, providing expert-level step-by-step solution annotations, and conducting zero-shot evaluations of existing large language models and multimodal models.
On Context Utilization in Summarization with Large Language Models
Mathieu Ravaut (Nanyang Technological University), Shafiq Joty (Nanyang Technological University)
GenerationLarge Language ModelTextBenchmark
🎯 What it does: This paper systematically investigates how large language models (LLMs) utilize input context in abstractive summarization tasks, revealing the 'position bias (middle hell)' phenomenon and proposing the MiddleSum dataset along with two reasoning methods (hierarchical reasoning and incremental reasoning) to alleviate this issue.
On Efficient and Statistical Quality Estimation for Data Annotation
Jan-Christoph Klie (UKP Lab, TU Darmstadt), Rahul Nair (Apple)
ClassificationRecognitionData-Centric LearningText
🎯 What it does: Propose and evaluate a statistical quality control method based on confidence intervals and acceptance sampling for error rate estimation in the data annotation process.
On Measuring Faithfulness or Self-consistency of Natural Language Explanations
Letitia Parcalabescu (Heidelberg University), Anette Frank (Heidelberg University)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper investigates the trustworthiness of natural language explanations generated by large language models, proposing a new metric called CC-SHAP and building a unified testing platform named CCB.
On the Hallucination in Simultaneous Machine Translation
Meizhi Zhong (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
GenerationTransformerText
🎯 What it does: Systematically analyze hallucinations in synchronous machine translation, investigating their frequency distribution, prediction distribution, and usage of target-side context; and attempt to alleviate hallucinations through a scheduled sampling method that adds noise to the target-side context.
On the Impact of Calibration Data in Post-training Quantization and Pruning
Miles Williams (University of Sheffield), Nikolaos Aletras (University of Sheffield)
CompressionData-Centric LearningTransformerTextBenchmark
🎯 What it does: Systematically investigate the impact of calibration data on the post-training quantization and pruning compression effects of large language models.
On the Multi-turn Instruction Following for Conversational Web Agents
Yang Deng (Singapore Management University), Tat-Seng Chua (National University of Singapore)
TransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Construct the MT-Mind2Web dataset and propose the Self-MAP framework to achieve multi-round instruction interaction for web navigation
On the Representational Capacity of Neural Language Models with Chain-of-Thought Reasoning
Franz Nowak (ETH Zurich), Ryan Cotterell (ETH Zurich)
Representation LearningRecurrent Neural NetworkTransformerChain-of-Thought
🎯 What it does: This paper formalizes chain-of-thought (CoT) within a probabilistic framework, elucidating the representational capabilities of CoT language models (LMs), and proves that CoT-trained RNNs and Transformers are weakly equivalent to probabilistic Turing machines (PTMs) at different precision levels, thereby demonstrating their theoretical Turing completeness.
On the Role of Long-tail Knowledge in Retrieval Augmented Large Language Models
Dongyang Li (East China Normal University), Jun Huang (Alibaba Group)
RetrievalComputational EfficiencyTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes a long-tail knowledge detection method called GECE and improves the RAG process by retrieving documents only when the query belongs to long-tail knowledge.
On the Semantic Latent Space of Diffusion-Based Text-To-Speech Models
Miri Varshavsky-Hassid (Verily AI), Ehud Rivlin (Verily AI)
GenerationRepresentation LearningConvolutional Neural NetworkDiffusion modelAudio
🎯 What it does: Investigate the latent space of diffusion models in text-to-speech systems, revealing that it contains semantic information, and achieve voice attribute editing in frozen models through supervised and unsupervised methods.
One Prompt To Rule Them All: LLMs for Opinion Summary Evaluation
Tejpalsingh Siledar (IIT Bombay), Nikesh Garera (Flipkart)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Created the SUMMEVAL-OP evaluation dataset and proposed OP-I-PROMPT and OP-PROMPTS for multi-dimensional opinion summary evaluation.
One-Shot Learning as Instruction Data Prospector for Large Language Models
Yunshui Li (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Yongbin Li (Alibaba Group)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the NUGGETS method, which filters the most valuable instruction set for subsequent instruction tuning without requiring additional annotations by evaluating a golden score for each instruction example through a single round of in-context learning from large language models.
