ACL 2026 Papers — Page 23
Annual Meeting of the Association for Computational Linguistics · 2296 papers
VLN-MME: Diagnosing MLLMs as Language-guided Visual Navigation Agents
Xunyi Zhao (Australian Institute for Machine Learning, Adelaide University), Qi Wu (Australian Institute for Machine Learning, Adelaide University)
Autonomous DrivingExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the VLN-MME framework for unified evaluation of multi-modal large language models (MLLMs) on vision-and-language navigation (VLN) tasks under zero-shot conditions without the need for a simulator.
VLN-NF: Feasibility-Aware Vision-and-Language Navigation with False-Premise Instructions
Hung-Ting Su (National Taiwan University), Winston H. Hsu (National Taiwan University)
Autonomous DrivingRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: A visual language navigation (VLN) benchmark named VLN-NF containing erroneous premise instructions was constructed, and a two-stage hybrid method named ROAM was proposed for evidence-driven target localization and NOT-FOUND judgment in partially observable environments.
Vocabulary Hijacking in LVLMs: Unveiling Critical Attention Heads by Excluding Inert Tokens to Mitigate Hallucination
Yangneng Chen (Harbin Institute of Technology Shenzhen), Jing Li (Harbin Institute of Technology Shenzhen)
Explainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelImageTextMultimodality
🎯 What it does: By conducting a Logit Lens analysis on the attention mechanism of visual language models, the study reveals and locates the Vocabulary Hijacking phenomenon that causes hallucinations.
Vocabulary Shapes Cross-Lingual Variation of Word-Order Learnability in Language Models
Jonas Mayer Martins (University of Göttingen), Lisa Beinborn (University of Göttingen)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsContrastive LearningText
🎯 What it does: Pre-trained a Transformer language model on synthetic variants of word order, investigating the impact of word order variation on model learnability.
VoxMind: An End-to-End Agentic Spoken Dialogue System
Tianle Liang (Zhejiang University), Zhou Zhao (Zhejiang University)
Autonomous DrivingOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Propose VoxMind, an end-to-end speech dialogue agent that combines the 'think first, then speak' mechanism and multi-agent dynamic tool management to achieve internal reasoning and tool invocation.
VOYAGER: A Training Free Approach for Generating Diverse Datasets using LLMs
Avinash Amballa (Samsung Research America), Srinivas Chappidi (Samsung Research America)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes VOYAGER, a training-agnostic and scalable framework for LLMs to generate diverse datasets.
VRPO: Rethinking Value Modeling for Robust RL under Noisy Supervision in LLM Post-Training
Dingwei Zhu (Fudan University), Xuanjing Huang (Fudan University)
OptimizationRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose the VRPO framework, which enhances the robustness of LLM training under noisy supervision by leveraging an improved value model.
Wait! There’s a Way Out: A Decision Mechanism for Forecasting Conversational Derailment
Laerdon Kim (Cornell University), Cristian Danescu-Niculescu-Mizil (Cornell University)
Recommendation SystemAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Propose a dialogue失控 warning system that separates prediction from trigger decision-making, and achieves selective delayed triggering by simulating future utterances.
WebAggregator: Enhancing Compositional Reasoning Capabilities of Deep Research Agent Foundation Models
Rui Wang (MoE Lab, Chinese University of Hong Kong), Kam-Fai Wong (MoE Lab, Chinese University of Hong Kong)
Data SynthesisExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the WebAggregator framework, which enhances the synthetic reasoning capabilities of deep research agents by automatically generating verifiable deep reasoning QA pairs;
WebClipper: Efficient Evolution of Web Agents with Graph-based Trajectory Pruning
Junjie Wang (Ant Group), Jinjie Gu (Ant Group)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Convert the execution trajectory of web agents into a state graph, and generate a more efficient trajectory by pruning it into a minimal necessary DAG; use the pruned trajectory to evolve existing agents, and propose the F-AE score to measure the balance between accuracy and efficiency.
WebCoderBench: Benchmarking Web Application Generation with Comprehensive and Interpretable Evaluation Metrics
Chenxu Liu (Peking University), Tao Xie (Peking University)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Construct and release the WebCoderBench benchmark, utilizing 1,572 real multimodal user requirements to evaluate the quality of LLM-generated web applications.
WebSTAR: Scalable Data Synthesis for Computer Use Agents with Step-Level Filtering
Yifei He (University Of Illinois Urbana Champaign), Xia Song (Microsoft)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextMultimodalityChain-of-Thought
🎯 What it does: Propose an scalable computer usage agent (CUA) data synthesis pipeline that transforms noisy teacher agent replays into high-quality training samples through thought augmentation and step-level filtering.
