π― What it does: Propose Value Flows, a reinforcement learning framework that utilizes flow-matching models to estimate the full future return distribution;
π― What it does: This paper proposes VARestorer, a single-step real image super-resolution framework based on a visual autoregressive model (VAR), which can directly generate high-quality images from low-quality images in a single forward pass.
π― What it does: Propose Variational Autoencoding Discrete Diffusion (VADD), integrating latent variable structures into the reverse process of Masked Diffusion Models (MDMs) to enable the model to implicitly capture correlations between dimensions, thereby improving the quality of samples generated with fewer steps.
Variational Deep Learning via Implicit Regularization
Jonathan Wenger (Columbia University), John Patrick Cunningham
CodeClassificationExplainability and InterpretabilityComputational EfficiencyImage
π― What it does: This paper proposes Implicit Bayesian Variational Inference (IBVI), which achieves variational deep learning by leveraging the implicit regularization of gradient descent, without explicitly using the KL regularization term.
CodeSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMeshBenchmark
π― What it does: Constructed the first 3D Visual Question Answering dataset for ancient Greek pottery, VaseVQA-3D, and trained a specialized Vision-Language Model (VLM) called VaseVLM based on this dataset.
π― What it does: Designed and implemented an online learning semantic cache system, vCache, which learns threshold values for each cache vector and guarantees error rates as specified by users, addressing the insufficient reliability and hit rate issues caused by traditional fixed thresholds.
CodeDrug DiscoveryGraph Neural NetworkLarge Language ModelWorld ModelBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose VCWorld, a cell-level white-box simulator based on a biological knowledge graph and LLM reasoning, for predicting drug-induced gene expression changes.
VEAttack: Downstream-agnostic Vision Encoder Attack against Large Vision Language Models
Hefei Mei (City University of Hong Kong), Chang Xu (University of Sydney)
CodeAdversarial AttackTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Proposed a gray-box attack method called VEAttack that targets only the visual encoder, achieving non-targeted, low-perturbation attacks on large vision-language models by minimizing the cosine similarity of visual features.
VeriEquivBench: An Equivalence Score for Ground-Truth-Free Evaluation of Formally Verifiable Code
Lingfei Zeng (Huazhong University of Science and Technology), Jie Fu (Shanghai Artificial Intelligence Laboratory)
CodeAI Code AssistantLarge Language ModelTextBenchmark
π― What it does: Constructed the VeriEquivBench benchmark, containing 2,389 complex algorithm problems along with their natural language descriptions, Python and Dafny implementations, unit tests, and formal specifications, and proposed an equivalence score evaluation method without relying on existing benchmarks.
Verification of the Implicit World Model in a Generative Model via Adversarial Sequences
AndrΓ‘s Balogh (University of Szeged), MΓ‘rk Jelasity (University of Szeged)
CodeGenerationAdversarial AttackTransformerLarge Language ModelWorld ModelSequential
π― What it does: Investigated the reliability of implicit world models in generative sequence models, proposing adversarial generation of legal game sequences to uncover model errors;
VerifyBench: Benchmarking Reference-based Reward Systems for Large Language Models
Yuchen Yan (Zhejiang University), Yueting Zhuang (Zhejiang University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Constructed and evaluated a reward system benchmark called VerifyBench and its more challenging variant VerifyBench-Hard, specifically designed for reference answer verification, to measure the accuracy of large language models in reasoning tasks for answer validation.
Verifying Chain-of-Thought Reasoning via Its Computational Graph
Zheng Zhao (FAIR at Meta), Nicola Cancedda (FAIR at Meta)
CodeExplainability and InterpretabilityAuto EncoderTextChain-of-Thought
π― What it does: Propose a white-box method called Circuit-based Reasoning Verification (CRV), which replaces the MLP of LLM with a sparse interpretable Transcoder, constructs attribution graphs for reasoning steps, extracts their structural features, and uses them to determine the correctness of reasoning steps.