Open Grounded Planning: Challenges and Benchmark Construction
Shiguang Guo (Chinese Information Processing Laboratory), Le Sun (Chinese Information Processing Laboratory)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the open grounded planning task based on an action set and constructed a corresponding benchmark
Open Ko-LLM Leaderboard: Evaluating Large Language Models in Korean with Ko-H5 Benchmark
Chanjun Park (Upstage AI), Hwalsuk Lee (Upstage AI)
Large Language ModelTextBenchmark
🎯 What it does: This study constructs the Open Ko-LLM Leaderboard and Ko-H5 Benchmark for Korean, employing a private test set and aligning with the English Open LLM Leaderboard to systematically evaluate the performance of Korean LLMs; conducts in-depth analyses of relevance, temporal evolution, and task saturation of the benchmark; proposes practical guidelines for expanding the benchmark based on score saturation; collects and statistics quality issues of submitted models; and encourages the community to collaboratively improve the leaderboard.
Open-Set Semi-Supervised Text Classification via Adversarial Disagreement Maximization
Junfan Chen (Beihang University), Chunming Hu (Beihang University)
ClassificationTransformerText
🎯 What it does: Proposes an open-set semi-supervised text classification method ADM based on adversarial discrepancy maximization, aiming to directly amplify the discrepancy between discriminative and anomaly detection measurements to enhance performance.
OpenToM: A Comprehensive Benchmark for Evaluating Theory-of-Mind Reasoning Capabilities of Large Language Models
Hainiu Xu (Kings College London), Yulan He (Kings College London)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper constructs the OpenToM benchmark, using long natural stories and diverse questions to evaluate LLMs' neurotheoretical mind (N-ToM) capabilities.
OPEx: A Component-Wise Analysis of LLM-Centric Agents in Embodied Instruction Following
Haochen Shi (Université de Montréal & Mila), Bang Liu (Université de Montréal & Mila)
Robotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelImageTextChain-of-Thought
🎯 What it does: Proposed a multi-component framework called OPEx for large language models, decomposing the Embodied Instruction Following task into three modules: Observer, Planner, and Executor. It introduces multi-agent LLM communication and prior knowledge integration, addressing bottlenecks such as the coupling between planning and execution, and perceptual errors in traditional methods.
Order-Agnostic Data Augmentation for Few-Shot Named Entity Recognition
Huiming Wang (Singapore University of Technology and Design), Lidong Bing (DAMO Academy, Alibaba Group)
RecognitionTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the Order-Agnostic Data Augmentation (OADA) method to enhance the performance of few-shot named entity recognition (NER) models.
OWSM-CTC: An Open Encoder-Only Speech Foundation Model for Speech Recognition, Translation, and Language Identification
Yifan Peng (Carnegie Mellon University), Shinji Watanabe (Carnegie Mellon University)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringAudio
🎯 What it does: Proposed a single-encoder speech foundation model called OWSM-CTC, capable of simultaneously performing multilingual automatic speech recognition (ASR), speech translation (ST), and language identification (LID) tasks.
PAGED: A Benchmark for Procedural Graphs Extraction from Documents
Weihong Du (Sichuan University), Wenqiang Lei (Sichuan University)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmark
🎯 What it does: Propose the PAGED benchmark, construct a large-scale high-quality program graph-document pair dataset, systematically evaluate existing methods, and explore the potential of LLMs in program graph extraction.
PairCFR: Enhancing Model Training on Paired Counterfactually Augmented Data through Contrastive Learning
Xiaoqi Qiu (Shenzhen University), Chunyan Miao (Nanyang Technological University)
Domain AdaptationRepresentation LearningData-Centric LearningTransformerContrastive LearningText
🎯 What it does: Propose the PairCFR framework, which combines contrastive learning and cross-entropy during training of Counterfactually Augmented Data (CAD), pairing original samples with their adversarial counterparts to enhance the model's learning of global features.
Parallel Structures in Pre-training Data Yield In-Context Learning
Yanda Chen (Columbia University), He He (New York University)
Representation LearningMeta LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigate how parallel structures (phrase pairs with the same template) in pre-trained text enhance language models' in-context learning (ICL) ability, and verify their importance through ablation experiments.
Paraphrasing in Affirmative Terms Improves Negation Understanding
MohammadHossein Rezaei (University of Arizona), Eduardo Blanco (University of Arizona)
Data-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Study how to use affirmative explanations (i.e., paraphrases without negation words) to enhance the model's understanding of negated sentences, and automatically generate these affirmative explanations; concatenate the generated affirmative explanations with the original sentence and input them into the model for training and inference.
Pareto Optimal Learning for Estimating Large Language Model Errors
Theodore Zhao (Microsoft), Hoifung Poon (Microsoft)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: This paper proposes the POLAR framework based on Pareto optimal learning, aimed at quantitatively estimating the error probability of large language model (LLM) outputs;
Parrot: Enhancing Multi-Turn Instruction Following for Large Language Models
Yuchong Sun (Renmin University of China), Kun Gai (Renmin University of China)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Enhance the instruction-following capability of large language models in multi-round interactions through the Parrot framework, with training and optimization focused on human-like multi-turn dialogues;
PathReasoner: Modeling Reasoning Path with Equivalent Extension for Logical Question Answering
Fangzhi Xu (Xi'an Jiaotong University), Jun Liu (Xi'an Jiaotong University)
Graph Neural NetworkTransformerDiffusion modelText
🎯 What it does: Proposes the PathReasoner framework, which converts text sentences into atoms and models reasoning paths to address insufficient logical consistency and structural awareness in logical reasoning tasks.