WebSynthesis: World Model-Guided Monte Carlo Tree Search for Efficient WebAgent Trajectory Synthesis
Yifei Gao (Beijing Jiaotong University), Jitao Sang (Beijing Jiaotong University)
Autonomous DrivingOptimizationComputational EfficiencyRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringWorld ModelTextSequentialBenchmark
🎯 What it does: Proposes the WebSynthesis framework, which achieves efficient offline synthesis of web interaction trajectories through an LLM-driven world model and Monte Carlo Tree Search.
WeightLoRA: Keep Only Necessary Adapters
Andrey Veprikov (Moscow Independent Research Institute of Artificial Intelligence), Aleksandr Beznosikov (Moscow Independent Research Institute of Artificial Intelligence)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes the WeightLoRA framework, which significantly reduces the number of trainable parameters by adaptively selecting the most important LoRA adapters during fine-tuning, while maintaining or improving model performance.
What About the Scene With the Hitler Reference? HAUNT: A Framework to Probe LLMs’ Self-consistency in Closed Domains Via Adversarial Nudge
Arka Dutta (Rochester Institute of Technology), Ashiqur R. KhudaBukhsh (Rochester Institute of Technology)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and evaluated a three-step framework, HAUNT, for detecting factual hallucinations in LLMs when they are subjected to dialogic induction in closed domains (movies and books).
What Deserves Memory: Adaptive Memory Distillation for LLM Agents
Wenquan Ma (Fudan University), WenLong Wu
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes NEMORI, an untrained adaptive memory distillation framework that uses prediction error to determine the future value of LLM agent experiences and uses this as an entry point for memory construction.
What Do LLMs Learn First? Asymmetric Learning Dynamics of Input Complexity and Output Ambiguity in Preference Alignment
Mengyang Li (Tianjin Normal University), Pinlong Zhao (Hangzhou Dianzi University)
OptimizationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Explored two sources of difficulty in DPO training: input complexity and output ambiguity, and proposed DECOPO for independent rhythm training.
What Do Prosody and Text Convey? Characterizing How Meaningful Information is Distributed Across Multiple Channels
Aditya Yadavalli (University of California San Diego), Alex Warstadt (University of California San Diego)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextMultimodalityAudio
🎯 What it does: The study investigates how to quantify the mutual information differences between prosody and text in expressing semantic features such as emotion, sarcasm, and interrogative intent, and constructs an information-theoretic framework to approximately estimate prosody information independent of text.
What Does LLM Refinement Actually Improve? A Systematic Study on Document-Level Literary Translation
Shaomu Tan (University of Amsterdam), Felix Hieber (Amazon AGI)
GenerationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Systematic experiments on iterative self-refinement in document-level literary translation, evaluating the effectiveness of different translation-refinement granularity combinations, prompting strategies, and LLM models.
What If Consensus Lies? Selective-Complementary Reinforcement Learning at Test Time
Dong Yan (University of Chinese Academy of Sciences), Tieniu Tan (University of Chinese Academy of Sciences)
TransformerReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the SCRL framework, achieving robust test-time reinforcement learning on unlabeled test streams through selective positive and negative pseudo-labels and dynamic reward shaping.
What is a protest anyway? Codebook conceptualization is still a first-order concern in LLM-era classification
Andrew Halterman (Michigan State University), Katherine A. Keith (Williams College)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper explores the importance of three stages in text classification in the era of LLMs: conceptualization, annotation, and downstream inference, and verifies the bias caused by conceptualization errors through simulation.
What Makes a Good Curriculum? Disentangling the Effects of Data Ordering on LLM Mathematical Reasoning
Yaning Jia (Dartmouth College), Soroush Vosoughi (Dartmouth College)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: This paper investigates the effectiveness of curriculum learning (CL) in large language models for mathematical reasoning tasks, constructing an offline multi-dimensional CL framework and systematically evaluating the impact of various difficulty metrics and data sorting directions (forward/backward) on model performance.
What Makes AI Research Replicable? Executable Knowledge Graphs as Scientific Knowledge Representations
Yujie Luo (Zhejiang University), Huajun Chen (Zhejiang University)
Representation LearningData-Centric LearningAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose an executable knowledge graph (XKG) and integrate it into three agent frameworks to enhance the effectiveness of replicating AI research from papers.
What Makes an Ideal Quote? Recommending “Unexpected yet Rational” Quotations via Novelty
Powei Chang (Fudan University), Jiaqing Liang (Fudan University)
Recommendation SystemTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a novel citation recommendation system called NOVELQR, aiming to provide users with quotes that are both contextually relevant and surprising;
What Makes Good Instruction-Tuning Data? An In-Context Learning Perspective
Guangzeng Han (University of Memphis), Xiaolei Huang (University of Memphis)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a data selection framework based on weighted contextual influence (wICI), which selects samples most beneficial for instruction fine-tuning using dynamic probe retrieval, influence weighting, and diversity constraints.
