AAAI Conference on Artificial Intelligence Β· 2140 papers
LLM-Guided Quantified SMT Solving over Uninterpreted Functions
Kunhang Lv (SKLCS and Key Laboratory of System Software, ISCAS), Jian Zhang (SKLCS and Key Laboratory of System Software, ISCAS)
CodeOptimizationComputational EfficiencyLarge Language ModelPrompt EngineeringBenchmarkChain-of-Thought
π― What it does: Leverage large language models (LLMs) to provide semantic guidance for quantified uninterpreted functions (UF) in nonlinear real arithmetic formulas, generating more effective instantiations and significantly improving SMT solving efficiency.
CodeKnowledge DistillationLarge Language ModelText
π― What it does: Proposed an adaptive knowledge distillation framework called AdaKD based on tokens, which dynamically adjusts the distillation process by leveraging the real-time learning difficulty of each token.
π― What it does: Propose LLMC+, a pluggable compression benchmark and toolbox supporting over 20 algorithms and five VLM families, with systematic evaluation of compression effectiveness in spatial/temporal redundancy, fine-grained tasks, and multi-turn dialogues.
π― What it does: Propose the LLMTM benchmark to evaluate the ability of LLMs in tasks such as identifying, detecting, and constructing temporal motifs in dynamic graphs, and develop tool-enhanced LLM agents with structure-aware schedulers to achieve self-adaptive balance between accuracy and computational cost.
Local Guidance for Configuration-Based Multi-Agent Pathfinding
Tomoki Arita (National Institute of Advanced Industrial Science and Technology), Keisuke Okumura (National Institute of Advanced Industrial Science and Technology)
CodeOptimizationBenchmark
π― What it does: Propose introducing local guidance in configuration-based multi-agent path planning, improving the LaCAM algorithm to enhance solution quality and real-time performance.
CodeComputational EfficiencyRepresentation LearningRobotic IntelligenceTransformerLarge Language ModelVision Language ModelNeural Radiance FieldSimultaneous Localization and MappingMultimodalityPoint Cloud
π― What it does: Propose LOG-Nav, a hierarchical planning framework for layout-aware object navigation, which enables efficient path planning and execution from global to local levels in unexplored multi-room indoor environments using LLM-driven agents.
Logic Unseen: Revealing the Logical Blindspots of Vision-Language Models
Yuchen Zhou (Sun Yat-sen University), Tat-Seng Chua (National University of Singapore)
CodeTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityBenchmark
π― What it does: This paper proposes LogicBench benchmark and LogicCLIP training framework to evaluate and enhance the understanding of logical relationships in vision-language models (VLM).
π― What it does: Proposes the LoGIC framework for joint structured sparsification of shared trunk and multi-task LoRA modules in multi-task Vision Transformers (ViT), enabling efficient deployment.
LoKI: Low-Damage Knowledge Implanting of Large Language Models
Runyu Wang, Tianbo Ji (South China University Of Technology)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Propose LoKI, a parameter-efficient fine-tuning framework based on the internal knowledge storage mechanism of Transformers, which identifies low-impact weights through knowledge vector attribution and fine-tunes these weights in a layer-balanced manner;
Long-form RewardBench: Evaluating Reward Models for Long-form Generation
Hui Huang (Harbin Institute of Technology), Tiejun Zhao (Harbin Institute of Technology)
CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose Longβform RewardBenchβthe first benchmark for evaluating reward models in long-text generation, covering five subtasks: QA, RAG, Chat, Writing, and Reasoning; simultaneously design the Longβform NeedleβinβaβHaystack test to investigate the impact of error positions and lengths on reward models.
LongLLaDA: Unlocking Long Context Capabilities in Diffusion LLMs
Xiaoran Liu (Fudan University), Xipeng Qiu (Fudan University)
CodeGenerationRetrievalTransformerLarge Language ModelDiffusion modelTextBenchmark
π― What it does: Systematically evaluate the performance of diffusion-based large language models (dLLMs) in long contexts and propose a training-free context extension method called LongLLaDA
LongT2IBench: A Benchmark for Evaluating Long Text-to-Image Generation with Graph-structured Annotations
Zhichao Yang (Xidian University), Leida Li (Xidian University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextGraphBenchmarkChain-of-Thought
π― What it does: Constructed a long-text-image generation alignment evaluation benchmark, LongT2IBench (14K long-text-image pairs with graph-structured annotations), and proposed the LongT2IExpert evaluator, which leverages a multimodal large language model to achieve alignment scoring and structured explanations through hierarchical chain-of-thought reasoning;
Look as You Think: Unifying Reasoning and Visual Evidence Attribution for Verifiable Document RAG via Reinforcement Learning
Shuochen Liu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeRetrievalExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the Chain-of-Evidence (CoE) reasoning paradigm and the Look-As-You-Think (LAT) reinforcement learning framework to achieve verifiable step-by-step visual evidence attribution for Visual Document Retrieval-Augmented Generation (VD-RAG).
