ACL 2026 Papers — Page 20
Annual Meeting of the Association for Computational Linguistics · 2296 papers
SRA: Span Representation Alignment for Large Language Model Distillation
Quoc Phong Dao (Hanoi University of Science and Technology), Trung Le (Monash University)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes the cross-tokenizer knowledge distillation framework SRA, which achieves alignment and knowledge transfer by dividing text into spans and using attention-weighted centroid representations.
SSA: Improving Performance With a Better Scoring Function
Omar Naim, Nicholas Asher
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Analyze the attention saturation and generalization failure caused by Softmax in Transformer, and propose the SSA (Scaled Signed Averaging) scoring function to improve the attention mechanism.
SSG: Logit-Balanced Vocabulary Partitioning for LLM Watermarking
Chenxi Gu (Monash University), John C. Grundy (Monash University)
GenerationSafty and PrivacyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes an LLM watermarking method called SSG based on vocabulary partitioning, aimed at improving the detectability of watermarks in low-entropy tasks.
SSSD: Simply-Scalable Speculative Decoding
Michele Marzollo, Lukas Cavigelli (Huawei)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Developed a training-free inference acceleration method called SSSD, which is based on lightweight n-gram matching and hardware-aware speculation.
Stability Implies Redundancy: Delta Attention Selective Halting for Efficient Long-Context Prefilling
Yujie Chen (Shanghai Jiao Tong University), Linfeng Zhang (University Of California San Diego)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: A new method called DASH (Delta Attention Selective Halting) is proposed to improve computational efficiency in long context pre-filling, reducing the computational cost of large language models and multimodal models.
Stable Language Guidance for Vision–Language–Action Models
Zhihao Zhan (Sun Yat-sen University), Guangrun Wang (Sun Yat-sen University)
Robotic IntelligenceTransformerLarge Language ModelPrompt EngineeringVision-Language-Action ModelDiffusion modelImageTextMultimodality
🎯 What it does: Propose the Residual Semantic Steering (RSS) framework, which enhances the robustness of VLA models to instructions by decoupling visual priors from linguistic semantics.
Stable Signer: Hierarchical Sign Language Generative Model
Sen Fang, Dimitris N. Metaxas (Georgia Institute Of Technology)
GenerationData SynthesisPose EstimationTransformerMixture of ExpertsVision Language ModelDiffusion modelContrastive LearningVideoTextMultimodality
🎯 What it does: Construct an end-to-end Sign Language Production (SLP) system called Stable Signer, which directly converts text into high-quality multi-style sign language videos.
Stable-RAG: Mitigating Retrieval-Permutation-Induced Hallucinations in Retrieval-Augmented Generation
Qianchi Zhang (Advanced Innovation Center for Future Blockchain and Privacy Computing), Zhiming Zheng (Advanced Innovation Center for Future Blockchain and Privacy Computing)
RetrievalExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose Stable-RAG in the RAG (Retrieval-Augmented Generation) system, which reduces hallucinations induced by document permutations by performing spectral clustering on the last layer hidden states of the LLM under different permutations of the top-5 retrieved documents, extracting dominant reasoning patterns, and using DPO (Direct Preference Optimization) to align answers from different permutations for consistency.
Standard-to-Dialect Transfer Trends Differ across Text and Speech: A Case Study on Intent and Topic Classification in German Dialects
Verena Blaschke (LMU Munich), Barbara Plank (LMU Munich)
ClassificationTransformerSupervised Fine-TuningTextMultimodalityAudio
🎯 What it does: A systematic evaluation of intent and topic classification for German and its non-standard dialects (Bavarian dialect, Swiss German) in three settings: text, speech, and speech→text concatenation, along with the first public release of corresponding parallel audio intent datasets.
STAPO: Selective Trajectory-Aware Policy Optimization for LLM Agent Training
Qiuyi Qi (Zhejiang University), Qiang Zhu (Zhejiang University)
Anomaly DetectionOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: This paper proposes a framework called STAPO, which is based on normalized entropy for trajectory missing localization and selective strategy optimization, aiming to improve the performance of LLM agents in long-term tasks.
StealthGraph: Exposing Domain-Specific Risks in LLMs through Knowledge-Graph-Guided Harmful Prompt Generation
Huawei Zheng (Zhejiang University), Dazhen Deng (Zhejiang University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextGraphFinance RelatedRetrieval-Augmented Generation
🎯 What it does: Propose the StealthGraph framework, using knowledge graphs to guide LLMs in generating domain-specific stealthy harmful prompts
Steganography Beyond Pixels: Reimagining Image Steganography as Cross-Modal Linguistic Communication
Ren Lijing, Denghui Zhang (Guangdong Mechanical & Electrical Polytechnic)
Safty and PrivacyTransformerLarge Language ModelDiffusion modelScore-based ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Propose a cross-modal steganography framework called VLace, which discretizes high-resolution images into binary indices through the reversible noise trajectory of diffusion models, and then embeds them into generated text via distribution alignment sampling, achieving image steganography and enabling the recovery of the original image without training.