What Makes LLMs Effective Sequential Recommenders? A Study on Preference Intensity and Temporal Context
Zhongyu Ouyang (Dartmouth College), Soroush Vosoughi (Dartmouth College)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningSequential
🎯 What it does: This paper proposes a preference optimization framework called RecPO, which leverages large language models to integrate preference intensity and temporal context information in sequence recommendation tasks, thereby improving the modeling of user preferences and recommendation performance.
What’s Left Unsaid? Detecting and Correcting Misleading Omissions in Multimodal News Previews
Fanxiao Li (Yunnan University), Min-Yen Kan (National University Of Singapore)
Explainability and InterpretabilityData-Centric LearningTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelGenerative Adversarial NetworkImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper constructs the MM-MISLEADING benchmark dataset and proposes the OMGUARD framework for detecting and correcting omissive misinformation in multimodal news previews.
When "Correct" Is Not Safe: Can We Trust Functionally Correct Patches Generated by Code Agents?
Yibo Peng (Carnegie Mellon University), Beidi Chen (Carnegie Mellon University)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: This paper investigates and verifies that code agents may generate functionally correct but vulnerable patches (FCV patches) in patching without explicit errors, and proposes a single-query black-box injection attack (FCV-Attack) to induce this behavior.
When Agents Look the Same: Quantifying Distillation-Induced Similarity in Tool-Use Behaviors
Chenghao Yang (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes two measurement methods, response pattern similarity (RPS) and action graph similarity (AGS), to quantify the behavioral homogenization caused by knowledge distillation in LLM agents during tool usage.
When Background Matters: Breaking Medical Vision Language Models by Transferable Attack
Akash Ghosh (Indian Institute of Technology Patna), Xiuying Chen (MBZUAI)
Safty and PrivacyExplainability and InterpretabilityAdversarial AttackTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasound
🎯 What it does: Propose a transferable, black-box multi-modal attack framework called MedFocusLeak, designed to induce visual language models to produce erroneous but credible diagnostic results in medical imaging.
When Benchmarks Leak: Inference-Time Decontamination for LLMs
Jianzhe Chai (Institute of Science Tokyo), Jun Sakuma (Institute of Science Tokyo)
Federated LearningSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Propose the DeconIEP framework, which introduces minor perturbations to input embeddings during inference to suppress memory shortcuts of LLMs on leakage benchmarks, thereby restoring genuine generalization ability.
When Bigger Isn’t Better: A Comprehensive Fairness Evaluation of Political Bias in Multi-News Summarisation
Nannan Huang (RMIT University), Junichi Yamagishi (National Institute of Informatics)
Recommendation SystemFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This study investigates the fairness of multi-document news summarization systems when handling different political perspectives, and systematically evaluates the performance of large language models (LLMs) in terms of political bias and the effectiveness of debiasing methods.
When Correct Beliefs Collapse: Epistemic Resilience of LLMs under Clinical Pressure
Boyu Xiao (Harbin Institute Of Technology), Bing Qin (Harbin Institute Of Technology)
Explainability and InterpretabilityComputational EfficiencyDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataElectronic Health RecordsReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper proposes the MED-STRESS framework to evaluate the belief stability of large models in multi-turn clinical dialogues, revealing a significant gap between knowledge and robustness, and significantly improving robustness through intervention in RBED during inference and fine-tuning with R-FT during training;
When Does Language Matter? Multilingual Instructions Reveal Step-wise Language Sensitivity in Vision-Language-Action Models
Xuan Dong (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningRobotic IntelligenceTransformerPrompt EngineeringVision-Language-Action ModelContrastive LearningTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Systematically evaluated the robustness of vision-language-action (VLA) models under multilingual instructions and found that non-English instructions significantly reduce success rates;
When Does Mixing Help? Analyzing Query Embedding Interpolation in Multilingual Dense Retrieval
Tongyao Zhu (National University of Singapore), Min-Yen Kan (National University of Singapore)
RetrievalTransformerPrompt EngineeringContrastive LearningTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper investigates the effect of mixed-language queries in multilingual dense retrieval, systematically evaluating the impact of controlling the mixing ratio by interpolating different language queries in the embedding space on retrieval performance.
When Efficiency Becomes a Vulnerability: Computational Cost Attacks on WebAgents
Liang-Bo Ning, Wenqi Fan (Hong Kong Polytechnic University)
Safty and PrivacyComputational EfficiencyAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose an attack framework named CostBomb, which generates and selects malicious prompts through LLM to attack the computational cost of WebAgent.
When Efficiency Meets Safety: A Benchmark Security Analysis of KV Cache Compression in Large Language Models
Xiaoxiao Ma (Tsinghua University), Shu-Tao Xia (Tsinghua University)
CompressionSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper systematically evaluates the security impact of KV cache compression on large language models (LLMs) when facing various jailbreak attacks, revealing the 'occasional robustness' and 'vulnerability paradox' brought by compression, and proposes a Safe-CAM defense scheme based on historical feedback.
