AAAI Conference on Artificial Intelligence · 2140 papers
JoDiffusion: Jointly Diffusing Image with Pixel-Level Annotations for Semantic Segmentation Promotion
Haoyu Wang (Northwestern Polytechnical University), Chen Ding (Northwestern Polytechnical University)
CodeSegmentationGenerationVision Language ModelDiffusion modelAuto EncoderImageText
🎯 What it does: Proposes JoDiffusion, a joint diffusion model capable of simultaneously generating images and corresponding pixel-level segmentation annotations.
Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models
Changyue Wang (Tsinghua University), Yiqun Liu (Tsinghua University)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed a black-box hallucination detection framework named RACE, specifically designed for large reasoning models (LRM), which detects factual hallucinations by jointly evaluating the consistency between the model's reasoning trajectory and the final answer.
Joint-GCG: Unified Gradient-Based Poisoning Attacks on Retrieval-Augmented Generation Systems
Haowei Wang (State Key Laboratory of Complex System Modeling and Simulation Technology), Qing Wang (State Key Laboratory of Complex System Modeling and Simulation Technology)
🎯 What it does: Propose the Joint-GCG framework to perform poisoning attacks on RAG systems by jointly manipulating gradients of the retriever and generator.
Judge Q: Trainable Queries for Optimized Information Retention in KV Cache Eviction
Yijun Liu (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
CodeOptimizationComputational EfficiencyLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose Judge Q, which inserts learnable soft tokens during the pre-filling phase and only fine-tunes the embedding layer, allowing these soft tokens to fit the attention maps of real decoding tokens, thereby better capturing global information during KV cache pruning;
Jump-teaching: Combating Sample Selection Bias via Temporal Disagreement
Kangye Ji (Xidian University), Bohu Huang (Xidian University)
CodeClassificationData-Centric LearningImage
🎯 What it does: Propose an efficient sample selection framework named Jump-teaching, which eliminates sample selection bias by leveraging temporal differences of a single network across different training iterations, and achieves finer-grained sample screening through single-sample loss splitting.
Jupiter: Enhancing LLM Data Analysis Capabilities via Notebook and Inference-Time Value-Guided Search
Shuocheng Li (Peking University), Dongmei Zhang (Microsoft)
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextTabular
🎯 What it does: This paper proposes the NbQA dataset and the JUPITER framework, aiming to enhance the performance of large language models (LLMs) in multi-step data analysis tasks.
K-12EduBench: A Benchmark for Evaluating Large Language Models’ Knowledge, Problem-Solving, and Educational Goal Cognition in K-12 Education
Yuqing Ye (Northeast Normal University), Dongdai Zhou (Cornell University)
CodeLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the K-12EduBench benchmark to evaluate three core capabilities of LLMs in K-12 education: knowledge mastery, interdisciplinary problem-solving, and educational goal cognition;
K-ProtoDiff: Key Prototypes-Guided Diffusion for Time Series Generation
Yuhang Duan (Dalian University of Technology), Xiaoshuai Wu (Dalian University of Technology)
CodeGenerationTransformerDiffusion modelTime Series
🎯 What it does: Propose a diffusion model based on key prototypes (K-ProtoDiff) for time series generation, which can maintain the global distribution while significantly preserving local key patterns.
🎯 What it does: Propose KALL-E, an autoregressive text-to-speech model based on Flow-VAE continuous latent representations, directly predicting the next-frame speech distribution.
KCLNet: Electrically Equivalence-Oriented Graph Representation Learning for Analog Circuits
Peng Xu (Chinese University of Hong Kong), Bei Yu (Chinese University of Hong Kong)
CodeClassificationObject DetectionRepresentation LearningGraph Neural NetworkContrastive LearningGraphPhysics Related
🎯 What it does: Proposes the KCLNet framework, which utilizes the current equivalence principle based on KCL and asynchronous current-oriented GNN to perform graph representation learning on analog circuits;
🎯 What it does: Propose the KeenKT model, which represents students' knowledge states as Normal-Inverse-Gaussian (NIG) distributions, enabling the capture of fluctuations in learning behaviors during each interaction and achieving distributed representation of states.
Yuxuan Tian (Peking University), Tong Yang (ByteDance)
CodeCompressionComputational EfficiencyTransformerLarge Language Model
🎯 What it does: Proposes the KeepKV method, which periodically compresses the KV cache of LLMs without sacrificing generation quality, addressing the attention inconsistency problem caused by traditional merging.