STEM: Structure-Tracing Evidence Mining for Knowledge Graphs-Driven Retrieval-Augmented Generation
Peng Yu (Kingsoft Corporation), Yinfei Xu (Kingsoft Corporation)
RetrievalExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose Structure-Tracing Evidence Mining (STEM), transforming multi-hop knowledge graph question answering into subgraph retrieval with structural matching and global guidance;
Step-GRPO: Internalizing Dynamic Early Exit for Efficient Reasoning
Benteng Chen (East China Normal University), Min Zhang (East China Normal University)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Integrate a dynamic early exit mechanism into the model through a post-training framework, achieving more efficient inference processes.
StepHint: Multi-level Stepwise Hints Enhance Reinforcement Learning to Reason
Kaiyi Zhang (Renmin University of China), Rui Yan (Wuhan University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose StepHint, an algorithm that uses multi-level step hints to guide large language models in reasoning within the RLVR framework.
Stereotype Bias in a Bilingual Setting: A Culturally Grounded Evaluation in Kazakhstan
Nurkhan Laiyk (Mohamed bin Zayed University of Artificial Intelligence), Fajri Koto (Nazarbayev University)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringScore-based ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Aqbileq dataset for evaluating stereotype bias in a bilingual (Kazakh-Russian) environment in Kazakhstan;
Still Between Us? Evaluating and Improving Voice Assistant Robustness to Third-Party Interruptions
Dongwook Lee (Seoul National University), Sungroh Yoon (Seoul National University)
RecognitionAnomaly DetectionComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityBenchmarkAudio
🎯 What it does: Proposed a complete framework for third-party interruption identification and response, and constructed the TPI-Train dataset and TPI-Bench benchmark.
STK-Adapter: Incorporating Evolving Graph and Event Chain for Temporal Knowledge Graph Extrapolation
Shuyuan Zhao (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
Federated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsTextGraphTime SeriesSequentialRetrieval-Augmented Generation
🎯 What it does: Propose STK-Adapter, which combines an evolutionary graph encoder with a large language model (LLM), achieving hierarchical spatiotemporal structure capture, event chain semantic modeling, and cross-modal alignment through three sets of Mixture-of-Experts (MoE) modules, thus completing the extrapolation task of Temporal Knowledge Graph (TKG).
Stop Fixating on Prompts: Reasoning Hijacking and Constraint Tightening for Red-Teaming LLM Agents
Yanxu Mao (Henan University), Datao You (Henan University)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextTabularBiomedical DataElectronic Health RecordsRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the JailAgent framework, which implicitly controls the reasoning trajectory and memory of LLM Agents through three stages: trigger extraction, reasoning hijacking, and constraint tightening, without modifying user prompts, to achieve red team attacks.
Stop Hardening Everything: A Training-Free Neuron-Level Defense for Neural Ranking Models
Yu-An Liu (State Key Laboratory of AI Safety), Xueqi Cheng (State Key Laboratory of AI Safety)
RetrievalAdversarial AttackTransformerPrompt EngineeringContrastive LearningText
🎯 What it does: To address the adversarial robustness defects of neural retrieval models, this paper proposes a training-free model-level defense framework called 'Search in the Model,' which enhances the model's attack resistance through fine-grained neuron identification and correction without retraining.
Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning
Renliang Sun (UCLA), Wei Wang (NEC Labs America)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a training-agnostic adaptive early stopping framework called REFRAIN, which can detect reflective redundancy during the LLM reasoning process and stop thinking at the appropriate time, thereby reducing over-reasoning;
StoryCoder: Narrative Reformulation for Structured Reasoning in LLM Code Generation
Geonhui Jang, YoungJoon Yoo (Chung Ang University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the STORYCODER framework, which rewrites brief instructions of code generation tasks into coherent natural language narratives to enhance model understanding and reasoning.
Stratagem: Learning Transferable Reasoning via Trajectory-Modulated Game Self-Play
Xiachong Feng (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
Computational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringWorld ModelTextSequentialChain-of-Thought
🎯 What it does: Propose the STRATAGEM framework, which learns transferable reasoning through trajectory advantage modulation during self-play in text games; use the reasoning transferability coefficient (ϕ) and reasoning evolutionary reward (ψ) to multidimensionally evaluate trajectories, and combine them with traditional terminal rewards to reinforce trajectories that demonstrate abstraction, adaptability, and logical coherence across multiple rounds of reasoning.
StratMem-Bench: Evaluating Strategic Memory Use in Virtual Character Conversation Beyond Factual Recall
Yerong Wu (Southeast University), Haofen Wang (Tongji University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed a new benchmark called StratMem-Bench to evaluate the strategic use of memory by virtual characters in dialogues, and proposed four specialized evaluation metrics; simultaneously conducted a unified evaluation of multiple LLMs.
STReasoner: Empowering LLMs for Spatio-Temporal Reasoning in Time Series via Spatial-Aware Reinforcement Learning
Juntong Ni (Emory University), Wei Jin (Emory University)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityTime SeriesBenchmarkChain-of-ThoughtStochastic Differential Equation
🎯 What it does: Proposes STReasoner, a multi-modal large language model designed for spatial-temporal reasoning in time series, and constructs the ST-Bench evaluation benchmark and an SDE-based multi-agent data synthesis pipeline.