When High Accuracy Hides Poor Calibration: Rethinking Confidence Evaluation in Transformer-Based Text Classification with Balanced Brier Score
Guilherme Fonseca (Federal University of Minas Gerais), Marcos André Gonçalves (Federal University of Minas Gerais)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper addresses the confidence evaluation of Fine-tuned Transformer in text classification, revealing that high accuracy conceals severe overconfidence, and proposes the Balanced Brier Score to more reliably measure calibration.
When Identity Skews Debate: Anonymization for Bias-Reduced Multi-Agent Reasoning
Hyeong Kyu Choi (University of Wisconsin-Madison), Sharon Li (University of Wisconsin-Madison)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the identity bias theory in Multi-Agent Debate (MAD) and reduce bias through response anonymization
When in Doubt, Consult: Expert Debate for Sexism Detection via Confidence-Based Routing
Anwar Alajmi (University of Warwick), Gabriele Pergola (University of Warwick)
ClassificationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: A two-stage framework was constructed, first using targeted training (class-balanced focal loss, class-aware batching, threshold calibration) to enhance the robustness of the specialized classifier, and then passing uncertain samples to a multi-role reasoning module (Collaborative Expert Judgment) via confidence routing for multi-perspective discussion and judgment.
When Is Thinking Enough? Early Exit via Sufficiency Assessment for Efficient Reasoning
Yang Xiang (Soochow University), Min Zhang (Soochow University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a dynamic early exit framework called DTSR based on reflective signal monitoring and thinking sufficiency checking, allowing large reasoning models to autonomously determine when to stop thinking while generating chain-of-thought (CoT) reasoning.
When KV Cache Reuse Fails in Multi-Agent Systems: Cross-Candidate Interaction is Crucial for LLM Judges
Sichu Liang (Southeast University), Deyu Zhou (Southeast University)
ClassificationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Investigate the failure phenomenon of KV cache reuse on the judge side in multi-agent LLM systems, propose the Judge Consistency Rate (JCR) metric, and reveal through experiments that the absence of cross-candidate attention leads to unstable judgment selection.
When LLMs Read Tables Carelessly: Measuring and Reducing Data Referencing Errors
Yuqing Yang (University of Southern California), Huzefa Rangwala (AWS AI Labs)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: Systematically evaluate the table data reference errors (DRE) that occur in LLMs on table tasks, and propose a Critic-based filtering and rejection sampling method to significantly reduce DRE and improve the final answer accuracy.
When Misinformation Speaks and Converses: Rethinking Fact-Checking in Audio Platforms
Chaewan Chun (Pennsylvania State University), Dongwon Lee (Pennsylvania State University)
RetrievalExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Propose and evaluate a complete fact-checking pipeline for audio dialogues, including claim detection, evidence retrieval, verification, and explanation generation.
When More Words Say Less: Decoupling Length and Specificity in Image Description Evaluation
Rhea Kapur (Stanford University), Elisa Kreiss (University of California, Los Angeles)
Image TranslationGenerationRetrievalExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: This paper investigates the relationship between the specificity and length of image descriptions, constructs a dataset where length and information content are independent, and quantifies the specificity of each description in the control set using CLIPScore. It verifies the consistency between human preference and this metric, and further analyzes the changes in the specificity of VLM-generated descriptions under different length constraints.
When Morphology Hides in Plain Sight: Breaking the Isolation in Vietnamese and Beyond
Anh Trac Duc Dinh (Ho Chi Minh City University of Technology), Tho Quan (Ho Chi Minh City University of Technology)
ClassificationRecognitionTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: This paper addresses the morphological structure challenges of isolating languages such as Vietnamese, proposing HuTieuBERT, a model that incorporates morphological boundary fusion and morphological-aware attention bias on a pre-trained Transformer.
When One LLM Drools, Multi-LLM Collaboration Rules
Shangbin Feng (University of Washington), Yulia Tsvetkov (University of Washington)
Recommendation SystemAnomaly DetectionOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: The paper proposes and systematically introduces a collaborative framework for multiple large language models (LLMs), elaborating on collaborative methods from the API layer to the weight layer, and demonstrating the advantages of multiple LLMs in terms of reliability, diversity, and efficiency.
When Personalization Legitimizes Risks: Uncovering Safety Vulnerabilities in Personalized Dialogue Agents
Jiahe Guo (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Studied the security risks of long-tail memory in personalized language model agents and proposed an intent legalization mechanism.
When Personalization Tricks Detectors: The Feature-Inversion Trap in Machine-Generated Text Detection
Lang Gao (MBZUAI), Xiuying Chen (MBZUAI)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: Proposes StyloBench as the first benchmark for detecting machine-generated text in personalized text scenarios, and finds that existing detectors perform significantly worse or even produce inverted predictions in personalized scenarios; further identifies the Feature-Inversion Trap phenomenon and develops StyloCheck to predict the robustness of detectors during personalized transfer;
When Seeing Is not Enough: Revealing the Limits of Active Reasoning in MLLMs
Hongcheng Liu (Shanghai Jiao Tong University), Yu Wang (Shanghai Jiao Tong University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper studies the capabilities of multi-modal large language models (MLLMs) in active reasoning scenarios, proposing to formalize active reasoning as an interactive target guessing problem, and designing a corresponding evaluation framework and process.