Zhe Ye (University of California, Berkeley), Dawn Song (University of California, Berkeley)
CodeGenerationTextBenchmark
π― What it does: Proposed VERINA, a high-quality Lean benchmark covering code, specifications, and proof generation for evaluating verifiable code generation;
VeriRole: Verifiable Role-Awareness through Hint-Guided Reinforcement Learning
Zongsheng Wang (Renmin University of China), Baoxun Wang (Tencent)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark
π― What it does: Propose the VeriRole framework, which extracts verifiable facts using a 'hint' mechanism and enhances the role awareness and consistency of role-play dialogue models through Verifiable Role Awareness Reward (VRAR) combined with reinforcement learning.
Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning
Hao Tan (University of Chinese Academy of Sciences), Zhen Lei (University of Chinese Academy of Sciences)
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodality
π― What it does: Constructed the HydraFake dataset containing diverse real and forged images, and developed the VERITAS deepfake detector based on a multi-modal large language model (MLLM).
VeriTrail: Closed-Domain Hallucination Detection with Traceability
Dasha Metropolitansky (Microsoft Research), Jonathan Larson (Microsoft Research)
CodeAnomaly DetectionLarge Language ModelTextBenchmark
π― What it does: This study addresses the problem of closed-domain hallucinations in multi-step (MGS) language model generation, proposing the VeriTrail method to achieve hallucination detection and traceability.
π― What it does: This paper proposes the Verifier-free Test-time Scalable Diffusion Model (VFScale), which achieves scalable inference capabilities during testing by increasing the number of samples, thereby enabling the completion of more complex reasoning tasks.
Video-KTR: Reinforcing Video Reasoning via Key Token Attribution
Ziyue Wang (ByteDance), Xudong Jiang (Nanyang Technological University)
CodeExplainability and InterpretabilityReinforcement LearningVision Language ModelVideo
π― What it does: Propose the Video-KTR framework, which identifies critical tokens in video reasoning by conducting comparative analysis of three factors: visual, temporal, and uncertainty, and updates only these tokens in reinforcement learning;
Video-LevelGauge: Investigating Contextual Positional Bias in Video Language Models.
Hou Xia (University of Science and Technology of China), Yongdong Zhang (University of Science and Technology of China)
CodeLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: This paper proposes and implements the Video-LevelGauge benchmark for systematically evaluating biases of large video-language models (LVLMs) at different context positions;
VideoAnchor: Reinforcing Subspace-Structured Visual Cues for Coherent Visual-Spatial Reasoning
Zhaozhi Wang (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
CodeRepresentation LearningTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: This paper investigates the shortcomings of multimodal large language models (MLLMs) in visual-spatial reasoning, proposing a no-training, plug-and-play VideoAnchor module. It leverages the self-expressive properties of sparse subspace clustering (SSC) to enhance attention during inference, thereby improving the consistency of visual cues and the accuracy of spatial reasoning.
VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling
Xinhao Li (Nanjing University), Limin Wang (Nanjing University)
CodeCompressionTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodality
π― What it does: Proposed a hierarchical video compression method called HiCo, combined with multi-stage short-long learning, the LongVid dataset, and the multi-hop pinhole video sandpile evaluation, to build the VideoChat-Flash video multimodal large language model, which can efficiently process long videos and achieve state-of-the-art performance on multiple benchmarks.
VideoNSA: Native Sparse Attention Scales Video Understanding
Enxin Song (University of California, San Diego), Zhuowen Tu (University of California, San Diego)
CodeRecognitionComputational EfficiencyTransformerVision Language ModelVideo
π― What it does: This work proposes VideoNSA, a framework that migrates Native Sparse Attention (NSA) to video-language models, significantly reducing attention computation while preserving video frame information;
VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning
Yang Ding (Tsinghua University), Yujiu Yang (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVideoMultimodality
π― What it does: Propose the VideoZoomer framework, enabling multimodal large language models to dynamically invoke time-domain scaling tools during long video reasoning processes, obtaining critical details through multi-round interactive communication.
ViPER: Empowering the Self-Evolution of Visual Perception Abilities in Vision-Language Models
Juntian Zhang (Renmin University of China), Rui Yan (Wuhan University)
CodeGenerationData SynthesisReinforcement LearningVision Language ModelDiffusion modelImageMultimodalityBenchmark
π― What it does: Proposed the ViPER framework, which enables self-evolution of visual perception capabilities in vision-language models (VLM) through self-generated data and two-stage reinforcement learning.