Stress Testing Factual Consistency Metrics for Long-Document Summarization
Zain Muhammad Mujahid (University of Copenhagen), Isabelle Augenstein (University of Copenhagen)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper systematically stress-tests six commonly used reference-free fact consistency metrics by evaluating their reliability in the context of long document summaries through fact consistency assessment.
STRIDE-ED: A Strategy-Grounded Stepwise Reasoning Framework for Empathetic Dialogue Systems
Hongru Ji (Northwestern Polytechnical University), Chao Gao (Northwestern Polytechnical University)
GenerationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposed the STRIDE-ED framework, which enhances empathetic dialogue generation quality by constructing a comprehensive emotional strategy system, step-by-step reasoning, and strategy-aware data processing.
StructMem: Structured Memory for Long-Horizon Behavior in LLMs
Buqiang Xu (Zhejiang University), Shumin Deng (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes StructMem, a hierarchical memory framework based on event centers, incorporating event-level binding and cross-event synthesis, aiming to enhance temporal reasoning and multi-hop answering capabilities in long-sequence dialogues.
Structure Guided Retrieval-Augmented Generation for Factual Queries
Miao Xie (China Agricultural University), Chunli Lv (Nanyang Technological University)
GenerationRetrievalRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a new structure-guided retrieval-augmented generation framework, SG-RAG, to address the problem of precise retrieval and answer generation for multi-constraint factual queries.
Structured Episodic Event Memory
Zhengxuan Lu (Southeast University), Baotian Hu (Harbin Institute of Technology)
RetrievalExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: Proposes SEEM, a hierarchical memory framework combining a graph memory layer and an episode memory layer, to enhance the reasoning and consistency of LLM agents during long-term interactions;
Structured Semantic Information Helps Retrieve Better Examples for In-Context Learning Applied to Few-Shot Relation Extraction
Aunabil Chakma (University of Arizona), Eduardo Blanco (University of Arizona)
TransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: A framework is proposed to automatically expand one-shot relational extraction examples, converting 1-shot tasks into 5/10-shot tasks by combining LLM generation, syntax-semantic rule retrieval, and diversity clustering to select examples.
Student Guides Teacher: Weak-to-Strong Inference via Spectral Orthogonal Exploration
Dayu Wang (Baidu Inc), Jizhou Huang (Baidu Inc)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Spectral Orthogonal Exploration (SOE) framework, which utilizes a weak student model as an orthogonal probe to inject orthogonal residuals into the teacher's reasoning process during inference, thereby breaking low-rank bias and improving reasoning diversity and accuracy.
Subject-level Inference for Realistic Text Anonymization Evaluation
Myeong Seok Oh (Tscientific), Hansaem Kim (Yonsei University)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the SPIA benchmark, using individuals as the unit of evaluation, to measure the inferability of individual information after text anonymization.
SubTokenTest: A Practical Benchmark for Real-World Sub-token Understanding
Shuyang Hou (Peking University), Muhan Zhang (Peking University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Designed and released a practical benchmark called SUBTOKENTEST to evaluate the ability of large language models in understanding sub-tokens, covering ten tasks and four domains, and focusing on stripping away the influence of complex reasoning to ensure the evaluation focuses on character-level processing; simultaneously conducted a systematic analysis of 'over-reasoning' and test-time scaling (TTSC) in models on sub-token tasks; further used linear probing to investigate how character information is encoded in the hidden layers of models.
Success and Cost Elicit Convention Formation for Efficient Communication
Saujas Vaduguru (Carnegie Mellon University), Daniel Fried (Carnegie Mellon University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: Proposes a method for training large multimodal models to form language conventions in order to improve communication efficiency with humans.
Suggest-Verify-Revise: A Three-Stage Document-Level Event Causality Identification with Narrative Consistency
Ya Su (Shanxi University), Hongye Tan (Shanxi University)
RecognitionGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphChain-of-Thought
🎯 What it does: Propose a three-phase framework named SVRECI based on suggestion-verification-revision, which uses LLM and hypergraph structures to generate candidate causal relationships, verifies narrative consistency using topological Hawkes processes, and iteratively revises through dynamic causal graphs and dual-layer structural contrastive learning, completing document-level event causality identification.
SURE or Not? Investigating Semantic Understanding in Dense Retrieval Models
Lingdi Kong (University of Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)
RetrievalTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Propose the SURE benchmark, construct controlled retrieval samples targeting three dimensions: semantic accuracy, abstraction, and equivalence, and systematically evaluate the semantic understanding capabilities of 10 dense retrieval models.
Surprisal Minimisation over Goal-directed Alternatives Predicts Production Choice in Dialogue
Thomas P. Utting (University of Aberdeen), Arabella Sinclair (University of Aberdeen)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Modeling the generation of a speaker's utterance in open-ended dialogue as probabilistic cost-sensitive selection over a set of contextual alternatives, distinguishing goal-relevant from goal-irrelevant alternatives;
Survey Response Generation: Generating Closed-Ended Survey Responses In-Silico with Large Language Models
Georg Ahnert (University of Mannheim), Markus Strohmaier (University of Mannheim)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Systematically evaluated the performance of 8 survey response generation methods when using large language models to generate closed-ended survey responses, conducting large-scale experiments (32M simulated responses) across different models, decoding strategies, prompt variations, and response option scales, providing empirical guidance for political attitude surveys.