When Seeing Overrides Knowing: Disentangling Knowledge Conflicts in Vision-Language Models
Francesco Ortu (University of Trieste), Alberto Cazzaniga (AREA Science Park)
Explainability and InterpretabilityRepresentation LearningTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Constructed the WHOOPS-AHA! dataset, and studied the internal mechanisms when visual information conflicts with internal knowledge by performing log-linear projection and target intervention on the attention heads of vision-language models (VLMs);
When TableQA Meets Noise: A Dual Denoising Framework for Complex Questions and Large-scale Tables
Shenghao Ye (University of Science and Technology of China), Xiaofeng Jiang (University of Science and Technology of China)
Recommendation SystemOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextTabularRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed a dual denoising framework called EnOtab, which first splits the problem at the evidence level through evidence-level denoising, then progressively prunes the table using an evidence tree, ultimately obtaining clean question-answer data.
When Vision-Language Models Judge Without Seeing: Exposing Informativeness Bias
Xiaohan Zou (Pennsylvania State University), Dan Roth (Oracle AI)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Analyze and eliminate the 'information bias' in visual language models during evaluation tasks, proposing the BIRCH framework that uses self-generated corrected anchors to make evaluations focus on image content.
Where and What: Reasoning Dynamic and Implicit Preferences in Situated Conversational Recommendation
Dongding Lin (Hong Kong Polytechnic University), Wenjie Li (Hong Kong Polytechnic University)
Recommendation SystemTransformerLarge Language ModelVision Language ModelTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a framework named SiPeR for scenario-based conversational recommendation, which can infer users' dynamic implicit preferences based on visual scenes and dialogues, and provide appropriate recommendations and responses.
Where Paths Split: Localized, Calibrated Control of Moral Reasoning in Large Language Models
Chenchen Yuan (Technical University of Munich), Gjergji Kasneci (Technical University of Munich)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Propose a local moral framework control method during inference called Convergent-Divergent Routing (CDR), combined with Dual Logit Calibration (DLC) to achieve calibrated moral reasoning steering.
Where the Cat Sat: A Multilingual Framework for Spatial Language Understanding
Demian Inostroza (University of Melbourne), Meladel Mistica (University of Melbourne)
RecognitionTransformerLarge Language ModelTextMultimodalityBenchmark
🎯 What it does: Built a cross-lingual spatial language understanding framework and created a corresponding benchmark dataset, systematically evaluating the spatial relation recognition capabilities of large language models in three languages (Spanish, Basque, Chinese).
Which Reasoning Trajectories Teach Students to Reason Better? A Simple Metric of Informative Alignment
Yuming Yang (Fudan University), Xuanjing Huang (Fudan University)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Studied the data-student adaptability of long-chain thinking (CoT) trajectories in knowledge distillation, and proposed the Rank-Surprisal Ratio (RSR) metric to measure the applicability of trajectories to specific students.
Who Plays Which Role When? Communication Role Dynamics for Peer Recognition and Team Performance Prediction
Yifan Song (University of Illinois Urbana-Champaign), Tal August (University of Illinois Urbana-Champaign)
RecognitionRecommendation SystemExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequentialReview/Survey Paper
🎯 What it does: The paper constructs an eight-category team communication role lexicon based on educational theory and annotates roles in Slack conversations from university computer science projects, investigating the relationship between role dynamics and team performance.
Who Wrote This Line? Evaluating the Detection of LLM-Generated Classical Chinese Poetry
Jiang Li (Inner Mongolia University), Xiangdong Su (Inner Mongolia University)
Anomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the ChangAn benchmark to systematically evaluate methods for detecting ancient Chinese poetry generated by LLMs
Whose Facts Win? LLM Source Preferences under Knowledge Conflicts
Jakob Schuster (Heidelberg University), Katja Markert (Heidelberg University)
Data SynthesisRetrievalExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper systematically investigates the source preference of large language models in multi-source retrieval augmented generation (RAG) pipelines by constructing synthetic knowledge conflict comparisons and synthetic sources. It evaluates the behaviors of 13 open-source weight models and reveals their hierarchical preferences for source credibility.
Why Are We Moral? An LLM-based Agent Simulation Approach to the Study of Moral Evolution
Zhou Ziheng (University of California Los Angeles), Fangwei Zhong (University of California Los Angeles)
Federated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningRobotic IntelligenceDrug DiscoveryReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built a proxy simulation framework based on large language models (LLM) (MORE+SOCIAL-EVOL) to study the dynamics of moral evolution in prehistoric hunter-gatherer societies.