Ming Li (University of Central Florida), Chen Chen (ByteDance Seed)
CodeGenerationData SynthesisOptimizationData-Centric LearningVision Language ModelDiffusion modelImageVideo
π― What it does: Propose the Poly-DPO method in visual generation and construct the ViPO large-scale preference dataset to address the noise and distribution imbalance issues in existing preference data.
Wei-Yao Wang (Sony Group Corporation), Yuki Mitsufuji (Sony Group Corporation)
CodeSegmentationRetrievalRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: Propose the VIRTUE framework, integrating the pre-trained segmentation model SAM2 with a vision-language model, enabling users to encode images at the entity level using visual prompts such as points, boxes, or masks while preserving global context, and constructing a large-scale SCaR retrieval benchmark to evaluate visual interaction capabilities.
CodeGenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Propose the VisCodex framework, which merges visual models and code models through task vector model merging to build a unified multimodal code generation model.
VisioMath: Benchmarking Figure-based Mathematical Reasoning in LMMs
Can Li (Beijing Normal University), Hua Huang (Beijing Normal University)
CodeSupervised Fine-TuningVision Language ModelMultimodalityBenchmarkChain-of-Thought
π― What it does: Constructed the VisioMath benchmark, containing 1,800 K-12 math problems, all with high similarity graphical options, testing multi-image reasoning.
Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models
Wenxuan Huang (East China Normal University), Shaohui Lin (East China Normal University)
CodeTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
π― What it does: Constructed a 200K multi-modal chain-of-thought (CoT) dataset without human annotation, and used this dataset to perform cold start initialization for large multi-modal language models (MLLMs); subsequently, adopted reinforcement learning (GRPO) combined with progressive thinking suppression training (PTST), significantly enhancing the model's complex reasoning capabilities through RL training using only 10K multi-modal math samples;
Vision-SR1: Self-Rewarding Vision-Language Model via Reasoning Decomposition and Multi-Reward Policy Optimization
Zongxia Li (Tencent AI Seattle Lab), Dong Yu (University of Maryland)
CodeReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
π― What it does: Propose Vision-SR1, a self-rewarding reinforcement learning framework that decomposes visual reasoning into visual description and language reasoning, and significantly enhances visual reasoning capabilities by leveraging model self-evaluation to obtain visual rewards.
Vision-Zero: Scalable VLM Self-Evolution via Multi-Agent Self-Play
Qinsi Wang (Duke University), Wentian Zhao (Adobe Inc)
CodeRepresentation LearningData-Centric LearningReinforcement LearningVision Language ModelImageMultimodality
π― What it does: Propose the Vision-Zero framework, leveraging a multi-agent 'Who's the Spy?' visual self-play game to achieve zero-shot self-evolution for Vision-Language Models (VLM).
VisionReasoner: Unified Reasoning-Integrated Visual Perception via Reinforcement Learning
Yuqi Liu (CUHK), Jiaya Jia (SmartMore)
CodeObject DetectionSegmentationReinforcement LearningVision Language ModelImageMultimodality
π― What it does: Propose VisionReasoner, a unified vision-language framework capable of performing multiple visual perception tasks such as detection, segmentation, and counting through a structured reasoning process.
CodeCompressionComputational EfficiencyTransformerVision Language ModelImageVideoBenchmark
π― What it does: Propose VisionTrim, a unified training-agnostic visual token compression framework for accelerating inference in multimodal large language models (MLLM);
VisJudge-Bench: Aesthetics and Quality Assessment of Visualizations
Yupeng Xie (Hong Kong University of Science and Technology), Yuyu Luo (Hong Kong University of Science and Technology)
CodeLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmark
π― What it does: Constructed the VISJUDGE-BENCH benchmark to evaluate the judgment capabilities of multimodal large language models (MLLM) in visual quality (credibility, expressiveness, aesthetics), and fine-tuned the VISJUDGE model based on this benchmark.
π― What it does: Proposed VAREdit, an instruction-guided image editing method based on a visual autoregressive (VAR) framework, employing a multi-scale 'next scale' prediction strategy.
Visual Multi-Agent System: Mitigating Hallucination Snowballing via Visual Flow
Xinlei Yu (National University of Singapore), Shuicheng YAN
CodeTransformerVision Language ModelImageVideoMultimodalityBenchmark
π― What it does: Investigate the 'visual hallucination snowball effect' in multi-agent systems and propose a lightweight, model-agnostic ViF method to alleviate this issue.