SWAN: Semantic Watermarking with Abstract Meaning Representation
Ziping Ye (Amazon), Ninareh Mehrabi (Meta)
GenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes the SWAN framework, which embeds watermark signatures into the semantic structure of sentences using Abstract Meaning Representation (AMR), supporting watermarking and detection in text generation without training.
SwissGov-RSD: A Human-annotated, Cross-lingual Benchmark for Token-level Recognition of Semantic Differences Between Related Documents
Michelle Wastl (University of Zurich), Rico Sennrich (University of Zurich)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityBenchmark
🎯 What it does: Constructed and released SwissGov-RSD, a human-annotated cross-lingual document-level semantic difference identification dataset, and evaluated multiple models on this dataset and its synthetic derivative sets.
Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards?
Lorenzo Molfetta (University of Bologna), Gianluca Moro (University of Bologna)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Evaluate the authority citation and temporal sensitivity of large language models in legal reasoning, proposing a four-axis diagnostic framework and a new benchmark LEGALLINK-EU.
SynthAgent: Adapting Web Agents with Synthetic Supervision
Zhaoyang Wang (UNC-Chapel Hill), Huaxiu Yao (UNC-Chapel Hill)
Data SynthesisDomain AdaptationReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the SynthAgent framework for unsupervised adaptation of web agents, using fully synthetic supervision to improve task and trajectory quality
Synthetic Data Generation for Training Diversified Commonsense Reasoning Models
Tianhui Zhang (University of Liverpool), Danushka Bollegala (University of Liverpool)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextChain-of-Thought
🎯 What it does: Constructed and released a synthetic dataset named CommonSyn for training models capable of generating diverse and high-quality common sense reasoning text, and improved the quality-diversity Pareto frontier of the model through a two-stage data generation and screening method.
Synthia: Scalable Grounded Persona Generation from Social Media Data
Vahid Rahimzadeh (Tehran Institute for Advanced Studies, Khatam University), Yadollah Yaghoobzadeh (University of Tehran)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the SYNTHIA framework, which utilizes a small LLM to transform real social media posts into virtual characters with complete life narratives, and generates synthetic populations for social scientific research through demographic matching and questionnaire simulation.
Systematicity between Forms and Meanings across Languages Supports Efficient Communication
Doreen Osmelak (Saarland University), Kate McCurdy (Saarland University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningRecurrent Neural NetworkContrastive LearningText
🎯 What it does: Systematically study the form-meaning mapping of verbs and pronouns across languages, evaluating their learnability and simplicity within an information efficiency framework.
T^\star: Progressive Block Scaling for Masked Diffusion Language Models Through Trajectory Aware Reinforcement Learning
Hanchen Xia (Shanghai Academy Of AI For Science), Siyu Zhu (Shanghai Academy Of AI For Science)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningDiffusion modelText
🎯 What it does: Proposed the T/Star progressive block size expansion training strategy, which scales the block size incrementally based on TRACERL reinforcement learning for Masked Diffusion Language Models (MDM), to enhance parallel decoding performance while maintaining inference accuracy.
TA-GRPO-d: Trajectory-Aware GRPO for Optimizing Denoising Trajectories in Diffusion LLMs
Gyunyeop Kim (Gachon University), Sangwoo Kang (Gachon University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningDiffusion modelText
🎯 What it does: This paper proposes a trajectory-aware strategy based on reinforcement learning to dynamically decide which masked positions in the decoding process of discrete diffusion large language models need to be unmasked, as well as how many positions should be unmasked at each step, thus achieving adaptive and low-latency decoding.
TabEmb: Joint Semantic-Structure Embedding for Table Annotation
Ehsan Hoseinzade, Anandharaju Durai Raju
ClassificationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTabularBenchmark
🎯 What it does: Proposes the TabEmb framework, which uses frozen LLMs to generate column semantic embeddings, and then performs structured information passing on the column graph through graph neural networks, completing three table annotation tasks: column type, column property, and table type.
Table Question Answering in the Era of Large Language Models: A Comprehensive Survey of Tasks, Methods, and Evaluation
Wei Zhou (Bosch Center for Artificial Intelligence), Mohsen Mesgar (Bosch Center for Artificial Intelligence)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityTabularReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Reviews the task setting, challenges, methods, and evaluation in the field of table question answering (TQA), focusing on the latest advances in the era of large language models (LLMs).
TabReX: Tabular Referenceless eXplainable Evaluation
Tejas Anvekar (Arizona State University), Vivek Gupta (Arizona State University)
Explainability and InterpretabilityData-Centric LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed TABREX, a no-reference, graph-based interpretable table evaluation framework, and created the TABREX-BENCH large-scale perturbation benchmark;
Tackling the Root of Misinformation by Teaching Laypeople about Logical Fallacies via Socratic Questioning and Critical Argumentation
Minjing Shi (ETH Zürich), Mrinmaya Sachan (ETH Zürich)
Human-Computer InteractionTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Design and implement a logic fallacy teaching system called LFTutor based on large language models, which uses Socratic questioning and critical argumentation to guide learners in identifying and correcting logical fallacies in everyday speech through multi-round dialogues.