PCAD: Towards ASR-Robust Spoken Language Understanding via Prototype Calibration and Asymmetric Decoupling
Xianwei Zhuang (Peking University), Yuexian Zou (Peking University)
Knowledge DistillationRepresentation LearningContrastive LearningTextAudio
🎯 What it does: Enhancing the robustness of speech understanding models under noisy speech recognition conditions through prototype calibration and asymmetric decoupled training.
Peacock: A Family of Arabic Multimodal Large Language Models and Benchmarks
Fakhraddin Alwajih (University of British Columbia), Muhammad Abdul-Mageed (University of British Columbia)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Built a series of multilingual large language models for Arabic called Peacock and proposed a new cultural evaluation benchmark named Henna.
Persuading across Diverse Domains: a Dataset and Persuasion Large Language Model
Chuhao Jin (Renmin University of China), Huan Chen (Meituan)
GenerationData SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Constructed a multi-domain persuasion dialogue dataset named DailyPersuasion spanning 35 daily domains, and proposed PersuGPT, a persuasion LLM integrating intent-strategy reasoning and multi-turn simulation preference optimization to enhance cross-domain persuasion capabilities.
Picturing Ambiguity: A Visual Twist on the Winograd Schema Challenge
Brendan Park (Brock University), Ali Emami (Brock University)
GenerationData SynthesisExplainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Propose the WINOVIS dataset and evaluation framework for testing the pronoun disambiguation capabilities of text-to-image models in multimodal contexts
PITA: Prompting Task Interaction for Argumentation Mining
Yang Sun (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
ClassificationRecognitionGraph Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the PITA model, which leverages generative prompt tuning through dynamic prompt templates and task interaction graphs to jointly decode three subtasks: argument component type identification, relation recognition, and relation type classification.
PixT3: Pixel-based Table-To-Text Generation
Iñigo Alonso (University of Basque Country), Mirella Lapata (University of Edinburgh)
GenerationTransformerVision Language ModelImageTextTabular
🎯 What it does: Propose PixT3, a pixel-level table-to-text generation model that treats tables as images and is pre-trained using a self-supervised structural learning curve, applicable to open-ended, controlled, and loosely controlled generation scenarios.
Planning Like Human: A Dual-process Framework for Dialogue Planning
Tao He (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Designed and implemented a dialogue planning framework based on dual-process theory (DPDP), combining a fast policy language model with a slow Monte Carlo Tree Search (MCTS), and proposed a non-parametric switching gate and two-phase training (offline reinforcement learning + MCTS-guided self-play).
PlatoLM: Teaching LLMs in Multi-Round Dialogue via a User Simulator
Chuyi Kong (Chinese University of Hong Kong Shenzhen), Benyou Wang (Chinese University of Hong Kong Shenzhen)
GenerationData SynthesisSupervised Fine-TuningText
🎯 What it does: Proposed a trainable user simulator called Socratic, which learns from real human questions to generate more human-like multi-turn dialogue data named SocraticChat. Subsequently, a response model PlatoLM was fine-tuned on this data.
PLUG: Leveraging Pivot Language in Cross-Lingual Instruction Tuning
Zhihan Zhang (University of Notre Dame), Francesco Barbieri (University of Notre Dame)
GenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes the PLUG (Pivot Language Guided Generation) method, which uses high-resource languages (e.g., English) as intermediaries. It first generates instructions and responses in the pivot language during a single inference step, then generates responses in the target low-resource language, thereby enhancing the instruction-following capability of large language models (LLMs) in low-resource languages, and constructs a multilingual evaluation benchmark called X-AlpacaEval.
PokeMQA: Programmable knowledge editing for Multi-hop Question Answering
Hengrui Gu (Jilin University), Xin Wang (Jilin University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the PokeMQA framework, which improves multi-hop question answering models using a programmable knowledge editing method, decoupling problem decomposition from knowledge conflict detection, and enhancing reasoning quality through external memory and knowledge prompting.
PolCLIP: A Unified Image-Text Word Sense Disambiguation Model via Generating Multimodal Complementary Representations
Qihao Yang (South China Normal University), Tianyong Hao (South China Normal University)
Representation LearningData-Centric LearningTransformerVision Language ModelDiffusion modelContrastive LearningMultimodality
🎯 What it does: Proposed the PolCLIP model, achieving unified processing of two subtasks of word sense disambiguation: text and image.