Look Closer! An Adversarial Parametric Editing Framework for Hallucination Mitigation in VLMs
Jiayu Hu (Chongqing University), Zhongshi He (Chongqing University)
CodeExplainability and InterpretabilityAdversarial AttackPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Proposed and implemented a framework called ALEAHallu based on adversarial parameter editing, aimed at reducing hallucinations (content inconsistent with visual inputs) generated by visual language models when producing text.
Look-Back: Implicit Visual Re-focusing in MLLM Reasoning
Shuo Yang (Peking University), Li Yuan (Peking University)
CodeLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelMultimodalityBenchmark
π― What it does: In this study, the authors propose an implicit visual refocusing method called LookβBack, which leverages multimodal large language models (MLLM) to automatically refocus image information during reasoning without the need for explicitly injecting visual prompts;
π― What it does: This paper proposes a training-free high-resolution image synthesis framework called LookFlow, which can rapidly and high-quality reconstruct low-resolution images into high-resolution images without additional training.
LoopLLM: Transferable Energy-Latency Attacks in LLMs via Repetitive Generation
Xingyu Li (National Interdisciplinary Research Center of Engineering Physics), Jia-Li Yin (Fuzhou University)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes LoopLLM, an energy-latency attack framework that induces large language models (LLMs) to generate repetitive content to maximize energy consumption and latency.
LORETTA: A Low Resource Framework to Poison Continuous Time Dynamic Graphs
Himanshu Pal (International Institute of Information Technology), Charu Sharma (International Institute of Information Technology)
CodeAdversarial AttackGraph
π― What it does: Proposes LORETTA, a model-free gradient, low-resource continuous-time dynamic graph poisoning framework that first sparsifies important edges and then inserts adversarial edges using degree-preserving negative sampling;
Lost in Benchmarks? Rethinking Large Language Model Benchmarking with Item Response Theory
Hongli Zhou (Harbin Institute of Technology), Tiejun Zhao (Harbin Institute of Technology)
CodeTextBenchmark
π― What it does: This paper proposes the PSN-IRT framework based on a pseudo twin network to evaluate the performance of large language models on various benchmark tests and conduct systematic analysis across 11 mainstream LLM benchmarks.
Lost in Time? A Meta-Learning Framework for Time-Shift-Tolerant Physiological Signal Transformation
Qian Hong (Renmin University of China), Zijing Zeng (OPPO)
CodeMeta LearningConvolutional Neural NetworkTime SeriesBiomedical DataElectronic Health Records
π― What it does: Propose a meta-learning framework called ShiftSyncNet that uses a time-shift correction network to perform waveform transformation on asynchronous physiological signals.
π― What it does: Proposed a zeroth-order optimization method called LOREN for LLM fine-tuning, which can significantly reduce memory usage and accelerate convergence without using backpropagation.
π― What it does: Proposed the LPPG-RL framework for lexicographic multi-objective reinforcement learning (Lexicographic Multi-Objective RL) in continuous spaces, achieving strict satisfaction of priority constraints through gradient projection and subproblem exploration.
LSAP-PV: High-Fidelity Palm Vein Image Synthesis via Layered Spectral Absorption Projection-Guided Diffusion Model
Sheng Shang (Hefei University Of Technology), Wei Jia (Hefei University Of Technology)
CodeGenerationData SynthesisDiffusion modelImage
π― What it does: Propose a 3D vascular tree model based on multi-layer spectral absorption projection (LSAP) and combine it with a conditional diffusion model to generate high-fidelity finger and palm vein images.