TagSpeech: End-to-End Multi-Speaker ASR and Diarization with Fine-Grained Temporal Grounding
Mingyue Huo (University of Illinois Urbana Champaign), Yuheng Zhang (University of Illinois Urbana Champaign)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextAudio
🎯 What it does: Propose TagSpeech, a unified end-to-end multi-speaker speech recognition and separation framework that can simultaneously output who spoke what and the specific time points.
TAIGR: Towards Modeling Influencer Content on Social Media via Structured, Pragmatic Inference
Nishanth Sridhar Nakshatri (Purdue University), Dan Goldwasser (Purdue University)
Recommendation SystemExplainability and InterpretabilityData-Centric LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringVideoTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Propose the TAIGR framework, which conducts a three-stage analysis of influencer content on social media: extracting core recommendations (takeaway), constructing an argumentation structure, and inferring credibility through probabilistic factor graphs;
Tailored Primitive Initialization is the Secret Key to Reinforcement Learning
Yihang Yao (Carnegie Mellon University), Chuang Gan (MIT-IBM Watson AI Lab)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose an automated process called Tailor, which initializes large language models (LLMs) by generating diverse and high-quality reasoning primitives before reinforcement learning, significantly improving the reinforcement learning performance on reasoning tasks.
Tailoring Diagnostic Modeling to Individual Learners: Personalized Distractor Generation via MCTS-Guided Reasoning Reconstruction
Tao Wu (Zhejiang University), Fei Wu (Zhejiang University)
Recommendation SystemExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelMixture of ExpertsTextTabularChain-of-Thought
🎯 What it does: By analyzing students' historical answer records, personalized misconception prototypes are constructed, and these prototypes are used to generate targeted distractors in new questions to improve diagnostic effectiveness.
TALAS: Teacher-Anchored Layer Alignment with Adaptive Sharpness-Aware Minimization for Embedding Distillation
Quoc Phong Dao (Hanoi University of Science and Technology), Trung Le (Monash University)
CompressionKnowledge DistillationRepresentation LearningTransformerAuto EncoderContrastive LearningSimultaneous Localization and MappingText
🎯 What it does: This paper proposes a compression method for sentence embedding models called TALAS, which aligns the final sentence embeddings of the teacher model with the student model across multiple layers, and further transfers semantic structure through bottom-up hierarchical self-distillation. It also employs adaptive sharpness-aware minimization (ASAM) to achieve a flatter loss surface. Additionally, SimCSE contrastive learning is incorporated as a regularization during training.
Talking to a Know-It-All GPT or a Second-Guesser Claude? How Repair reveals distinct Multi-Turn Behavior in LLMs
Clara Lachenmaier (Bielefeld University), Sina Zarrieß (Bielefeld University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Studying how LLMs perform repair in multi-turn dialogues and whether they proactively or passively initiate repair when faced with unanswerable questions.
TalkLoRA: Communication-Aware Mixture of Low-Rank Adaptation for Large Language Models
Lin Mu (Anhui University), Yiwen Zhang (Anhui University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose a communication-aware LoRA Mixture-of-Experts framework called TalkLoRA, which uses a lightweight dialogue module to achieve information interaction between low-rank experts, thereby enhancing the parameter-efficient fine-tuning effect of large language models.
TAMA: Target-Aware Multilingual Abuse Detection by Cascaded Conditional Multi-Task Learning
JiYan Liu, Diana Maynard (University of Sheffield)
ClassificationRecognitionAnomaly DetectionTransformerSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Proposed a multilingual targeted abuse detection benchmark called TAMA for public figures, and designed a three-task joint learning framework named Cascaded-MTL;
TAMAS: Benchmarking Adversarial Risks in Multi-Agent LLM Systems
Ishan Kavathekar (International Institute of Information Technology Hyderabad), Tanuja Ganu (Microsoft Research India)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed a multi-agent LLM security evaluation benchmark named TAMAS, which includes 5 high-risk domains, 6 attack types, 300 attack instances, and 100 harmless tasks. It evaluated the performance of 10 mainstream LLMs and 3 multi-agent collaboration configurations (Autogen and CrewAI) on metrics such as ARIA, PNA, and ERS.
TamEdit: Trajectory-Aware Meta-Learning for Specificity-Preserving Continual Knowledge Editing
Shiqiang Tian (East China Normal University), Liang He (East China Normal University)
Meta LearningTransformerSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: Propose TamEdit, a trajectory-aware meta-learning framework for continuous knowledge editing
Taming "Zombie" Agents: A Markov State-Aware Framework for Resilient Multi-Agent Evolution
Taolin Zhang (Hefei University of Technology), Richang Hong (Hefei University of Technology)
OptimizationFederated LearningComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed the AgentRevive framework to achieve dynamic soft pruning in multi-agent systems based on Markov states, allowing 'zombie' agents to recover and contribute in subsequent rounds;
Taming Actor-Observer Asymmetry in Agents via Dialectical Alignment
Bobo Li (National University of Singapore), Wynne Hsu (National University of Singapore)
OptimizationFederated LearningExplainability and InterpretabilityKnowledge DistillationRepresentation LearningData-Centric LearningMeta LearningDrug DiscoveryAI Code AssistantNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextTabularBenchmarkFinance RelatedChain-of-Thought
🎯 What it does: This study investigates the Actor-Observer Asymmetry (AOA) imbalance that occurs when large language model agents perform role-playing, and proposes the ReTAS method based on Dialectical Alignment to eliminate this bias.