Yi Xu (University of Cambridge), Ivan VuliΔ (University of Cambridge)
CodeTransformerReinforcement LearningImage
π― What it does: This paper proposes the 'visual planning' paradigm, which uses large visual models to perform multi-step planning through pure image sequences, and designs a two-stage training framework VPRL based on reinforcement learning;
π― What it does: Proposes the Prompt-Agnostic Evolution (PAE) framework, treating Visual Prompt Tuning (VPT) as a Koopman-Lyapunov discrete dynamical system, adopting task-aware frequency-domain initialization and shared evolution operators to significantly accelerate convergence and improve accuracy.
Visual Self-Refine: A Pixel-Guided Paradigm for Accurate Chart Parsing
Jinsong Li (Chinese University of Hong Kong), Dahua Lin (Chinese University of Hong Kong)
CodeRecognitionVision Language ModelImageMultimodalityBenchmark
π― What it does: Propose the Visual Self-Refinement (VSR) framework and implement ChartVSR for chart parsing, generating pixel-level localization, visual feedback, iterative correction, then decoding the final structured data, and constructing a challenging ChartP-Bench benchmark;
VisualPrompter: Semantic-Aware Prompt Optimization with Visual Feedback for Text-to-Image Synthesis
Shiyu Wu (Institute of Automation, Chinese Academy of Sciences), Jing Liu (Institute of Automation, Chinese Academy of Sciences)
CodeGenerationLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: Developed a no-training, training-free prompt engineering framework called VisualPrompter, which fine-grained improves the model adaptability and semantic consistency of user prompts through visual feedback self-reflection (SERE) and task-specific optimization (TSPO);
π― What it does: Constructed the VisuRiddles benchmark and synthesizer, and proposed the Perception-Augmented Visual Reasoner (PAVR) model, aiming to enhance the performance of multimodal large language models on abstract visual reasoning (AVR) tasks.
π― What it does: Vivid-VR is a generative video restoration method based on Diffusion Transformer (DiT), which utilizes ControlNet to perform controllable generation on low-quality videos, enhancing texture realism and temporal consistency.
Vlaser: Vision-Language-Action Model with Synergistic Embodied Reasoning
Ganlin Yang (University of Science and Technology of China), Zhi Hou (Shanghai AI Laboratory)
CodeRobotic IntelligenceTransformerLarge Language ModelMixture of ExpertsVision-Language-Action ModelFlow-based ModelImageTextMultimodality
π― What it does: Proposed Vlaser, a foundation model integrating vision-language-action capabilities, combining high-level reasoning with end-to-end robot control;
VLM-SubtleBench: How Far Are VLMs from Human-Level Subtle Comparative Reasoning?
Minkyu Kim (Krafton), Dongmin Park (Krafton)
CodeAutonomous DrivingPrompt EngineeringVision Language ModelImageTextBenchmarkChain-of-Thought
π― What it does: Designed and released the VLM-SubtleBench benchmark, comprising 13K image pairs with subtle differences, along with question-answering and description tasks, to evaluate visual language models' performance in fine-grained comparative reasoning.
VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety
Shruti Palaskar (Apple), Joseph Yitan Cheng
CodeData SynthesisSafty and PrivacyImageTextMultimodalityBenchmark
π― What it does: Proposed the VLSU framework, constructed a dataset of 8,187 real image-text pairs tailored for multimodal safety, and conducted systematic evaluation of model performance on safety assessment tasks.
VoG: Enhancing LLM Reasoning through Stepwise Verification on Knowledge Graphs
Wenxin Zhao (Hong Kong University of Science and Technology), Lei Chen (Hong Kong University of Science and Technology)
CodeGraph Neural NetworkTransformerLarge Language ModelAgentic AITextGraphRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes the VoG framework, which enhances the accuracy and robustness of multi-hop question answering through a cyclic process of progressive retrieval, verification, and revision of knowledge graphs guided by LLM-generated reasoning plans.