LSHFed: Robust and Communication-Efficient Federated Learning with Locally-Sensitive Hashing Gradient Mapping
Guanjie Cheng (Zhejiang University), Shuiguang Deng (Zhejiang University)
CodeClassificationCompressionFederated LearningSafty and PrivacyImage
π― What it does: Propose a federated learning framework named LSHFed, which achieves malicious gradient detection and communication compression through LSH gradient mapping while ensuring gradient privacy.
LUCID: Learning-Enabled Uncertainty-Aware Certification of Stochastic Dynamical Systems
Ernesto Casablanca (Newcastle University), Sadegh Soudjani (Max Planck Institute for Software Systems)
CodeAutonomous DrivingOptimizationBenchmark
π― What it does: Propose the LUCID tool, which learns control barrier certificates from limited transfer data of black-box systems to quantify safety guarantees.
LUMIN: A Longitudinal Multi-modal Knowledge Decomposition Network for Predicting Breast Cancer Recurrence
Chunyao Lu (Netherlands Cancer Institute), Ritse Mann (Netherlands Cancer Institute)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkTransformerContrastive LearningMultimodalityTime SeriesBiomedical DataElectronic Health Records
π― What it does: Developed a longitudinal multi-modal knowledge decomposition network called LUMIN, utilizing follow-up mammograms and electronic health records (EHR) to predict breast cancer recurrence.
LungNoduleAgent: A Collaborative Multi-Agent System for Precision Diagnosis of Lung Nodules
Cheng Yang (Hangzhou Dianzi University), Ruiquan Ge (Chongqing University of Posts and Telecommunications)
CodeClassificationObject DetectionAgentic AIPrompt EngineeringMixture of ExpertsVision Language ModelImageTextComputed TomographyRetrieval-Augmented Generation
π― What it does: Propose a collaborative multi-agent system named LungNoduleAgent for precise diagnosis of lung nodules in lung CT scans, covering three modules: nodule localization, local report generation, and malignancy assessment.
π― What it does: Propose a lightweight remote sensing visual backbone called LWGANet, which uniformly models and optimizes spatial and channel redundancies in remote sensing images.
π― What it does: This paper proposes a metric based on unlabeled data called M-Loss, which quantifies the gap between parameter average merging and output average ensembling, and uses it to guide model merging and pruning;
M2FMoE: Multi-Resolution Multi-View Frequency Mixture-of-Experts for Extreme-Adaptive Time Series Forecasting
Yaohui Huang (Central South University), Ruipeng Dong (Central South University)
CodeMixture of ExpertsTime Series
π― What it does: Proposes an extreme event adaptive time series prediction model called M2FMoE, which captures differences between regular and extreme patterns through multi-resolution, multi-perspective frequency Mixture-of-Experts;
M3SR: Multi-Scale Multi-Perceptual Mamba for Efficient Spectral Reconstruction
Yuze Zhang, Victor C. M. Leung (Shenzhen University)
CodeRestorationAuto EncoderImageBenchmark
π― What it does: Propose a multi-scale, multi-perceptual Mamba architecture (M3SR), which extracts features simultaneously in the spatial, frequency, and spectral domains through the MPF (multi-perceptual fusion) module, and integrates the U-Net structure to achieve global-middle-local multi-scale information fusion for RGB image to hyperspectral image reconstruction.
M3UCD: A Multi-task Multimodal Metaphor Understanding Challenge Dataset for LLMs
Tianlong Zheng (Xinjiang Technical Institute of Physics and Chemistry, Chinese Academy of Sciences), Turghun Osman (Xinjiang Technical Institute of Physics and Chemistry, Chinese Academy of Sciences)
CodeClassificationTransformerLarge Language ModelSupervised Fine-TuningImageTextMultimodalityBenchmark
π― What it does: Constructed the largest-scale, fine-grained annotated multimodal metaphor understanding dataset M3UCD, containing 15,345 samples and 12 manually annotated attributes; simultaneously systematically evaluated LLMs' performance on multimodal tasks and proposed a unified multitask collaborative learning framework MCLF to enhance the metaphor understanding capability of multimodal LLMs.
π― What it does: This paper proposes MacPrompt, a cross-lingual black-box attack method targeting text-to-image models, capable of bypassing safety filters and generating NSFW or restricted images by constructing hybrid-language 'macaronic words.'
π― What it does: Propose the MACS framework, implementing a two-stage audio-image conversion method that first separates multi-source audio and then generates images.