Key Decision-Makers in Multi-Agent Debates: Who Holds the Power?
Qian Zhang (Tianjin University), Lanjun Wang (Tianjin University)
CodeLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper systematically investigates the impact of role allocation strategies in multi-agent debate (MAD) on reasoning task performance, and proposes the MADC consistency ranking strategy based on path consistency without requiring prior knowledge of the truth.
🎯 What it does: Designed a graph contrastive learning framework called Khan-GCL based on the Kolmogorov-Arnold network (KAN), and proposed a method for critical feature identification and hard negative sample generation using KAN coefficients
🎯 What it does: Propose a direct alignment recovery method based on k-nearest neighbors (KNNDA), which simultaneously accomplishes alignment recovery and consistent representation learning in partially viewed multi-view clustering (PVC) without requiring pre-training.
Knowledge Graph Guided Heterogeneity-Informed Diffusion Model for Spatio-Temporal Generation
Zi'ang Wang, Yu Zhao (Beihang University)
CodeGenerationData SynthesisMixture of ExpertsDiffusion modelGraphTime Series
🎯 What it does: This paper proposes a knowledge graph-based heterogeneous-aware diffusion model (KGDiff) for generating urban spatiotemporal data with spatial structure and temporal heterogeneity.
Knowledge-Enhanced Explainable Prompting for Vision-Language Models
Yequan Bie (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
CodeExplainability and InterpretabilityLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Designed and implemented a framework called KEEP to enhance prompt learning of visual-language models (e.g., CLIP) with fine-grained domain knowledge, providing both visual and textual interpretability.
KSS-MoE: Knowledge Space Synergy Framework in Mixture of Experts for Continual Visual Instruction Tuning
Lingyun Song (Northwestern Polytechnical University), Xuequn Shang (Northwestern Polytechnical University)
CodeMixture of ExpertsMultimodalityBenchmark
🎯 What it does: Studied a novel knowledge space collaborative framework KSS-MoE, using Mixture of Experts in continuous visual instruction tuning to alleviate catastrophic forgetting.
KTV: Keyframes and Key Tokens Selection for Efficient Training-Free Video LLMs
Baiyang Song (Xiamen University), Jianyuan Guo (Xiamen University)
CodeComputational EfficiencyRepresentation LearningTransformerVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: Designed and implemented the KTV framework, which leverages an untrained vision-language model for video understanding. It first selects keyframes through clustering, then extracts key visual tokens from each frame to reduce spatiotemporal redundancy and improve inference efficiency.
KVmix: Gradient-Based Layer Importance-Aware Mixed-Precision Quantization for KV Cache
Fei Li (Xi'an Jiaotong University), Jinyu Wang (Xi'an Jiaotong University)
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: To address the excessive memory consumption of KV cache in large language models, the KVmix method is proposed, which leverages gradient importance analysis to achieve hierarchical mixed-precision quantization, combined with dynamic critical context selection and efficient CUDA kernel fusion, significantly compressing the KV cache and enhancing inference throughput.
LaF-GRPO: In-Situ Navigation Instruction Generation for the Visually Impaired via GRPO with LLM-as-Follower Reward
Yi Zhao (Hong Kong Polytechnic University), Jing Li (Hong Kong Polytechnic University)
CodeAutonomous DrivingReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the LaF-GRPO framework, which utilizes LLM to simulate responses of visually impaired users as a reward for post-training Vision-Language Models, generating precise, context-aware step-by-step navigation instructions, and constructing the NIG4VI dataset with 27k samples.
🎯 What it does: Propose the Lambda two-phase HPO framework, which first learns a reliable prior in low-fidelity (LF) tasks and then uses this prior to guide high-fidelity (HF) search.
CodeClassificationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the LAMDAS method, which utilizes a pre-trained LLM as an implicit classifier to select domain-specific data, addressing the challenges of scarce high-quality data and noise in large volumes of unverified data.
LAMIC: Layout-Aware Multi-Image Composition via Scalability of Multimodal Diffusion Transformer
Yuzhuo Chen, Weiming Zhang (University Of Science And Technology Of China)
CodeGenerationTransformerVision Language ModelDiffusion modelAuto EncoderImageTextMultimodality
🎯 What it does: Propose a zero-training multi-graph layout-aware synthesis framework named LAMIC, achieving joint generation of multiple reference images and spatial layouts.