Political Compass or Spinning Arrow? Towards More Meaningful Evaluations for Values and Opinions in Large Language Models
Paul Röttger (Bocconi University), Dirk Hovy (Bocconi University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper evaluates the values and opinions of large language models (LLMs) using the Political Compass Test (PCT), systematically comparing differences in model responses under various evaluation settings, including restricted multiple-choice, restricted multiple-choice with forced prompts, prompt tuning, and open-ended answers, and exploring the impact of prompting methods, forced approaches, and robustness on evaluation outcomes.
POMP: Probability-driven Meta-graph Prompter for LLMs in Low-resource Unsupervised Neural Machine Translation
Shilong Pan (National University of Defense Technology), Dongsheng Li (National University of Defense Technology)
GenerationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: This paper proposes an unsupervised low-resource language translation method called POMP, which constructs language-specific meta-graphs to dynamically sample multiple translation paths and uses these paths to organize auxiliary languages as prompts to drive large language models (LLMs) for translation; meanwhile, a probabilistic reverse graph evolution mechanism is introduced to continuously update the sampling probabilities of auxiliary languages.
Pouring Your Heart Out: Investigating the Role of Figurative Language in Online Expressions of Empathy
Gyeongeun Lee (University of Illinois at Chicago), Natalie Parde (University of Illinois at Chicago)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper constructs the AcnEmpathize dataset for acne forum empathetic expressions and investigates the role of metaphorical language, such as metaphors, idioms, and hyperbole, in online empathy detection.
Predicting Text Preference Via Structured Comparative Reasoning
Jing Nathan Yan (Cornell University), Michael Bendersky (Google)
Recommendation SystemTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a structured comparison reasoning model SC² to enhance the accuracy of text preference prediction
PreFLMR: Scaling Up Fine-Grained Late-Interaction Multi-modal Retrievers
Weizhe Lin (University of Cambridge), Bill Byrne (University of Cambridge)
RetrievalTransformerVision Language ModelMultimodalityBenchmark
🎯 What it does: Developed PreFLMR, a fine-grained late-interaction based multi-modal retriever, and constructed M2KR as a unified retrieval benchmark, improving retrieval and answering performance in KB-VQA.
PRewrite: Prompt Rewriting with Reinforcement Learning
Weize Kong (Google DeepMind), Michael Bendersky (Google DeepMind)
ClassificationGenerationOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a method that uses a reinforcement learning (RL)-trained large language model (LLM) as a prompt rewriter to automatically rewrite initial prompts, thereby enhancing the performance of LLMs on downstream tasks.
Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinement
Wenda Xu (University of California, Santa Barbara), William Wang (Carnegie Mellon University)
TransformerLarge Language ModelText
🎯 What it does: This paper quantifies self-bias in large language models during the self-refinement process and conducts systematic evaluations on tasks such as translation, constrained text generation, and mathematical reasoning.
PrivLM-Bench: A Multi-level Privacy Evaluation Benchmark for Language Models
Haoran Li (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
Safty and PrivacyTransformerSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Propose the PrivLM-Bench benchmark, which uses a unified workflow to evaluate multi-layer privacy leakage in language models, including both training and inference stages, and conducts empirical evaluation combined with various privacy attacks.
Probing Language Models for Pre-training Data Detection
Zhenhua Liu (Soochow University), Wenliang Chen (Soochow University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes an internal activation analysis method based on detection technology to determine whether a given text is included in the pre-training data of large language models; meanwhile, a new academic-level dataset named ArxivMIA is constructed as a more challenging benchmark for pre-training data detection.
Probing the Multi-turn Planning Capabilities of LLMs via 20 Question Games
Yizhe Zhang (Apple), Navdeep Jaitly (Apple)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Designed an 'Entity Reasoning Game' (20 questions) as a benchmark to evaluate the multi-round planning and reasoning capabilities of large language models (LLMs), and systematically assessed the performance of various LLMs on this task.
Progressively Modality Freezing for Multi-Modal Entity Alignment
Yani Huang (Beihang University), Jaein Kim (Beihang University)
Representation LearningContrastive LearningMultimodality
🎯 What it does: Propose the Progressive Modality Freezing (PMF) method for multi-modal entity alignment, progressively freezing irrelevant modal features and fusing useful features.
PRoLoRA: Partial Rotation Empowers More Parameter-Efficient LoRA
Sheng Wang (University of Hong Kong), Chuan Wu (University of Hong Kong)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes PRoLoRA, a parameter-efficient fine-tuning method that achieves internal layer sharing based on LoRA.