MacVQA: Adaptive Memory Allocation and Global Noise Filtering for Continual Visual Question Answering
Zhifei Li (Hubei University), Bing Yang (Hubei University)
CodeTransformerVision Language ModelAuto EncoderMultimodality
π― What it does: Proposed the MacVQA framework for continual visual question answering tasks, aiming to simultaneously address knowledge retention, adaptation to new tasks, and visual noise interference issues.
MagicPaint: Operate Anything for Image Inpainting with Diffusion Model
Qinhong Yang (University of Science and Technology of China), Nenghai Yu (Beijing Electronic Science and Technology Institute)
CodeRestorationVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: Propose a unified diffusion model called MagicPaint, supporting three restoration tasks: object addition, removal, and unconditional restoration under text and image conditions;
π― What it does: Designed and implemented a new equivariant parameter-efficient fine-tuning method called Magnitude-Modulated Equivariant Adapter (MMEA), which fine-tunes equivariant graph neural networks based on spherical harmonics by using a lightweight scalar gate to modulate the amplitude of each multiplicity channel while preserving rotational equivariance.
π― What it does: Develop the MAISI-v2 3D medical image synthesis framework, employing rectified flow to achieve 33Γ sampling acceleration, and introduce region-specific contrast loss to enhance conditional consistency and detail quality.
CodeComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
π― What it does: Propose a new self-attention mechanism called SPAttention, which divides the attention distance spectrum into non-overlapping segments, allowing each attention head to focus on different distance ranges and thus eliminating redundant computations in multi-head attention.
CodeClassificationExplainability and InterpretabilityKnowledge DistillationText
π― What it does: This paper proposes ProtoSurE, a post-hoc explanation framework based on sentence-level prototypes, to provide interpretable and faithful reasoning processes for text classification in large language models.
MambaSeg: Harnessing Mamba for Accurate and Efficient Image-Event Semantic Segmentation
Fuqiang Gu (Chongqing University), Zhenliang Ni (Chongqing University)
CodeSegmentationTransformerMultimodalityBenchmark
π― What it does: Propose the MambaSeg dual-branch framework for multimodal semantic segmentation of RGB and event data, and design the dual-dimensional interaction module (DDIM) to finely fuse spatial and temporal information.
Managing Infinite Abstractions in Numeric Pattern Database Heuristics
Markus Fritzsche, Alexander Shleyfman (Technion Israel Institute Of Technology)
CodeOptimizationBenchmark
π― What it does: Proposed a numerical planning pattern database (PDB) heuristic method for infinite abstractions, with improvements in A* guided goal-oriented abstract exploration, fragmented abstract heuristic enhancements, and failure query backup strategies.
Manipulation Intention Understanding for Zero-Shot Composed Image Retrieval
Yuanmin Tang (Chinese Academy of Sciences), Qi Wu (Augusta University)
CodeRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityChain-of-Thought
π― What it does: Proposes an intent-centric image-text dataset and the De-MINDS framework to enhance user intent understanding in zero-shot compositional image retrieval.
Many Minds, One Path: LLM-Augmented Consensus Decision for Distributed Control in Multi-Agent Collaborative Stable Scenarios
Zhuohao Yu (University of Chinese Academy of Sciences), Qing Wang (University of Chinese Academy of Sciences)
CodeExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt Engineering
π― What it does: Propose the LLMASC framework, utilizing a semantic-aware encoder, an LLM-driven consensus decision module, and a policy execution controller to achieve long-term stable control in distributed multi-agent systems.
π― What it does: This paper proposes a Multi-Activation Plane Interaction Graph Neural Network (MAPI-GNN), which dynamically constructs multi-graphs for each patient using a multi-dimensional feature discriminator and fuses them in two hierarchical levels to achieve multi-modal medical diagnosis.
MAPS: Multi-Agent Personality Shaping for Collaborative Reasoning
Jian Zhang (Xi'an Jiaotong University), Jun Liu (Nanyang Technological University)
CodeTransformerLarge Language ModelAgentic AIMultimodalityRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This study proposes a multi-agent collaboration framework called MAPS based on the Big Five personality theory for multimodal complex reasoning.