🎯 What it does: Propose a visual Transformer PTQ method based on hierarchical mixed-precision quantization (LAMPQ), which can assign different bit-widths to each layer, significantly improving the accuracy of low-precision models
Landsat30-AU: A Vision-Language Dataset for Australian Landsat Imagery
Sai Ma (Australian National University), John A. Taylor (Australian National University)
CodeLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the Landsat30-AU audio-visual language dataset (with two subsets: CAP and VQA), covering four satellites (Landsat 5/7/8/9), spanning 36 years, and with 30-meter resolution, and conducted benchmark evaluations on existing Vision-Language Models (VLMs).
LangGPS: Language Separability Guided Data Pre-Selection for Joint Multilingual Instruction Tuning
Yangfan Ye (Harbin Institute of Technology), Bing Qin (Harbin Institute of Technology)
CodeData-Centric LearningText
🎯 What it does: Propose a two-stage lightweight data pre-selection framework called LangGPS, which uses language separability as a guidance signal to filter multilingual instruction-tuning data;
🎯 What it does: Investigated the language drift phenomenon in multilingual retrieval-augmented generation (RAG) and proposed a lightweight decoding control method called Soft Constrained Decoding (SCD) to alleviate drift.
🎯 What it does: Proposed a Large-Scale Brain Connectivity Model (LCM) using a decoder-only Transformer architecture, pre-trained on a large number of functional connectivity matrices (fMRI) through multi-task learning, and fine-tuned semi-supervisedly to achieve various clinical and behavioral prediction tasks.
Large Language Models Meet Extreme Multi-label Classification: Scaling and Multi-modal Framework
Diego Ortego (NielsenIQ), Juan C. SanMiguel (NielsenIQ)
CodeClassificationTransformerLarge Language ModelContrastive LearningTextMultimodality
🎯 What it does: In the extreme multi-label classification (XMC) task, this paper investigates how to effectively utilize large decoder language models and visual metadata, and proposes a multimodal framework called ViXML;
LAS: Loss-less ANN-SNN Conversion for Fully Spike-Driven Large Language Models
Long Chen (Sichuan University), Yanan Sun (Sichuan University)
CodeComputational EfficiencySpiking Neural NetworkTransformerLarge Language ModelTextMultimodality
🎯 What it does: This paper proposes the LAS framework, which losslessly converts pre-trained large language models into fully spiking neural networks, achieving inference with lower energy consumption.
Latent Self-Consistency for Reliable Majority-Set Selection in Short- and Long-Answer Reasoning
Jungsuk Oh (Seoul National University), Jay-Yoon Lee (Seoul National University)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the Latent Self-Consistency (LSC) method, which appends learnable summary tokens after generation by large language models, utilizing contrastive learning to obtain semantic embeddings, thereby achieving unified consistency selection for both short and long answers.
LatentLLM: Activation-Aware Transform to Multi-Head Latent Attention
Toshiaki Koike-Akino (Mitsubishi Electric Research Laboratories), Matthew Brand
CodeCompressionTransformerTextMultimodality
🎯 What it does: Perform training-free compression on pre-trained large language models and multi-modal models, converting them into a low-dimensional multi-head latent attention (MLA) structure.
LaTeX2Layout: High-Fidelity, Scalable Document Layout Annotation Pipeline for Layout Detection
Feijiang Han (University of Pennsylvania), Lyle Ungar (University of Pennsylvania)
CodeObject DetectionData SynthesisSupervised Fine-TuningVision Language ModelImage
🎯 What it does: Built a pipeline that directly generates pixel-level layout annotations using a LaTeX compiler, and fine-tuned a general-purpose vision-language model by programatically generating synthetic data to accomplish document layout parsing.
🎯 What it does: LayerEdit achieves untrained multi-object text-driven image editing through multi-layer decomposition, conflict-aware editing, and transparency-guided fusion.
Laytrol: Preserving Pretrained Knowledge in Layout Control for Multimodal Diffusion Transformers
Sida Huang (Northwestern Polytechnical University), Hongyuan Zhang (University Of Hong Kong)
CodeGenerationData SynthesisTransformerPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposed the Laytrol network for layout-to-image generation, leveraging parameter copying to retain pre-trained knowledge, constructed the LaySyn dataset, and designed specialized initialization, object-level RoPE, and random prompt dropping mechanisms.
Learn from Global Correlations: Enhancing Evolutionary Algorithm via Spectral GNN
Kaichen Ouyang, Dayu Hu (Sun Yat-sen University)
CodeOptimizationGraph Neural NetworkBenchmark
🎯 What it does: Proposed a evolutionary algorithm framework GNE based on spectral graph neural networks, modeling the population as a graph and utilizing spectral filtering to update individuals, thereby achieving global information learning.