Prompt Expansion for Adaptive Text-to-Image Generation
Siddhartha Datta (Google DeepMind), Peter Anderson (Google DeepMind)
GenerationTransformerSupervised Fine-TuningPrompt EngineeringDiffusion modelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the Prompt Expansion framework, which generates diverse and high-quality images using expanded text prompts, significantly reducing the need for users to refine prompts.
Prompt Optimization via Adversarial In-Context Learning
Xuan Long Do (National University of Singapore), Junxian He (Hong Kong University of Science and Technology)
OptimizationLarge Language ModelPrompt EngineeringGenerative Adversarial NetworkText
🎯 What it does: Propose an adversarial learning-based prompt optimization framework, adv-ICL, which iteratively optimizes on LLMs using a generator, discriminator, and prompt modifier;
Prompt Refinement with Image Pivot for Text-to-Image Generation
Jingtao Zhan (Tsinghua University), Tao Mei (HiDream.ai)
GenerationData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Proposed and implemented a prompt refinement framework called PRIP based on image latent representations, which can automatically convert user natural language prompts into technically refined prompts suitable for text-image generation systems.
Prompted Aspect Key Point Analysis for Quantitative Review Summarization
An Quang Tang (RMIT University), Erik Cambria (Nanyang Technological University)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the Prompted Aspect Key Point Analysis (PAKPA) framework, which utilizes large language models to perform aspect-based sentiment analysis on reviews and generates and quantifies key points.
ProtLLM: An Interleaved Protein-Language LLM with Protein-as-Word Pre-Training
Le Zhuo (Beihang University), Wentao Zhang (Peking University)
Protein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBiomedical DataBenchmark
🎯 What it does: Designed and trained a cross-modal large language model called PROTLLM, capable of processing natural language and interleaved protein inputs of arbitrary quantity, supporting protein-centric and protein-language tasks.
Prototypical Reward Network for Data-Efficient RLHF
Jinghan Zhang (Portland State University), Kunpeng Liu (Georgia Institute of Technology)
Reinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningText
🎯 What it does: Propose Proto-RM, which leverages prototype networks to enhance RLHF reward models, maintaining efficient learning and improving LLM generation quality when human feedback samples are scarce.
ProtT3: Protein-to-Text Generation for Text-based Protein Understanding
Zhiyuan Liu (National University of Singapore), Tat-Seng Chua (National University of Singapore)
GenerationRetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose the ProtT3 framework, which integrates protein language models and text language models through a cross-modal projector to accomplish protein sequence-to-text generation and retrieval tasks.
ProxyQA: An Alternative Framework for Evaluating Long-Form Text Generation with Large Language Models
Haochen Tan (City University of Hong Kong), Linqi Song (City University of Hong Kong)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the PROXYQA framework, using proxy questions to evaluate the knowledge coverage and information volume of LLM-generated long texts
PRP-Graph: Pairwise Ranking Prompting to LLMs with Graph Aggregation for Effective Text Re-ranking
Jian Luo (University of Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
RetrievalGraph Neural NetworkLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose a graph aggregation method called PRP-Graph based on an improved contrastive ranking prompt (scoring PRP) to enhance the performance of large language models (LLMs) in zero-shot document re-ranking tasks.
PRP: Propagating Universal Perturbations to Attack Large Language Model Guard-Rails
Neal Mangaokar (University of Michigan), Atul Prakash (University of Michigan)
Adversarial AttackTransformerPrompt EngineeringTextBenchmark
🎯 What it does: Design and verify a two-step attack called PRP targeting Guard-Railed LLMs, which first constructs a generic attack prefix to mislead the Guard model, then injects this prefix into the baseline LLM's response, thereby enabling the leakage of harmful content.
PsychoGAT: A Novel Psychological Measurement Paradigm through Interactive Fiction Games with LLM Agents
Qisen Yang (Tsinghua University), Gao Huang (01.AI)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AITextChain-of-Thought
🎯 What it does: Proposed PsychoGAT, an interactive gamified psychological assessment framework based on large language models (LLM) multi-agent systems, converting traditional self-report scales into immersive narrative games, where players' choices during the game serve as the assessment results.
PsySafe: A Comprehensive Framework for Psychological-based Attack, Defense, and Evaluation of Multi-agent System Safety
Zaibin Zhang (Shanghai Artificial Intelligence Laboratory), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Adversarial AttackLarge Language ModelAgentic AIText
🎯 What it does: This paper proposes the PsySafe framework, which attacks, evaluates, and defends multi-agent systems from a psychological perspective to enhance their security.