MARE: Multimodal Analogical Reasoning for Disease Evolution-Aware Radiology Report Generation
Qingqing Gao (Beijing University of Technology), Zhaohui Liu (Beijing University of Technology)
CodeGenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBiomedical DataChain-of-Thought
π― What it does: Propose an end-to-end multimodal analogy reasoning framework called MARE for generating radiology reports based on the longitudinal evolution of medical imaging.
Margin-Aware Preference Optimization for Aligning Diffusion Models Without Reference
Jiwoo Hong (KAIST AI), Jongheon Jeong (Theia Insights)
CodeGenerationReinforcement Learning from Human FeedbackSupervised Fine-TuningDiffusion modelImageText
π― What it does: This paper proposes a reference-free text-to-image diffusion model alignment method (MaPO), which directly learns model generation preferences by performing marginal optimization based on the Bradley-Terry model on the likelihood gap between selected and rejected samples.
MARPO: A Reflective Policy Optimization for Multi-Agent Reinforcement Learning
Cuiling Wu (Beijing Institute of Technology), Ying Fu (Beijing Institute of Technology)
CodeOptimizationReinforcement Learning
π― What it does: Proposed a Multi-Agent Reflective Policy Optimization (MARPO) framework that combines trajectory reflection mechanisms and KL-driven asymmetric clipping to improve sample efficiency and training stability.
π― What it does: Proposes a multi-agent, Socratic-style guided automated prompt optimization framework called MARS, which uses a Planner to generate task-specific optimization paths and refines prompts through a Teacher-Critic-Student three-party dialogue loop, with the final performance evaluated by the Target agent;
MartDE: A Privacy-Preserving and Cost-Efficient Evaluation Framework for Data Marketplaces
Xinyuan Qian (University of Electronic Science and Technology of China), Guowen Xu (University of Electronic Science and Technology of China)
CodeSafty and PrivacyComputational EfficiencyImage
π― What it does: This study proposes the MartDE framework, achieving efficient and fair model update transactions in data markets through privacy-preserving model evaluation.
Mask the Redundancy: Evolving Masking Representation Learning for Multivariate Time-Series Clustering
Zexi Tan (Guangdong University of Technology), Yiqun Zhang (Guangdong University of Technology)
CodeRepresentation LearningConvolutional Neural NetworkAuto EncoderContrastive LearningTime Series
π― What it does: Proposes the EMTC method, integrating dynamic masking with multi-view learning for unsupervised clustering of multivariate time series data.
π― What it does: Propose MaskAD, a parallel multi-branch masked autoencoder for multi-class unsupervised anomaly detection, which locates anomalies by leveraging reconstruction differences under different masks.
MASP: Multi-Aspect Guided Emotion Reasoning with Soft Prompt Tuning In Vision-Language Models
SangEun Lee (Electronics and Telecommunications Research Institute), Wonseok Chae (Electronics and Telecommunications Research Institute)
CodeClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
π― What it does: Propose the MASP framework, which extracts emotion-related visual features through a multi-perspective cross-attention module and achieves visual emotion recognition by optimizing the language model with soft prompts.
π― What it does: Designed a batch concept elimination method based on concept hierarchy, utilizing parent-child structures to perform group-level suppression on related sub-concepts, and proposed the SuPLoRA mechanism to preserve the generation capability of parent concepts during sub-concept elimination.
Massively Parallel Proof-Number Search for Impartial Games and Beyond
TomΓ‘Ε‘ ΔΓΕΎek (Charles University), Martin Schmid (Charles University)
CodeOptimization
π― What it does: Proposed a two-level parallel proof-number search algorithm (PNS-PDFPN), integrating Grundy numbers to reduce game trees, and implemented solutions for impartial games like Sprouts on a distributed cluster.
MathSE: Improving Multimodal Mathematical Reasoning via Self-Evolving Iterative Reflection and Reward-Guided Fine-Tuning
Jinhao Chen (Beihang University), Jie Tang (Beihang University)
CodeReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelMultimodalityChain-of-Thought
π― What it does: Propose a mathematical self-evolution framework named MathSE, which continuously enhances the mathematical reasoning ability of multi-modal large language models through multi-round supervised fine-tuning, reward-guided feedback, and self-reflection loops.