VPI-Bench: Visual Prompt Injection Attacks for Computer-Use Agents
Tri Cao (National University of Singapore), Bryan Hooi (Cyber Emerging Tech and R&D)
CodeSafty and PrivacyAdversarial AttackLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringImageTextBenchmark
π― What it does: Constructed the VPI attack model and introduced the VPI-Bench benchmark to evaluate the robustness of Computer-Use Agents and Browser-Use Agents under visual prompt injection (VPI) attacks.
π― What it does: Proposes a negative prompt guidance method called Value Sign Flip (VSF), which dynamically suppresses unwanted content in diffusion or flow-matching image/video generation models within a few steps (1-8 steps) by flipping attention values.
VTool-R1: VLMs Learn to Think with Images via Reinforcement Learning on Multimodal Tool Use
Mingyuan Wu (University of Illinois Urbana-Champaign), Klara Nahrstedt (University of Illinois Urbana-Champaign)
CodeReinforcement LearningAgentic AIPrompt EngineeringVision Language ModelMultimodalityTabularChain-of-Thought
π― What it does: Fine-tune visual language models using reinforcement learning to automatically generate and utilize intermediate visual steps during reasoning, enabling the ability to 'think with images.'
π― What it does: Propose WAFT, an iterative optical flow estimation framework based on warping, which removes the high-cost cost volume and performs multi-step iterative updates of the flow using high-resolution feature warping.
Viraj Prabhu (Salesforce Research), Ran Xu (Salesforce Research)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This paper proposes the WALT framework, which automatically generates callable tools by reverse engineering website functions (such as search, filtering, and publishing), replacing traditional low-level UI operations;
WARP: Weight Teleportation for Attack-Resilient Unlearning Protocols
Mohammad mahdi Maheri (Imperial College London), Hamed Haddadi (Imperial College London)
CodeSafty and PrivacyConvolutional Neural NetworkTransformerImage
π― What it does: Proposed the WARP (Weight Jumping) defense mechanism to enhance privacy security during the approximate machine unlearning process, reducing the success rates of membership inference and data reconstruction attacks.
WATS: Wavelet-Aware Temperature Scaling for Reliable Graph Neural Networks
Xiaoyang Li (Independent Researcher), Chang Xu (University of Sydney)
CodeClassificationGraph Neural NetworkGraph
π― What it does: Propose a post-calibration framework named WATS, which utilizes graph heat kernel wavelet features to predict temperatures for each node, thereby calibrating the confidence of GNN.
CodeRetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningVideoTextMultimodalityAudio
π― What it does: Propose the WAVE model, constructing a unified audio-visual-text embedding space and supporting cross-modal retrieval and prompt-aware embeddings.
π― What it does: Propose WISDOM, which captures multi-scale features of task evolution through a learnable wavelet transform network, achieving fast adaptation in non-stationary reinforcement learning.
wd1: Weighted Policy Optimization for Reasoning in Diffusion Language Models
Xiaohang Tang (University College London), Ilija Bogunovic (UniversitΓ€t Basel)
CodeComputational EfficiencyTransformerLarge Language ModelReinforcement LearningDiffusion modelText
π― What it does: Proposed a ratio-free weighted policy optimization method (wd1) for fine-tuning discrete diffusion large language models to enhance reasoning capabilities
Weak-to-Strong Generalization with Failure Trajectories
Ruimeng Ye (University Of Tulsa), Bo Hui (Northwestern University)
CodeOptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningText
π― What it does: Propose a generalization framework from weak models to strong models, leveraging the success and failure trajectories generated by weak models to construct a hierarchical trajectory tree, and fine-tuning the strong model using TreeDPO or MCTS to enhance its reasoning and decision-making capabilities.
WearVox: An Egocentric Multichannel Voice Assistant Benchmark for Wearables
Zhaojiang Lin (Meta), Xin Luna Dong (Meta)
CodeTransformerLarge Language ModelBenchmarkAudio
π― What it does: Propose the WearVox benchmark, collecting 3,842 multi-channel egocentric audio samples from AI glasses and evaluating five wearable speech tasks.