Learner-Tailored Program Repair: A Solution Generator with Iterative Edit-Driven Retrieval Enhancement
Zhenlong Dai (Zhejiang University), Jingyuan Chen (Zhejiang University)
CodeAI Code AssistantTransformerLarge Language ModelTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the Learner-Customized Program Repair Task (LPR) and designed the LSGEN framework, which generates repair code and corresponding bug descriptions by leveraging retrieval databases, edit-based retrieval, differential analysis, and iterative retrieval enhancement;
Learning Adaptive and Expandable Mixture Model for Continual Learning
Fei Ye (University of Electronic Science and Technology of China), ShiJie Zhou (University of Electronic Science and Technology of China)
CodeClassificationDomain AdaptationTransformerMixture of ExpertsImageBenchmark
🎯 What it does: In the multi-domain task incremental learning scenario, we propose a continuous learning framework based on pre-trained models, which includes dual representation backbone networks, an expandable mixture of experts module, and an adaptive fusion with dynamic knowledge calibration mechanism.
🎯 What it does: In semi-supervised semantic segmentation, a pseudo-label refinement framework called ViLaDiff is constructed by generating image descriptions, fusing them with visual features, and introducing mixed noise diffusion in the label space.
🎯 What it does: Proposes the Tree-Gate branching strategy TGPPO based on PPO, which directly learns variable selection during the Branch-and-Bound process.
Learning Cell-Aware Hierarchical Multi-Modal Representations for Robust Molecular Modeling
Mengran Li (Sun Yat-sen University), Stan Z. Li (Sun Yat-sen University)
CodeRepresentation LearningDrug DiscoveryGraph Neural NetworkContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Propose a multi-modal representation framework (CHMR) that jointly models molecular structure, cell phenotypes, and gene expression to address missing cell modalities and hierarchical dependency issues.
🎯 What it does: Investigated a dual-branch neural SDF network that utilizes a shared compact latent space and spatial voxel grid features, enabling the representation of multiple 3D shapes on a single latent code, while enhancing high-frequency detail reconstruction through a balanced sampling strategy.
🎯 What it does: Utilizes deep learning to predict conjugate direction fields (CDF) on freeform surfaces and directly generates initial layouts suitable for constructing planar quadrilateral meshes, avoiding the traditional costly nonlinear optimization process.
🎯 What it does: This study proposes a new perspective based on word counting for learning deterministic finite automata (DFA) using only positive examples, and provides the corresponding NP-completeness proof; meanwhile, it designs an integer linear programming (ILP) solver and a heuristic preprocessing algorithm;
🎯 What it does: Proposed a closed-loop self-regulating framework AEED, which uses entropy flow to monitor learning progress and dynamically adjust class weights, thereby improving the performance of long-tailed action recognition.
🎯 What it does: Propose a SiGMoID framework based on HyperPINN and Wasserstein GAN for simultaneously estimating system parameters, quantifying noise, and reconstructing unobserved system states from noise-sparse, partially observable data.
CodeTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Designed and implemented the LEAP system, which leverages students' long-term learning processes for adaptive tutoring planning, and constructed the LEAD dataset based on real student multi-round submissions.
Learning from Scoring Disagreements: Contrastive Error Mining for Efficient and Robust LLM-based Assessment
Lei Chen (Jinan University), Weiqi Luo (Jinan University)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose the CEM-FT framework, which automatically identifies high-value hard samples where the score differences between fully fine-tuned models and few-shot models are significant, and applies lightweight LoRA fine-tuning on these samples to improve the accuracy and consistency of LLM automatic scoring.
Learning from the Undesirable: Robust Adaptation of Language Models Without Forgetting
Yunhun Nam (Korea University), Jongheon Jeong (Korea University)
CodeDomain AdaptationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a regularization method called Learning-from-the-Undesirable (LfU) for supervised fine-tuning of large language models under limited data conditions, significantly reducing overfitting while retaining pre-trained knowledge.
🎯 What it does: This paper proposes a heuristic function based on graph neural networks (GNN) for solving numerical planning problems, evaluated on multiple IPC 2023 numerical planning benchmarks.