MathSmith: Towards Extremely Hard Mathematical Reasoning by Forging Synthetic Problems with a Reinforced Policy
Shaoxiong Zhan (Tsinghua University), Fei Tan (East China Normal University)
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
π― What it does: Proposes the MathSmith framework, which can randomly sample concept-explanation pairs from PlanetMath and autonomously generate high-difficulty math problems;
π― What it does: Studied the estimation of matrix twoβinfinity (β₯Β·β₯βββ) and oneβtwo (β₯Β·β₯βββ) norms in a matrix-free setting, proposing TwINEst and TwINEst++ algorithms using random Rademacher vectors and Hutchinson's diagonal estimation.
π― What it does: Proposes MAUGen, a diffusion-based multimodal framework capable of synthesizing realistic facial images with precise Action Unit (AU) labels from a single text prompt, across multiple identities.
MCMoE: Completing Missing Modalities with Mixture of Experts for Incomplete Multimodal Action Quality Assessment
Huangbiao Xu (Fuzhou University), Jinglin Xu (University of Science and Technology Beijing)
CodeRestorationTransformerMixture of ExpertsMultimodality
π― What it does: This paper proposes the MCMoE framework, which utilizes a hybrid expert and an adaptive gating modality generator to simultaneously learn unimodal and cross-modal features in a single-stage training, addressing the modality missing problem in action quality assessment.
MCP-AgentBench: Evaluating Real-World Language Agent Performance with MCP-Mediated Tools
Zikang Guo (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
CodeLarge Language ModelAgentic AITextBenchmarkChain-of-Thought
π― What it does: This paper constructs MCP-AgentBench, a language agent evaluation benchmark based on the Model Context Protocol (MCP), which includes 33 executable, stateless, text-based interactive MCP servers (with 188 tools in total) and 600 well-designed queries (categorized into six types based on single-machine/multi-machine and parallel/sequential invocation), and proposes MCP-Evalβa terminal task success rate evaluation framework based on LLM-as-a-judge.
MCPTox: A Benchmark for Tool Poisoning on Real-World MCP Servers
Zhiqiang Wang (University of Science and Technology of China), Xiangyang Li
CodeAdversarial AttackLarge Language ModelBenchmark
π― What it does: Constructed the MCPTox benchmark to systematically evaluate the robustness of LLM agents against tool poisoning attacks on real MCP servers, generating 1,348 malicious test cases;
MCTSr-Zero: Self-Reflective Psychological Counseling Dialogues Generation via Principles and Adaptive Exploration
Hao Lu (JianChengXingYun Technology Co., Ltd.), Chen Li (JianChengXingYun Technology Co., Ltd.)
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Introducing the self-reflective MCTS+LLM framework MCTSr-Zero for psychological counseling dialogue generation, which generates high-quality dialogues through mechanisms such as domain alignment, regeneration, and meta-prompt adaptation.
MdaIF: Robust One-Stop Multi-Degradation-Aware Image Fusion with Language-Driven Semantics
Jing Li (East China Normal University), Bin Yang (Central University of Finance and Economics)
CodeRestorationConvolutional Neural NetworkTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageMultimodality
π― What it does: Propose a multi-degradation-aware single-site image fusion framework (MdaIF) based on the semantic prior of a vision-language model (VLM), utilizing a Mixture of Experts (MoE) and a degradation-aware channel attention module (DCAM) to achieve infrared-visible image fusion under different weather degradation conditions such as fog, rain, and snow.
CodeClassificationRecurrent Neural NetworkTransformerLarge Language ModelTextMultimodalityBenchmarkAudio
π― What it does: Designed and implemented the MDF framework, which first decomposes audio into text and acoustic components to generate cross-modal heterogeneity, then enhances heterogeneous features using the CHE module, and completes sentiment analysis by adaptively fusing multimodal information through MAW.
π― What it does: Proposed and implemented a scene text recognition framework based on the Mask Diffusion Model (MDiff4STR), which balances recognition accuracy and inference speed.
MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models
Pengfei Zhou (National University of Singapore), Kaipeng Zhang (Shanghai Innovation Institute)
CodeTransformerLarge Language ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: Propose MDK12-Bench, a large-scale multimodal evaluation benchmark based on real K-12 exams, comprising 141K instances across six subjects and a six-level knowledge hierarchy.
MDMLP-EIA: Multi-domain Dynamic MLPs with Energy Invariant Attention for Time Series Forecasting
Hu Zhang (Changsha University), Yongfang Xie (Central South University)
CodeTime SeriesBenchmark
π― What it does: A time series forecasting framework named MDMLP-EIA based on multi-domain dynamic MLP and energy-invariant attention is constructed.