Web-CogReasoner: Towards Multimodal Knowledge-Induced Cognitive Reasoning for Web Agents
Yuhan Guo (Southwestern University of Finance and Economics), Yong Dai (Fudan University)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmarkChain-of-Thought
π― What it does: Proposed the Web-CogKnowledge framework, decomposing the learning of Web agents into two stages: knowledge acquisition and cognitive processes, and constructed Web-CogDataset and Web-CogBench, training a multimodal Web agent Web-CogReasoner based on knowledge-driven Chain-of-Thought (CoT);
WebWatcher: Breaking New Frontiers of Vision-Language Deep Research Agent
Xinyu Geng (The Hong Kong University of Science and Technology), Jingren Zhou (Tongyi Lab, Alibaba Group)
CodeRetrievalLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVision-Language-Action ModelMultimodalityBenchmark
π― What it does: Proposed a multimodal deep research agent called WebWatcher, capable of performing joint reasoning on visual and textual information and utilizing multiple tools to accomplish complex information retrieval and inference tasks.
WeTok: Powerful Discrete Tokenization for High-Fidelity Visual Reconstruction
Shaobin Zhuang (Shanghai Jiao Tong University), Yali Wang (Shenzhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
π― What it does: Designed and implemented a discrete visual tokenizer, WeTok, with high compression ratio and high-quality reconstruction, supporting zero-shot image reconstruction and generation at compression ratios of 400% and above.
CodeOptimizationFlow-based ModelRectified FlowBiomedical Data
π― What it does: Propose the WFR-FM framework, which utilizes flow matching methods to jointly learn velocity fields and growth rates, solving dynamic unbalanced Wasserstein-Fisher-Rao optimal transport without simulation, applied to single-cell trajectory inference.
CodeOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Propose the AutoGEO framework, which leverages large language models to automatically extract preference rules from generative search engines. The framework employs rule-driven prompt models (AutoGEOAPI) and reinforcement learning models (AutoGEOMini) to enhance document visibility across multiple LLM engines while maintaining response quality.
Junxi Yan (Tsinghua University), Jingtao Zhan (Tsinghua University)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Proposes a tri-decomposition of cross-entropy, identifying error entropy as the true quantity that varies with model scale, and provides the error entropy scaling law.
What's In My Human Feedback? Learning Interpretable Descriptions of Preference Data
Rajiv Movva (University of California Berkeley), Emma Pierson (University of California Berkeley)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackLarge Language ModelAuto EncoderText
π― What it does: Propose the WIMHF method, which automatically discovers interpretable measurable and expressive preference features from human feedback data using sparse autoencoders, aiding in understanding and utilizing preference data;
Whatever Remains Must Be True: Filtering Drives Reasoning in LLMs, Shaping Diversity
GermΓ‘n Kruszewski (Naver Labs Europe), Marc Dymetman (Independent Researcher)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Propose the DMVR framework, which uses verifiable rewards to filter out incorrect answers, defines a target distribution, and approximates it through Ξ±- DPG, thereby enhancing output diversity while maintaining correctness.
When Agents βMisrememberβ Collectively: Exploring the Mandela Effect in LLM-based Multi-Agent Systems
Naen Xu (Zhejiang University), Shouling Ji (Zhejiang University)
CodeLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextBenchmark
π― What it does: This paper constructs the MANBENCH benchmark targeting the Mandela Effect in LLM multi-agent systems, systematically evaluating its occurrence and persistence across multi-task and multi-protocol scenarios.
CodeData SynthesisTransformerLarge Language ModelAgentic AITextBenchmarkFinance Related
π― What it does: This paper investigates the risks of multi-agent systems driven by large language models (LLMs) collaborating to commit financial fraud on social platforms. It constructs a full lifecycle fraud simulation benchmark named MAFF-Bench, and evaluates the amplification effect of multi-agent collaboration on fraud success rate and population impact rate on this benchmark. Additionally, it explores the effectiveness of three categories of defense strategies: content warnings, agent monitoring, and community resistance.
π― What it does: To address the model collapse issue in long-term test-time adaptation (TTA), this paper proposes the Adaptive and Selective Reset (ASR) framework, combined with importance-aware knowledge recovery regularization and prediction inconsistency-driven online adaptation adjustment, achieving robust adaptation in continuous domain drift scenarios.
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningChain-of-Thought
π― What it does: Studied the exploration behavior of large language models in multi-armed bandit (MAB) tasks, compared supervised fine-tuning (SFT) and reinforcement learning (RL) training methods, designed new reward signals, and evaluated their performance across different environments and time durations.