Taming Extreme Tokens: Covariance-Aware GRPO with Gaussian-Kernel Advantage Reweighting
Cheng Wang (National University of Singapore), Muhao Chen (University of California, Davis)
OptimizationRepresentation LearningTransformerLarge Language ModelReinforcement LearningScore-based ModelGaussian SplattingTextBenchmark
🎯 What it does: Propose a covariance-aware variant of GRPO, which uses a Gaussian kernel to weight advantages, thereby dynamically balancing exploration and exploitation in the inference tasks of large language models.
Taming System Complexity: Demystifying Software Engineering Agents in Diagnosing Linux Kernel Faults
Zhenhao Zhou (Fudan University), Yiling Lou (Fudan University)
Anomaly DetectionAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Built a defect localization (FL) benchmark for the Linux kernel (LINUXFLBENCH), evaluated the performance of existing LLM agents on this benchmark, found that the effectiveness was lower than that of general software, and then proposed the LINUXFL+ framework to improve localization accuracy.
TARE: Lightweight Token-Aware Representation Editing for Fine-tuning Transformer-like Models
Yulong Wang (Beijing University of Posts and Telecommunications), Siyu Zhao (Beijing University of Posts and Telecommunications)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a lightweight Token-Aware Representation Editing (TARE) method to achieve parameter-efficient fine-tuning of Transformers, keeping the backbone frozen, and only using a small number of trainable parameters to perform fine-grained dimension-wise scaling and bias adjustment on the hidden representations of each token.
Task-Aware LLM Routing with Multi-Level Task-Profile-Guided Data Synthesis for Cold-Start Scenarios
Hui Liu (City University of Hong Kong), Haoliang Li (City University of Hong Kong)
Data SynthesisRecommendation SystemFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a multi-level task hierarchy and a data synthesis framework based on task characteristics to generate diverse question-answer pairs, thereby training LLM routers in cold-start scenarios where domain-specific training data is absent; meanwhile, it introduces TRouter, which treats potential task types as intermediate variables and incorporates prior regularization to improve routing performance.
TaxPraBen: A Scalable Benchmark for Structured Evaluation of LLMs in Chinese Real-World Tax Practice
Gang Hu (Yunnan University), Kun Yue (Wuhan University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and released TaxPraBen — an end-to-end evaluation benchmark tailored for Chinese tax practice, integrating 14 self-built datasets (totaling 7.3k instances), covering 10 traditional application tasks and 3 real-world scenarios (tax risk prevention, tax audit analysis, tax planning), and conducting zero-shot and one-shot evaluations on 19 mainstream LLMs.
TEA-Bench: A Systematic Benchmarking of Tool-enhanced Emotional Support Dialogue Agent
Xingyu Sui (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes TEA-Bench, an interactive benchmark for tool-enhanced emotional support dialogues, and constructs the TEA-Dialog dataset, which includes 81 scenarios and 365 high-quality tool-enhanced dialogues, to evaluate the tool calling and emotional expression capabilities of large language models in emotional support.
Teach a Reward Model to Correct Itself: Reward Guided Adversarial Failure Discovery for Robust Reward Modeling
Pankayaraj Pathmanathan (University of Maryland College Park), Furong Huang (University of Maryland College Park)
Safty and PrivacyExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a control decoding method based on the reward model itself (REFORM), which identifies misjudgment patterns of the reward model by generating samples that are inconsistent with the reward model but still belong to the same preference category, and then self-improves;
Teaching Language Models to Check Grounded Claim Factuality with Human Test-Taking Strategies
Yuxuan Ye (University of Bristol), Edwin Simpson (University of Bristol)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Reformulate grounded claim factuality checking as a reading comprehension task, and guide LLMs to perform claim decomposition and atomic fact checking by designing multi-step reasoning prompts based on human testing strategies; subsequently, use LLMs as teachers and adopt a two-stage training process (SFT and DPO) to obtain an SLM that can generate reasoning explanations with lower costs than large models.
Teaching LLMs Human-Like Editing of Inappropriate Argumentation via Reinforcement Learning
Timon Ziegenbein (Leibniz University Hannover), Henning Wachsmuth (Leibniz University Hannover)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Train LLM to generate self-contained, meaning-preserving sentence-level editing suggestions through reinforcement learning, improving the appropriateness of argumentative texts.
Team-Based Self-Play With Dual Adaptive Weighting for Fine-Tuning LLMs
Wu Li (Harbin Institute of Technology), Jing Li (Harbin Institute of Technology)
OptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: The paper proposes a fully self-supervised team-based self-play algorithm called TPAW, which uses historical checkpoints as both opponents and main players to align large language models.
TeamFusion: Supporting Open-ended Teamwork with Multi-Agent Systems
Jiale Liu (Pennsylvania State University), Qingyun Wu (Pennsylvania State University)
Autonomous DrivingOptimizationFederated LearningRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityReview/Survey PaperChain-of-Thought
🎯 What it does: Propose TeamFusion, a multi-agent framework that generates team deliverables retaining diverse perspectives and achieving greater consensus through structured discussions among agent proxies.