π― What it does: Proposes the MDND framework, integrating non-differentiable iterative optimization with deep functional mapping to achieve unsupervised learning.
π― What it does: Propose a source-agnostic domain adaptation method ME-SFDA based on a pyramidal Atkinson-Shiffrin memory model, achieving effective transfer to the target domain through two-step splitting, memory fusion, and adversarial clustering.
Measuring the Unmeasurable: Unveiling Latent Cognitive Capabilities of LLM
Cui Danxin (Fudan University), Yilun Liu (Huawei)
CodeLarge Language ModelTextBenchmarkChain-of-Thought
π― What it does: Constructed a hierarchical classification of cognitive abilities based on the ACT-R cognitive architecture, and introduced the multilingual CogProbe benchmark and corresponding CogEval dataset to fine-grained evaluate LLMs' performance on 16 cognitive operations.
π― What it does: Studies the automation of mechanical mechanism design, proposing the MechaFormer model, which transforms the mechanism synthesis problem into a conditional sequence generation task. The model generates DSL strings describing mechanism topology and geometric parameters using a Transformer, achieving one-time complete design;
Mechanistic Dissection of Cross-Attention Subspaces in Text-to-Image Diffusion Models
Jun-Hyun Bae (Kyungpook National University), Heechul Jung (Kyungpook National University)
CodeGenerationExplainability and InterpretabilityComputational EfficiencyTransformerDiffusion modelImageText
π― What it does: This paper performs singular value decomposition (SVD) on the cross-attention output-value (OV) circuit in text-to-image diffusion models, revealing that semantic concepts are encoded in low-dimensional spectral subspaces, and verifies their functionality by intervening in these subspaces during the generation process.
MedGR2: Breaking the Data Barrier for Medical Reasoning via Generative Reward Learning
Weihai Zhi (Guangdong Institute of Intelligence Science and Technology), Shangyang Li (Guangdong Institute of Intelligence Science and Technology)
CodeData SynthesisData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelMultimodalityBiomedical Data
π― What it does: MedGR 2 automatically generates high-quality medical VLM training data through a self-improving generative-reward learning framework, and then uses reinforcement learning to enhance cross-modal and cross-task reasoning based on this data.
MedLA: A Logic-Driven Multi-Agent Framework for Complex Medical Reasoning with Large Language Models
Siqi Ma (Westlake University), Zelin Zang (CAIR, Hong Kong Institute of Science and Innovation, Chinese Academy of Sciences)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBiomedical DataBenchmark
π― What it does: Proposed a multi-agent framework called MEDLA based on a logical tree, which can decompose medical question-answering into three-step reasoning (major premise, minor premise, conclusion), and let agents iteratively refine the reasoning tree through multi-round graph-guided discussions, ultimately achieving consistent and traceable diagnostic conclusions.
MedMKEB: A Comprehensive Knowledge Editing Benchmark for Medical Multimodal Large Language Models
Dexuan Xu (Peking University), Yu Huang (China Academy of Chinese Medical Sciences)
CodeAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBiomedical DataBenchmark
π― What it does: This paper proposes and implements MedMKEBβa comprehensive benchmark for knowledge editing in medical multimodal large language models (MLLMs), covering five dimensions: reliability, locality, generalizability, transferability, and robustness.
MedSΒ³: Towards Medical Slow Thinking with Self-Evolved Soft Dual-sided Process Supervision
Shuyang Jiang (Fudan University), Yu Wang (Shanghai Jiao Tong University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataBenchmarkChain-of-Thought
π― What it does: Proposed MedS 3, a self-evolving slow thinking framework that enables small medical language models to generate reliable step-by-step reasoning paths through MCTS, thereby self-improving and ultimately achieving high-performance clinical reasoning.
MedSAMix: A Training-Free Model Merging Approach for Medical Image Segmentation
Yanwu Yang (University Hospital Tubingen), Thomas Wolfers (University Hospital Tubingen)
CodeSegmentationHyperparameter SearchTransformerImageBiomedical Data
π― What it does: A training-agnostic model merging framework called MedSAMix is proposed for medical image segmentation, which can enhance performance by automatically searching for merging configurations to combine the general foundation model SAM with specialized models MedSAM and MedicoSAM;
π― What it does: Propose Medverse, a general context learning (ICL) framework capable of performing 3D medical image segmentation, transformation, and enhancement tasks within the same model.