Quality-Aware Translation Models: Efficient Generation and Quality Estimation in a Single Model
Christian Tomani (Google), Daniel Cremers (Google)
GenerationComputational EfficiencyTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Integrating quality labels during training of Neural Machine Translation (NMT) models enables the model to self-assess translation quality and leverage this information during decoding to improve translation quality
Quantifying Contamination in Evaluating Code Generation Capabilities of Language Models
Martin Riddell (Yale University), Arman Cohan (Yale University)
Data-Centric LearningAI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated the data pollution problem in code generation models within evaluation benchmarks, and quantified the overlap between the MBPP and HumanEval benchmarks and public pre-training corpora (PILE and STARCODERDATA).
Quantifying Generalizations: Exploring the Divide Between Human and LLMs’ Sensitivity to Quantification
Claudia Collacciani (University of Bologna), Marianna Bolognesi (University of Bologna)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigating the sensitivity and default interpretation capabilities of large language models (LLMs) in understanding generic sentences.
Quantifying the Persona Effect in LLM Simulations
Tiancheng Hu (University of Cambridge), Nigel Collier (University of Cambridge)
TransformerLarge Language ModelPrompt EngineeringTextTabular
🎯 What it does: This study systematically evaluates the role of persona variables (including demographics, attitudes, and behaviors) in the prediction of large language models (LLMs) on subjective NLP tasks, and explores the improvement of model performance and its applicability under zero-shot prompting (persona prompting).
Quantifying Uncertainty in Answers from any Language Model and Enhancing their Trustworthiness
Jiuhai Chen (Cleanlab, University of Maryland), Jonas Mueller (Cleanlab)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the BSDETECTOR method, which enhances the reliability of LLM outputs and the accuracy of automated evaluations by generating diverse answers through multiple calls to a black-box LLM, and estimating answer confidence through observation consistency and self-reflection.
Quantized Side Tuning: Fast and Memory-Efficient Tuning of Quantized Large Language Models
Zhengxin Zhang (Carnegie Mellon University), Zhihao Jia (Carnegie Mellon University)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the Quantization-Side Tuning (QST) framework, which first quantizes the weights of large language models to 4 bits, and then introduces an independent side network for task-specific fine-tuning through low-rank adapters and a gradient-free downsampling module, updating only the side network parameters to achieve significant memory compression and acceleration without significant performance loss.
QueryAgent: A Reliable and Efficient Reasoning Framework with Environmental Feedback based Self-Correction
Xiang Huang (State Key Laboratory for Novel Software Technology, Nanjing University), Yuzhong Qu (State Key Laboratory for Novel Software Technology, Nanjing University)
Computational EfficiencyTransformerLarge Language ModelAgentic AITextGraphChain-of-Thought
🎯 What it does: Propose the QueryAgent framework, which generates KB semantic queries through step-by-step tool execution and incorporates a self-correction process.
RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models
Cheng Niu (NewsBreak), Tong Zhang (University of Illinois Urbana Champaign)
ClassificationAnomaly DetectionTransformerSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper constructs a large-scale, fine-grained word-level RAGTruth hallucination corpus, collecting approximately 18,000 naturally generated RAG responses and performing manual annotations; subsequently, it evaluates and compares multiple hallucination detection methods based on this data, and achieves competitive detection performance by fine-tuning a small model.
RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors
Liam Dugan (University of Pennsylvania), Chris Callison-Burch (University of Pennsylvania)
Anomaly DetectionAdversarial AttackLarge Language ModelTextBenchmark
🎯 What it does: Developed RAID, the largest and most challenging shared benchmark dataset for evaluating machine-generated text detectors, and systematically evaluated 12 open-source, metric-based, and commercial detectors on this dataset.
RAM-EHR: Retrieval Augmentation Meets Clinical Predictions on Electronic Health Records
Ran Xu (Emory University), Carl Yang (Emory University)
RetrievalGraph Neural NetworkTransformerLarge Language ModelBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: Propose RAM-EHR, a framework for EHR prediction that utilizes multi-source knowledge retrieval enhancement and consistency regularization.
RAVEL: Evaluating Interpretability Methods on Disentangling Language Model Representations
Jing Huang (Stanford University), Atticus Geiger (Pr(Ai) R Group 2)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderTextBenchmark
🎯 What it does: Developed and released the RAVEL benchmark to evaluate the disentanglement capability of explanation methods for entity attributes in language models, and conducted experimental comparisons of multiple methods (including MDAS) on Llama2-7B.
RDRec: Rationale Distillation for LLM-based Recommendation
Xinfeng Wang (University of Yamanashi), Fumiyo Fukumoto (University of Yamanashi)
Recommendation SystemKnowledge DistillationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the RDRec model, which uses a large language model to perform chain-of-thought reasoning on user reviews to extract user preferences and product attributes, and then trains a small model using these rationales for top-N and sequential recommendations.