Tears or Cheers? Benchmarking LLMs via Culturally Elicited Distinct Affective Responses
Chongyuan Dai (Hefei University of Technology), Jinpeng Hu (Hefei University of Technology)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and constructed CEDAR — a multimodal benchmark specifically designed to evaluate the performance of large language models in cross-cultural emotional understanding;
TeCES: Collaborative Geometric Knowledge Representation Framework under Evolving Fact Snapshots
Jiujiang Guo (Tianjin University), Yiwei Wei (China University of Petroleum-Beijing)
Federated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningGraphTabularTime Series
🎯 What it does: The paper proposes the TeCES framework, which utilizes snapshots and 2-level geometric algebra to collaboratively model temporal knowledge graphs, achieving a unified characterization of the semantics of static and dynamic facts.
Ted-Tok: Maintaining an Evolving Vocabulary for Lifelong Learning
Jiameng Huang (Peking University), Di He (Peking University)
CompressionFederated LearningRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Proposes Temporal Drift Tokenizer (Ted-Tok), a BPE tokenizer that can adaptively evolve its vocabulary during the lifelong learning process.
TED-TTS: Training-Free Intra-Utterance Emotion and Duration Control for Text-to-Speech Synthesis
Qifan Liang (National University of Singapore), Ye Wang (National University of Singapore)
GenerationTransformerLarge Language ModelPrompt EngineeringFlow-based ModelTextAudio
🎯 What it does: Propose a training-free TTS framework TED-TTS, which can achieve fine-grained control of emotion and speaking speed within a single sentence;
TellWhisper: Tell Whisper Who Speaks When
Yifan Hu (Inner Mongolia University), Rui Liu (Tencent Technology Co.Ltd)
RecognitionTransformerSupervised Fine-TuningContrastive LearningAudio
🎯 What it does: Propose a unified multi-speaker automatic speech recognition framework called TellWhisper, which models time and speaker dynamics simultaneously within the encoder, achieving end-to-end task of 'who said what and when'.
TEMA: Anchor the Image, Follow the Text for Multi-Modification Composed Image Retrieval
Zixu Li (Shandong University), Liqiang Nie (Harbin Institute of Technology)
RetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes the TEMA image retrieval framework assisted by multi-modified semantic text (MMT), and constructs two multi-modified datasets, M-FashionIQ and M-CIRR.
Temp-R1: A Unified Autonomous Agent for Complex Temporal KGQA via Reverse Curriculum Reinforcement Learning
Zhaoyan Gong (Zhejiang University), Wen Zhang (Zhejiang University)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsTextGraphTime SeriesRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a self-contained end-to-end agent named Temp-R1 for handling time knowledge graph question answering tasks that require multi-hop and multi-constraint reasoning.
Template-assisted Contrastive Learning of Task-oriented Dialogue Sentence Embeddings
Minsik Oh (Stanford University), Guoyin Wang (Alibaba Qwen Pilot)
Data SynthesisRepresentation LearningTransformerPrompt EngineeringContrastive LearningText
🎯 What it does: A study proposes an unsupervised task-oriented dialogue sentence embedding method called TaDSE, which utilizes template and slot information for data augmentation, contrastive learning, and reasoning.
Temporal Evidence Chain for Temporal Knowledge Graph Question Answering with Large Language Models
Shihao Liu (Chinese Academy of Sciences), Geyuan Zhang (Chinese Academy of Sciences)
Recommendation SystemAutonomous DrivingFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationData-Centric LearningMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelGraphTime SeriesSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the TECQA framework for the Temporal Knowledge Graph Question Answering (TKGQA) task, aiming to construct temporal evidence chains and assist large language models (LLMs) in reasoning.
Temporal Leakage in Search-Engine Date-Filtered Web Retrieval: A Retrospective Forecasting Case Study
Ali El Lahib (University of California, San Diego), Xinyu Pi (University of California, San Diego)
RetrievalExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextTime SeriesBenchmarkRetrieval-Augmented Generation
🎯 What it does: Systematically audit the date filtering functions of Google Search and DuckDuckGo, discovering that most search results contain later information leakage, and evaluate the impact of these leaks on the prediction accuracy of LLMs.
Temporal Precision Matters: Brain-Tuning Speech Language Models with Millisecond-Resolution Neural Signals
Zhejun Zhang (Beijing University of Posts and Telecommunications), Lei Li (Beijing University of Posts and Telecommunications)
RecognitionRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningTime SeriesBiomedical DataAudio
🎯 What it does: This paper enhances the alignment between the model and human brain activity by using millisecond-level temporal information of electrocorticography (ECoG) signals to fine-tune a speech-language model.
Temporal Sampling for Forgotten Reasoning in LLMs
Yuetai Li (University of Washington), Radha Poovendran (University of Washington)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMixture of ExpertsTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper investigates the Temporal Forgetting phenomenon that occurs during the fine-tuning of large language models, and proposes a method to recover the forgotten reasoning path by sampling answers from multiple checkpoints in the training trajectory during inference through Temporal Sampling.