When LLMs get significantly worse: A statistical approach to detect model degradations
Jonas M. KΓΌbler (Amazon), George Karypis (Amazon)
CodeTransformerLarge Language ModelText
π― What it does: To address the potential accuracy degradation in large language models (LLMs) after inference optimization, this paper proposes a statistical hypothesis testing framework based on the McNemar test to determine whether observed accuracy drops are real or due to noise.
When MLLMs Meet Compression Distortion: A Coding Paradigm Tailored to MLLMs
Jinming Liu (Shanghai Jiao Tong University), Yan Lu (Eastern Institute of Technology)
CodeCompressionTransformerVision Language ModelAuto EncoderContrastive LearningMultimodality
π― What it does: Study the impact of compression distortion on multimodal large language models (MLLMs) and propose the CoTAM compression algorithm, which utilizes shallow CLIP attention to guide bit allocation, lightweight adapters, and multi-level reconstruction loss to preserve low-level details while ensuring high-level semantics.
When More is Less: Understanding Chain-of-Thought Length in LLMs
Yuyang Wu (Peking University), Yisen Wang (Peking University)
CodeLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: Systematically investigate the impact of Chain-of-Thought (CoT) length on large language model reasoning performance, and prove that there exists an 'optimal' CoT length; reveal the scaling laws of optimal length with task difficulty, model size, and per-step computational demands through controlled experiments, theoretical error propagation analysis, and RL-based adaptive regulation; further propose practical methods based on optimal length training and length-filtering voting.
π― What it does: Investigated the vulnerability of pre-trained models to unlearnable examples (UEs) and proposed the BAIT framework to overcome the failure of UEs caused by pre-trained priors
When Reasoning Meets Compression: Understanding the Effects of LLMs Compression on Large Reasoning Models
Nan Zhang (Pennsylvania State University), Rui Zhang (Pennsylvania State University)
CodeCompressionExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelTextBenchmark
π― What it does: Quantize, distill, and prune large reasoning models (LRMs), and analyze the impact of compression on inference capability using performance benchmarks and fine-grained mechanisms.
When Silence Is Golden: Can LLMs Learn to Abstain in Temporal QA and Beyond?
Xinyu Zhou (HKUST), Seyed Ali Bahrainian (University of TΓΌbingen)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextTime SeriesChain-of-Thought
π― What it does: This paper systematically studies how to train large language models to learn self-denial (abstention) in time-sensitive question-answering tasks.
When Style Breaks Safety: Defending LLMs Against Superficial Style Alignment
Yuxin Xiao (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
π― What it does: Studied the impact of style patterns on the safety of large language models, and found that style patterns can lead to ASR (attack success rate) inflation, proposing SafeStyle to counteract the security risks caused by surface style alignment;
When Thinking Backfires: Mechanistic Insights into Reason-induced Misalignment
Hanqi Yan (King's College London), Yulan He (King's College London)
CodeSafty and PrivacyExplainability and InterpretabilityLarge Language ModelContrastive LearningTextChain-of-Thought
π― What it does: Explores and reveals the 'Reasoning-Induced Misalignment (RIM)' phenomenon, where enhancing the reasoning capabilities of Large Language Models (LLMs) (via Chain-of-Thought prompts or training) paradoxically increases the model's tendency to respond to malicious requests.
When to Ensemble: Identifying Token-Level Points for Stable and Fast LLM Ensembling
Heecheol Yun (KAIST), Eunho Yang (KAIST)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
π― What it does: Proposed the SAFE framework, which dynamically determines when to perform probability-level fusion during long-text generation based on tokenization mismatch and model consistency, thereby enhancing the stability and efficiency of LLMs.
When to use Graphs in RAG: A Comprehensive Analysis for Graph Retrieval-Augmented Generation
Zhishang Xiang, Jinsong Su (Xiamen University)
CodeRetrievalGraph Neural NetworkLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented Generation
π― What it does: Proposed GraphRAGBench, a comprehensive benchmark to evaluate the performance of Graph Retrieval-Augmented Generation (GraphRAG) compared to traditional RAG, and conducted systematic experiments on multiple GraphRAG frameworks.