Re-Tuning: Overcoming the Compositionality Limits of Large Language Models with Recursive Tuning
Eric Pasewark (Washington University in St. Louis), Chenguang Wang (Washington University in St. Louis)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Employs recursive fine-tuning (Re-Tuning) to enable large language models to automatically split problems into smaller subproblems during inference, recursively solving them and returning results, significantly enhancing the model's capability for compositional tasks.
Re3: A Holistic Framework and Dataset for Modeling Collaborative Document Revision
Qian Ruan (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
Graph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Proposed the Re3 framework and constructed the Re3-Sci dataset, systematically analyzing the collaborative peer review, revision, and response processes in academic papers.
REANO: Optimising Retrieval-Augmented Reader Models through Knowledge Graph Generation
Jinyuan Fang (University of Glasgow), Craig Macdonald (University of Glasgow)
GenerationRetrievalGraph Neural NetworkTransformerTextGraphRetrieval-Augmented Generation
🎯 What it does: In open-domain question answering tasks, the authors propose REANO, which enhances answer quality by incorporating a knowledge graph generation module into a retrieval-augmented reader.
Reasoning in Conversation: Solving Subjective Tasks through Dialogue Simulation for Large Language Models
Xiaolong Wang (Tsinghua University), Yang Liu (Tsinghua University)
ClassificationRecognitionExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a reasoning method based on dialogue simulation (RiC), which enhances the performance of large language models on subjective tasks through keyword extraction, dialogue generation, and dialogue-enhanced reasoning.
Reasoning in Flux: Enhancing Large Language Models Reasoning through Uncertainty-aware Adaptive Guidance
Zhangyue Yin (Fudan University), Xipeng Qiu (Fudan University)
Explainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the Uncertainty-aware Adaptive Guidance (UAG) method, dynamically monitoring and correcting uncertainty errors during LLM reasoning processes.
RecGPT: Generative Pre-training for Text-based Recommendation
Hoang Ngo (VinAI Research), Dat Quoc Nguyen (VinAI Research)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: This study trains a domain-adapted large language model for text recommendation, named RecGPT-7B, and further obtains RecGPT-7B-Instruct through instruction fine-tuning. It also publicly releases the pre-training and fine-tune datasets to verify its performance on rating prediction and sequential recommendation tasks.
ReConcile: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs
Justin Chen (UNC Chapel Hill), Mohit Bansal (UNC Chapel Hill)
Explainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a multi-model, multi-agent 'roundtable meeting' framework called RECONCILE, which enhances reasoning ability through multi-round discussions among LLMs from different families
Reducing Privacy Risks in Online Self-Disclosures with Language Models
Yao Dou (Georgia Institute of Technology), Wei Xu (Georgia Institute of Technology)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper develops a set of technologies for automatically detecting and abstracting online self-disclosure, helping users reduce privacy risks.
REFINESUMM: Self-Refining MLLM for Generating a Multimodal Summarization Dataset
Vaidehi Patil (UNC Chapel Hill), Markus Dreyer (Amazon AGI)
GenerationData SynthesisLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Proposed and implemented a self-refinement process that leverages a multi-modal large language model (MLLM) to automatically generate image-text summaries. High-quality summaries are then filtered using a multi-dimensional critic, which fine-tunes the MLLM. Finally, a high-quality image-text summary dataset called REFINESUMM was generated and released.
Reflect-RL: Two-Player Online RL Fine-Tuning for LMs
Runlong Zhou (University of Washington), Beibin Li (Microsoft Research)
Supervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes Reflect-RL, a two-player online reinforcement learning fine-tuning framework that first performs supervised fine-tuning (SFT) on language models, then conducts reinforcement learning (RLFT) in interactive environments, enabling models to achieve self-reflection and action in multi-round decision-making tasks.
ReFT: Reasoning with Reinforced Fine-Tuning
Luong Trung, Hang Li (ByteDance Research)
OptimizationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose the ReFT (Reinforced Fine-Tuning) method, which first uses supervised fine-tuning (SFT) as a warm-up, then employs PPO reinforcement learning to sample multiple Chain-of-Thought (CoT) paths on the same training set for self-learning, thereby enhancing the performance of large language models on mathematical reasoning tasks.
RelayAttention for Efficient Large Language Model Serving with Long System Prompts
Lei Zhu (City University of Hong Kong), Rynson Lau (City University of Hong Kong)
Computational EfficiencyTransformerText
🎯 What it does: To address computational bottlenecks caused by long system prompts in large language model services, the RelayAttention scheme is proposed to achieve efficient attention computation without additional training.