Why Do Emotions Change? Appraisal-Guided Reasoning for Emotion–Cause Triplet Extraction in Conversations
Qiao Liang (Tongji University), Lin Zhang (IEIT SYSTEMS Co Ltd)
Explainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Propose a reasoning framework called ECFlow based on cognitive appraisal, which is used to extract emotion-cause triplets in multi-modal multi-turn dialogues and ensure emotional consistency at the dialogue level and coherence of core events.
Why Do LLM-based Web Agents Fail? A Hierarchical Planning Perspective
Mohamed Aghzal (George Mason University), Ziyu Yao (George Mason University)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringWorld ModelTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a hierarchical planning-based evaluation framework, systematically decomposing the performance of LLM Web agents across three levels: high-level planning, low-level execution, and replanning, and conducting experiments on the extended Mind2WebLive benchmark.
Why Do More Experts Fail? A Theoretical Analysis of Model Merging
Zijing Wang (Northeastern University), Hinrich Schuetze
Federated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsContrastive LearningTextTabularSequentialBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper explains the reason for performance degradation of model fusion when the number of experts increases through theoretical analysis, and proposes a reparameterized heavy-tailed method (RHT) to expand the parameter space of the fusion model, thereby improving performance.
Why Does Reinforcement Learning Generalize? A Feature-Level Mechanistic Study of Post-Training in Large Language Models
Dan Shi (Tianjin University), Deyi Xiong (Tianjin University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAuto EncoderText
🎯 What it does: After training with the same data on the same base model, comparing SFT and RL, using a feature-level interpretable framework to analyze how the two training methods reshape internal representations, and revealing the mechanism by which RL achieves cross-task generalization.
Why LLM Safety Guardrails Collapse After Fine-tuning: A Similarity Analysis Between Alignment and Fine-tuning Datasets
Lei Hsiung (Dartmouth College), Yaoqing Yang (Dartmouth College)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: The study investigates the collapse mechanism of safety guardrails after downstream fine-tuning of safety-aligned models, using representation similarity analysis to examine the similarity between upstream alignment data and downstream tasks.
Why LLMs Hallucinate on Structured Knowledge: A Mechanistic Analysis of Reasoning over Linearized Representations
Shanghao Li (University Of Illinois Chicago), Philip S. Yu (University Of Illinois Chicago)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextGraphTabular
🎯 What it does: Investigate the internal mechanisms by which LLMs generate hallucinations when reasoning over linearized structured knowledge (such as graphs and tables), propose and quantify two diagnostic metrics—Structural Shortcut Dependency (SSR) and Semantic Alignment Score (SAS), and build a lightweight hallucination detector based on them.
Why Mean Pooling Works: Quantifying Second-Order Collapse in Text Embeddings
Tomomasa Hara (Tohoku University), Sho Yokoi (NINJAL)
Explainability and InterpretabilityRepresentation LearningTransformerContrastive LearningText
🎯 What it does: This paper investigates the problem of second-order statistical information loss caused by the use of mean pooling in text embeddings, and proposes a metric called SOCM (Second-Order Collapse by Mean pooling) to quantify this 'second-order collapse'.
Why Multimodal In-Context Learning Lags Behind? Unveiling the Inner Mechanisms and Bottlenecks
Yu Wang (University of Wisconsin-Madison), Sharon Li (University of Wisconsin-Madison)
Anomaly DetectionExplainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Studied the internal mechanisms of in-context learning (ICL) in multi-modal large language models during inference, systematically decomposing it into two steps: task mapping construction and task mapping transfer, and proposed a mapping-guided inference (MGI) enhancement method during the inference phase;
Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics
Ziwen Xu (Zhejiang University), Ningyu Zhang (Zhejiang University)
OptimizationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Unify LoRA, partial weight fine-tuning, and activation layer intervention into a dynamic weight update framework, and propose a preference-utility (log-odds) analysis, discovering a consistent trade-off pattern between preference and utility across different control methods, then design the SPLIT objective function to enhance preference while maintaining utility.
Why Supervised Fine-Tuning Fails to Learn: A Systematic Study of Incomplete Learning in Large Language Models
Chao Xue (University of New South Wales), Flora D. Salim (University of New South Wales)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextBenchmarkFinance RelatedRetrieval-Augmented Generation
🎯 What it does: This paper systematically studies the incomplete learning phenomenon (ILP) that occurs during supervised fine-tuning (SFT), and constructs a diagnostic framework to identify unlearned samples and their causes.
WildFeedback: Aligning LLMs With In-situ User Interactions And Feedback
Taiwei Shi (University of Southern California), Jennifer Neville (Microsoft Corporation)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequentialRetrieval-Augmented Generation
🎯 What it does: Propose the WILDFEEDBACK framework, which automatically identifies user satisfaction/dissatisfaction feedback in real conversations, constructs dialogue samples for preference training, and evaluates the alignment of the model with real user preferences through checklist-driven evaluation methods.