TensorLens: End-to-End Transformer Analysis via High-Order Attention Tensors
Ido Andrew Atad (Tel Aviv University), Lior Wolf (Tel Aviv University)
Explainability and InterpretabilityRepresentation LearningTransformerAuto EncoderContrastive LearningImageText
🎯 What it does: This paper proposes TensorLens, a linear operator method that represents the entire Transformer model as a high-order attention tensor, achieving unified modeling and interpretability of the model's global behavior;
TeRA: Vector-based Random Tensor Network for High-Rank Adaptation of Large Language Models
Yuxuan Gu (Imperial College London), Danilo Mandic (Imperial College London)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Designed TeRA, a method for efficiently fine-tuning large-scale language models using tensor networks to achieve high-rank parameter updates.
Test of Time: Rethinking Temporal Signal of Benchmark Contamination
Terry Jingchen Zhang (Jinesis Lab, University of Toronto & Vector Institute), Zhijing Jin (Jinesis Lab, University of Toronto & Vector Institute)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextTime SeriesSequentialBenchmark
🎯 What it does: Studied the sensitivity of LLMs to performance degradation caused by post-truncation in time-split evaluation, and demonstrated the impact of question generation methods on temporal patterns.
Test-Time Reasoners Are Strategic Multiple-Choice Test-Takers
Nishant Balepur (University of Maryland), Rachel Rudinger (University of Maryland)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Examines the performance of large language models in multiple-choice questions under two settings: full input and options-only input, and explores their reasoning strategies during test-time reasoning (TTR).
TexOCR: Advancing Document OCR Models for Compilable Page-to-LaTeX Reconstruction
Chengye Wang (Zhejiang University), Yilun Zhao (Yale University)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a complete framework for reconstructing scientific PDF pages into compilable LaTeX. It first constructs the TEXOCR-Bench benchmark and the TEXOCR-Train large-scale training set, then trains the TEXOCR model, and verifies its effectiveness under multi-dimensional evaluation.
Text-Attributed Knowledge Graph Enrichment with Large Language Models for Medical Concept Representation
Mohsen Nayebi Kerdabadi (University of Kansas), Zijun Yao (University of Kansas)
Representation LearningDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextGraphElectronic Health Records
🎯 What it does: Construct a global heterogeneous medical knowledge graph (diagnosis, drugs, surgeries), generate node descriptions and edge reasoning texts using large language models, and produce unified medical concept embeddings through LLM-GNN co-learning for improving EHR sequence prediction.
Text-Guided Multi-Scale Frequency Representation Adaptation
Weicai Yan (Zhejiang University), Tao Jin (Zhejiang University)
RetrievalDomain AdaptationRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed FreqAdapter, which utilizes text guidance to perform multi-scale adaptive optimization of visual features in the frequency domain, thereby enhancing the performance of multimodal models;
Text-to-Distribution Prediction with Quantile Tokens and Neighbor Context
Yilun Zhu (Amazon.com, Inc.), Shervin Malmasi (Amazon.com, Inc.)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose a quantile token regression model based on retrieval augmentation for predicting the full conditional distribution from text inputs.
Text2Tabular – Reconstructing Tabular Research Data from Scientific Publications
Jonas Gottal (Technical University of Munich), Florian Matthes (Technical University of Munich)
RestorationData SynthesisTransformerLarge Language ModelGaussian SplattingTextTabularBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a system called Text2Tabular that automatically extracts statistical summaries from scientific papers and reconstructs complete tabular data.
Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models
Woody Haosheng Gan (University of Southern California), Willie Neiswanger (University of Southern California)
Image TranslationRestorationObject DetectionSegmentationDepth EstimationSuper ResolutionCompressionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelAuto EncoderImageTextMultimodality
🎯 What it does: This paper enhances the model's visual reasoning performance by extracting the steering vector from the text language model and injecting it into the internal representation of the multimodal large model.
TH-RAG : Topic-Based Hierarchical Knowledge Graphs for Robust Multi-hop Reasoning in Graph-based RAG Systems
JungHyoun Kim (Hanyang University), Yong Suk Choi (Hanyang University)
RetrievalRecommendation SystemExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraphAgriculture RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the TH-RAG framework, which constructs a three-layer hierarchical knowledge graph consisting of topics-subtopics-entities in one go. It utilizes LLMs to extract topics/subtopics and triplets, and employs topic-driven graph traversal and sentence-level retrieval to achieve a retrieval-augmented generation system for multi-hop reasoning.
The “Knowledge–Behavior Gap” in Cultural Taboo Safety of Large Language Models
Ying He (Fudan University), Yanghua Xiao (Fudan University)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed and implemented the first benchmark for evaluating cultural taboos safety in large language models, CulShield, covering 77 countries and over 2020 taboos;
The African Languages Lab: A Collaborative Approach to Advancing Low-Resource African NLP
Sheriff Issaka (University of California Los Angeles), Saadia Gabriel (University of California Los Angeles)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Built the African Languages Lab (All Lab), which includes a brand-new multimodal data collection platform called All Voices. It collected and validated 19 billion tokens of textual data and 12,628 hours of audio data across 40 African languages, and used this data to fine-tune the Llama-3.2-1B model to generate a translation model.