Relying on the Unreliable: The Impact of Language Models’ Reluctance to Express Uncertainty
Kaitlyn Zhou (Stanford University), Maarten Sap (Carnegie Mellon University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper analyzes the answers generated by multiple publicly deployed language models on multiple-choice questions, combined with user experiments, to explore the tendency and effectiveness of models in using 'epistemic markers' (uncertainty markers) in their responses.
RepCodec: A Speech Representation Codec for Speech Tokenization
Zhichao Huang (ByteDance), Tom Ko (ByteDance)
CompressionRepresentation LearningConvolutional Neural NetworkAuto EncoderAudio
🎯 What it does: Proposes RepCodec, an end-to-end speech representation compression model that maps speech waveforms to low-bitrate semantic discrete codebooks for speech processing in large language models.
Rephrasing the Web: A Recipe for Compute and Data-Efficient Language Modeling
Pratyush Maini (Carnegie Mellon University), Navdeep Jaitly (Apple)
Data SynthesisComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the WRAP method, which uses instruction-tuned LLMs to rephrase raw web-crawled text into multiple high-quality styles (such as Wikipedia style, Q&A style, etc.), and then mixes the original text with synthetic text for LLM pretraining, significantly improving pretraining efficiency and downstream performance.
Representation Learning with Conditional Information Flow Maximization
Dou Hu (Chinese Academy of Sciences), Songlin Hu (Chinese Academy of Sciences)
ClassificationRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose a conditionally information flow maximization (CIFM) framework based on information theory to learn sufficient and noise-invariant representations that contain both input and target information;
Resisting the Lure of the Skyline: Grounding Practices in Active Learning for Morphological Inflection
Saliha Muradoglu (Australian National University), Mans Hulden (University of British Columbia)
Data-Centric LearningTransformerText
🎯 What it does: The study conducts experiments using active learning for morphological inflection in linguistic corpora, comparing the impact of different distributions of unlabeled sample pools on model performance.
Respond in my Language: Mitigating Language Inconsistency in Response Generation based on Large Language Models
Liang Zhang (Renmin University of China), Furu Wei (Microsoft Research Asia)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper addresses the problem of language inconsistency caused by monolingual instruction fine-tuning in large language models. It proposes using pseudo-inconsistency penalty (PIP) during training to suppress the model's preference for English responses, and enhances multilingual instruction-following consistency during inference by combining prior-enhanced decoding (PED) with the language prior of the base model.
Retaining Key Information under High Compression Ratios: Query-Guided Compressor for LLMs
Zhiwei Cao (Xiamen University), Jinsong Su (Xiamen University)
CompressionTransformerSupervised Fine-TuningText
🎯 What it does: Proposes a Query-Guided Compressor (QGC) that compresses documents in the long context of LLM inputs by query guidance, preserving key information while supporting high compression ratios;
Rethinking Task-Oriented Dialogue Systems: From Complex Modularity to Zero-Shot Autonomous Agent
Heng-Da Xu (Beijing Institute of Technology), Heyan Huang (Beijing Institute of Technology)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: Proposed AutoTOD, a fully zero-shot autonomous task-oriented dialogue agent that directly utilizes instruction-following large language models (LLMs) to complete dialogues, invoke APIs, and generate responses without training any modules.
Rethinking the Bounds of LLM Reasoning: Are Multi-Agent Discussions the Key?
Qineng Wang (Zhejiang University), Yangqiu Song (Zhejiang University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: This paper proposes and systematically evaluates a novel multi-agent discussion framework called CMD, exploring its performance differences with single-agent systems in reasoning tasks;
Rethinking the Multimodal Correlation of Multimodal Sequential Learning via Generalizable Attentional Results Alignment
Tao Jin (Zhejiang University), Zhou Zhao (Zhejiang University)
ClassificationRecognitionRetrievalTransformerContrastive LearningMultimodality
🎯 What it does: Proposes the Multimodal Contextual Contrast (MCC) framework, which aligns local and global attention results in multimodal Transformers to enhance the representational power of multimodal sequence learning.
RetinaQA: A Robust Knowledge Base Question Answering Model for both Answerable and Unanswerable Questions
Prayushi Faldu (Indian Institute of Technology Delhi), Mausam .
TransformerContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes RetinaQA, a knowledge base question answering model that addresses both answerable and unanswerable questions.
Retrieval Augmented Fact Verification by Synthesizing Contrastive Arguments
Zhenrui Yue (University of Illinois Urbana Champaign), Dong Wang (University of Illinois Urbana Champaign)
ClassificationRetrievalTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposed a retrieval-augmented fact verification framework called RAFTS, which determines the truthfulness of claims by retrieving relevant documents, constructing contrastive arguments, and prompting with a few examples.