WildReward: Learning Reward Models from In-the-Wild Human Interactions
Hao Peng (Tsinghua University), Juanzi Li (Tsinghua University)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the WILDFB dataset and trained the reward model WILDREWARD by leveraging interactions between real users and large language models, directly learning preferences from interactions rather than traditional preference pairs;
WIST: Web-Grounded Iterative Self-Play Tree for Domain-Targeted Reasoning Improvement
Fangyuan Li (Harbin Institute of Technology), Yuqiang Li (Shanghai Innovation Institute)
OptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose WIST, a web-based iterative self-play tree framework, aimed at enhancing the reasoning capabilities of large language models without relying on manually curated corpora.
Would LLMs be Good Historical Linguists and Chinese Dialect Learners?
Yicheng Liu (Beijing Institute of Technology), Xingchen Zhang (Beijing Institute of Technology)
Representation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: By introducing historical phonological information of Middle Chinese into large language models, two tasks (conditional phonological change rule induction and sentence-level dialect IPA transcription) are designed, a multi-source dataset is constructed, and LoRA fine-tuning combined with GRPO reinforcement learning is adopted to enhance pronunciation modeling for low-resource Chinese dialects.
Writing-RL: Advancing Long-form Writing via Adaptive Curriculum Reinforcement Learning
Xuanyu Lei (Tsinghua University), Yang Liu (Tsinghua University)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the Writing-RL framework, significantly improving the generation quality of long-form writing models through adaptive curriculum reinforcement learning.
WSDPO: A Generative Word Sense Disambiguation Framework with Chain-of-Thought and Preference Optimization
Kunpeng Kang (Qilu University of Technology), Wenpeng Lu (Qilu University of Technology)
RecognitionOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposed the WSDPO framework, using chain-of-thought and preference optimization to enhance the performance of generative word sense disambiguation models.
XMark: Reliable Multi-Bit Watermarking for LLM-Generated Texts
Jiahao Xu (University of Nevada, Reno), Zikai Zhang (University of Nevada, Reno)
GenerationSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a multi-bit reversible watermarking method called XMARK for text generated by large language models, which can achieve high message decoding accuracy while maintaining text quality.
XOXO: Stealthy Cross-Origin Context Poisoning Attacks against AI Coding Assistants
Adam Štorek (Columbia University), Suman Jana (Columbia University)
Adversarial AttackAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the XOXO attack, which silently pollutes the context of AI programming assistants by performing semantically-preserving code transformations, such as variable renaming in shared code, leading to the generation of code with vulnerabilities or errors.
XToM: Exploring the Multilingual Theory of Mind for Large Language Models
Chunkit Chan (HKUST), Yangqiu Song (HKUST)
TransformerLarge Language ModelPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper constructs a multilingual Theory of Mind (ToM) benchmark called XToM and uses it to systematically evaluate the ToM reasoning capabilities of various large language models in different language environments.
XtraGPT: Context-Aware and Controllable Academic Paper Revision via Human-AI Collaboration
Nuo Chen (National University of Singapore), Bingsheng He (National University of Singapore)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextReview/Survey PaperBenchmarkRetrieval-Augmented Generation
🎯 What it does: Built a specialized framework and tool called XtraGPT for academic paper paragraph revision, supporting human-machine collaboration, context-awareness, and fine-grained revisions according to writing guidelines.
XY-Tokenizer: Mitigating the Semantic-Acoustic Conflict in Low-Bitrate Speech Codecs
Yitian Gong (Fudan University), Xipeng Qiu (Fudan University)
CompressionTransformerLarge Language ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningAudio
🎯 What it does: Propose a low-bitrate (≈1 kbps) speech codec called XY-Tokenizer, which can simultaneously achieve semantic alignment and audio reconstruction;
YIELD: A Large-Scale Dataset and Evaluation Framework for Information Elicitation Agents
Victor De Lima (Georgetown University), Grace Hui Yang (Georgetown University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes an Information Extraction-based Dialogue Agent (IEA), and for the first time constructs a large-scale, ethically sourced dialogue dataset called YIELD. The large language model is fine-tuned on this dataset using POMDP modeling and offline reinforcement learning (AWR).
You Can Have a Second Chance: Unbiased and Multi-bit Watermarking for Diffusion Language Models with Regret-based Remasking
Ke Yang (Institute of Information Engineering Chinese Academy of Sciences), Chi Chen (Institute of Information Engineering Chinese Academy of Sciences)
GenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented UMR, an unbiased multi-bit watermarking scheme for discrete diffusion language models, which combines Regret-based Remasking to provide re-generation opportunities for unwatermarked tokens during the generation process, thereby improving watermark coverage while maintaining generation quality.