Mem-PAL: Towards Memory-based Personalized Dialogue Assistants for Long-term User-Agent Interaction
Zhaopei Huang (Renmin University of China), Qin Jin (Renmin University of China)
CodeGenerationData SynthesisTransformerLarge Language ModelTextGraphBenchmarkRetrieval-Augmented Generation
π― What it does: This paper constructs a Chinese long-term user-assistant interaction dataset PAL-Set and benchmark PAL-Bench, and proposes a hierarchical heterogeneous memory framework H Memory 2 to enhance personalized dialogue generation.
Membership Inference Attack Against Large Language Model-Based Recommendation Systems: A New Distillation-Based Paradigm
Cuihong Li, Jitao Sang (Beijing Jiaotong University)
CodeRecommendation SystemKnowledge DistillationAdversarial AttackTransformerLarge Language ModelText
π― What it does: Studied a membership inference attack against LLM-based recommendation systems and proposed a novel attack paradigm based on knowledge distillation.
MemeBQ:Memory Efficient Binary Quantization of LLMs
Yuanhui Wang (Sanya Nanhai Innovation and Development Base of Harbin Engineering University), Qinghao Hu (Institute of Automation Chinese Academy of Sciences)
CodeComputational EfficiencyTransformerText
π― What it does: Propose a binary post-training quantization framework called MemeBQ, which reduces additional bitmap memory and improves quantization quality by leveraging row similarity clustering to share bitmaps and performing fine-grained segmentation within each group via k-means.
MemGuide: Intent-Driven Memory Selection for Goal-Oriented Multi-Session LLM Agents
Yiming Du (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the MemGuide framework, achieving efficient dialogue in multi-session task-oriented conversations through intent-driven memory selection.
MemoryART: Enhancing LLMs via Multi-Memory Models with Adaptive Resonance Theory for Healthcare Agents
Renke Dai (South Central Minzu University), Ah-Hwee Tan (Singapore Management University)
CodeTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose MemoryART, a multi-memory framework based on Adaptive Resonance Theory (ART), to enhance long-term memory and retrieval capabilities of large language models (LLMs) in multi-turn medical dialogues.
Mental Model-based Generation of Lies for Insider Threat Modeling
Brittany Cates (Colorado State University), Sarath Sreedharan (Colorado State University)
CodeAdversarial AttackBenchmark
π― What it does: Studied deception mechanisms in insider threat attacks, proposing a framework that generates lies and plans through model reconstruction and planning search, enabling attackers to complete hidden objectives while maintaining the supervisor's belief that they are performing expected tasks.
π― What it does: Proposed a grid-cell-based search algorithm called MeshA*, which solves path planning problems with motion primitive constraints by searching on the grid while recording motion primitives passing through cells.
Yiqing Zou (Beijing Institute of Technology), Sijie Ruan (Beijing Institute of Technology)
CodeRecurrent Neural NetworkGraph Neural NetworkTime Series
π― What it does: Propose the MetaDG framework, which utilizes dynamic node embeddings to generate time-varying graph structures for unified spatiotemporal modeling of traffic flow.
π― What it does: Propose a meta-learning-based cross-modal hashing framework called MGSH, which enhances retrieval performance in the presence of label noise.
MetaAct-RL: Training Language Models for Reasoning Through Meta-Action-Based Reinforcement Learning
Zhiheng Xi (Fudan University), Xuanjing Huang (Fudan University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
π― What it does: Propose the MetaAct-RL framework, which enables language models to self-select and execute high-level meta-actions (forward reasoning, criticism, and refinement) during inference, optimized via reinforcement learning; simultaneously constructs diverse meta-action trajectories as SFT data and introduces length rewards and key-state restart mechanisms;
MetaDiT: Enabling Fine-grained Constraints in High-degree-of Freedom Metasurface Design
Hao Li (Harbin Engineering University), Andrey Bogdanov (Harbin Engineering University)
CodeGenerationOptimizationTransformerDiffusion modelContrastive LearningImageSequentialPhysics Related
π― What it does: Designed a generative framework named MetaDiT to simultaneously optimize all structural parameters in high-freedom metasurfaces while precisely satisfying given high-resolution optical spectrum constraints.