Where Did It Go Wrong? Attributing Undesirable LLM Behaviors via Representation Gradient Tracing
Zhe Li (Singapore Management University), Jun Sun (Singapore Management University)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
π― What it does: Propose a framework based on Representation Gradient Tracing (RepT) to diagnose and attribute harmful, inaccurate, or backdoor-polluted output behaviors in large language models.
Why Keep Your Doubts to Yourself? Trading Visual Uncertainties among Vision-Language Models
Jusheng Zhang (Sun Yat-sen University), Keze Wang (Sun Yat-sen University)
CodeReinforcement LearningAgentic AIMixture of ExpertsVision Language ModelMultimodalityBenchmark
π― What it does: This paper proposes the Agora framework, which quantifies the cognitive uncertainty of multimodal vision-language models as tradable assets and achieves multi-agent collaboration through market mechanisms.
π― What it does: This paper proposes a new benchmark framework for estimating Local Intrinsic Dimension (LID), which can perform consistent evaluation of the same manifold across different domains;
Wide-In, Narrow-Out: Revokable Decoding for Efficient and Effective DLLMs
Feng Hong (Shanghai Jiao Tong University), Jiangchao Yao (Shanghai Jiao Tong University)
CodeComputational EfficiencyLarge Language ModelTextMultimodality
π― What it does: Proposed the WINO algorithm, which performs reversible draft-verification parallel decoding on the decoding process of DLLM, significantly improving speed and quality.
π― What it does: Propose a real-time online 3D reconstruction framework called WinT3R, which can instantaneously predict camera poses and point cloud maps from continuous image streams.
π― What it does: Strategy optimization of vision-language-action (VLA) models based on pixel-level video generation world models, performing on-policy RL entirely in an 'imagined' environment, eliminating the high sampling costs of real robot interactions.
WorldTree: Towards 4D Dynamic Worlds from Monocular Video using Tree-Chains
Qisen Wang (Beihang University), Jia Li (Beihang University)
CodeGenerationGaussian SplattingOptical FlowVideo
π― What it does: Propose the WorldTree framework to achieve 4D dynamic reconstruction from monocular videos, incorporating the Temporal Partition Tree (TPT) and Spatial Ancestor Chain (SAC) to enable coarse-to-fine temporal optimization and multi-level spatial representation;
WOW-Seg: A Word-free Open World Segmentation Model
Danyang Li (Nankai University), Xiang Li (NKIARI)
CodeSegmentationTransformerLarge Language ModelImageBenchmark
π― What it does: Proposed a lexical-free open-world segmentation model called WOW-Seg, which automatically identifies arbitrary objects through visual prompts and outputs masks and categories.
WSVD: Weighted Low-Rank Approximation for Fast and Efficient Execution of Low-Precision Vision-Language Models
Haiyu Wang (New York University), Sai Qian Zhang (New York University)
CodeCompressionComputational EfficiencyVision Language ModelMultimodalityBenchmark
π― What it does: This work proposes the WSVD (Weighted Low-Rank Approximation) framework to compress and accelerate the inference of vision-language models (VLMs) through techniques such as weighted low-rank approximation, head-wise SVD, and quantization-aware fine-tuning, particularly achieving efficient decoding at low precision.
π― What it does: Compare the scaling behavior of xLSTM and Transformer during training and inference, plot the compute-loss Pareto frontier, compute the optimal model and its dependence on context length, and model inference time.
XModBench: Benchmarking Cross-Modal Capabilities and Consistency in Omni-Language Models
Xingrui Wang, Zicheng Liu
CodeTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBenchmarkAudio
π― What it does: Proposed XModBench, a specialized benchmark for evaluating cross-modal consistency in multimodal large language models, covering six modality combinations of audio, visual, and text;
π― What it does: This study proposes an online adaptation interactive medical image segmentation framework (OAIMS), which enhances segmentation performance under distribution shifts by adaptively updating the model on a per-image and per-click basis using user click feedback.
Your Language Model Secretly Contains Personality Subnetworks
Ruimeng Ye (University of Tulsa), Bo Hui (University of Tulsa)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelContrastive LearningText
π― What it does: Extract sparse subnetworks specific to different personas from pre-trained LLMs using activation-guided structured pruning without training steps.