Your Inference Request Will Become a Black Box: Confidential Inference for Cloud-based Large Language Models
Chung-ju Huang (Key Lab of High Confidence Software Technologies (Peking University), Ministry of Education), Leye Wang (Key Lab of High Confidence Software Technologies (Peking University), Ministry of Education)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Developed Talaria, a framework that achieves confidentiality during large language model inference in the cloud;
Your Reasoning Benchmark May Not Test Reasoning: Revealing Perception Bottleneck in Abstract Reasoning Benchmarks
Xinhe Wang (Carnegie Mellon University), Jiaqi W. Ma (University of Illinois Urbana-Champaign)
Explainability and InterpretabilityTransformerVision Language ModelImageTextBenchmarkChain-of-Thought
🎯 What it does: The paper verifies that the perceptual bottleneck is the main reason for the difficulty of ARC-style benchmarks by decomposing the visual abstract reasoning task into two stages: perception and reasoning.
Your Reasoning Model is Secretly a Reward Model - Optimization-Free Verification from Experience
Zhenwen Liang (Tencent AI Lab), Dong Yu (Tencent AI Lab)
ClassificationOptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes an untrained hidden state verification method called CLUE, which uses the activation difference in the internal reasoning trajectory of the LLM for binary classification, thereby more reliably selecting the correct answer in high-branch scenarios.
Your Reasoning Model Knows What Counts: Self-Guided Chain-of-Thought Pruning for Efficient Reasoning
Zi-Ao Ma (Beijing Institute Of Technology), Zhijing Wu (Beijing Institute Of Technology)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose a self-guided pruning framework named SGP-CoT, which utilizes the model's own likelihood signals to identify and remove redundant reasoning steps, generating more concise chain-of-thought reasoning.
Your Students Don’t Use LLMs Like You Wish They Did
Sebastian Kobler (University of Sydney), Jonathan K. Kummerfeld (University of Sydney)
ClassificationRecommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Six computational metrics are proposed and studied to evaluate the pedagogical alignment between students and AI dialogue systems, and their effectiveness is tested in 500 course dialogues.
Z3D: Zero-Shot 3D Visual Grounding from Images
Nikita Drozdov (Lomonosov Moscow State University), Maksim Kolodiazhnyi (Lomonosov Moscow State University)
RecognitionObject DetectionSegmentationPose EstimationDepth EstimationTransformerAgentic AIVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTextPoint CloudMeshRetrieval-Augmented Generation
🎯 What it does: Proposes Z3D — a zero-shot 3D visual grounding pipeline capable of locating target objects in 3D scenes using only multi-view images (optional camera poses and depth maps).
ZARA: Training-Free Motion Time-Series Reasoning via Evidence-Grounded LLM Agents
Zechen Li (University of New South Wales), Flora D. Salim (University of New South Wales)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIAuto EncoderTabularTime SeriesRetrieval-Augmented Generation
🎯 What it does: This paper proposes a zero-training motion temporal reasoning framework called ZARA, which interprets multi-sensor temporal data through a statistical knowledge base and retrieval-enhanced LLM agent.
Zero-Shot Detection of LLM-Generated Text using Temperature Sensitivity
Shixuan Ma (Beijing University of Posts and Telecommunications), Quan Wang (Beijing University of Posts and Telecommunications)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a zero-shot LLM-generated text detection method that utilizes temperature sensitivity features to determine the text source.
Zero-shot Jianzi Recognition as Structured Visual Information Extraction in Open Compositional Symbolic Systems
Zehan Li (Northeastern University), Jingwei Cheng (Northeastern University)
RecognitionData SynthesisRecurrent Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringImageTextSequentialRetrieval-Augmented Generation
🎯 What it does: Studied zero-shot guqinjianzi recognition, modeling it as a structured visual information extraction task, and proposed the JZ-Glyph framework.
Zero-shot Large Language Models for Automatic Readability Assessment
Riley Grossman (New Jersey Institute of Technology), Yi Chen (New Jersey Institute of Technology)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextTabularSequentialBenchmark
🎯 What it does: Proposed a zero-shot prompting method and the LAURAE integrated model for unsupervised readability assessment, with systematic evaluation on 14 diverse datasets.
Zero-Shot Multimodal Retrieval with Multi-Scale Contextual Representations
Sourajit Saha (University of Maryland Baltimore County), Tejas Gokhale (University of Maryland Baltimore County)
RetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageVideoTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Propose a two-stage zero-shot multi-modal retrieval framework called Multi-Score, which first uses multi-scale Matryoshka embeddings for efficient candidate filtering, and then performs fine-grained re-ranking by combining bidirectional chain-of-thought embeddings scores and question-answer relevance scores;
ZoomR: Memory Efficient Reasoning through Multi-Granularity Key Value Retrieval
David H. Yang (Rensselaer Polytechnic Institute), Pin-Yu Chen (IBM Research)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose ZoomR, which achieves memory-efficient inference for long-text reasoning by letting LLM generate idea summaries and dynamically selecting the granularity levels of KV cache.