Learning Latent Imaging Biomarkers for Interpretable Microvascular Invasion Prediction in Hepatocellular Carcinoma
Ji Rao (Tongji University), Ye Luo (Tongji University)
CodeClassificationExplainability and InterpretabilityTransformerContrastive LearningMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose a two-stage interpretable framework IRCL, which first learns potential image features through dual-layer contrastive learning and clusters them to generate imaging biomarkers, then predicts microvascular invasion (MVI) by aligning these markers with patient features in the original images, and achieves spatial interpretation of image markers through a learnable mask.
🎯 What it does: Propose a network decomposition method called MIND based on graph neural networks and reinforcement learning, completely independent of manual features;
🎯 What it does: Proposes a method for training neural operators under partially observed data, incorporating mask prediction training strategies and a physics-aware latent autoregressive propagation module.
🎯 What it does: Propose a real-time personality recognition framework based on audio-visual behavior simulation of individual internal cognition, achieving regression of real personality traits by analyzing cognitive graphs using a personalized network weight generator and 2D graph neural network.
Learning Procedural-Aware Video Representations Through State-Grounded Hierarchy Unfolding
Jinghan Zhao (Beihang University), Feng Lu (Beihang University)
CodeRepresentation LearningConvolutional Neural NetworkLarge Language ModelVideoTextMultimodality
🎯 What it does: Propose a Task-Step-State (TSS) three-layer framework, integrating a visual state layer into the task-step hierarchy to address the semantic gap between abstract descriptions and visual data, and design a progressive pre-training strategy to gradually unfold this hierarchical structure;
Learning Protein–Ligand Binding in Hyperbolic Space
Jianhui Wang (Tsinghua University), Yanyan Lan (Tsinghua University)
CodeDrug DiscoveryGraph Neural NetworkTransformerContrastive LearningBiomedical Data
🎯 What it does: Propose HypSeek, a three-tower architecture that embeds ligands, protein pockets, and protein sequences into the hyperbolic space of the Lorentz model, to unify virtual screening and affinity ranking.
Learning Subgroups with Maximum Treatment Effects Without Causal Heuristics
Lincen Yang (Leiden University), Saber Salehkaleybar (Leiden University)
CodeExplainability and InterpretabilitySupervised Fine-TuningTabular
🎯 What it does: Under the structural causal model framework, the problem of identifying subgroups with maximum average treatment effect is transformed into a standard supervised learning problem, and subgroup discovery is achieved through CART trees;
🎯 What it does: Investigate the problem of expanding learning systems, proposing the HaT framework to achieve efficient heterogeneous perception knowledge transfer.
Xi Ding (Griffith University), Yongsheng Gao (Griffith University)
CodeClassificationAnomaly DetectionTransformerImageVideoTime Series
🎯 What it does: Propose a SEQ learning framework that aligns predicted sequences with class-specific time prototypes using soft DTW, enabling a standard feedforward classifier to possess temporal reasoning capabilities.
Learning to Collaborate: An Orchestrated-Decentralized Framework for Peer-to-Peer LLM Federation
Inderjeet Singh (Fujitsu Research of Europe), Motoyoshi Sekiya (Fujitsu Research of Europe)
CodeFederated LearningSafty and PrivacyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a decentralized LLM federated learning framework KNEXA-FL, which assigns optimal peer-to-peer knowledge distillation tasks to heterogeneous LLM agents through a centralized matcher, enabling multi-institutional data collaboration without sharing raw data.
🎯 What it does: Proposes a dual-agent reinforcement learning framework called Cutter for compressing large-scale graph structures while preserving graph robustness features.
Learning to Generate and Extract: A Multi-Agent Collaboration Framework for Zero-Shot Document-Level Event Arguments Extraction
Guangjun Zhang, Ru Li (Shanxi University)
CodeData SynthesisTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Propose a multi-agent collaborative framework that iteratively generates high-quality synthetic document-level event argumentation data through the interaction between generation agents and evaluation agents, thereby improving the performance of zero-shot event argument extraction.
Learning to Tell Apart: Weakly Supervised Video Anomaly Detection via Disentangled Semantic Alignment
Wenti Yin (Huazhong University of Science and Technology), Nong Sang (Huazhong University of Science and Technology)
CodeAnomaly DetectionGraph Neural NetworkTransformerVision Language ModelContrastive LearningVideo
🎯 What it does: This paper proposes the DSANet framework, which performs coarse-to-fine granularity detection and classification of video anomalies in a weakly supervised manner;
Learning Topology-Aware Dynamic Associations for Robust Multi-Person Pose Estimation
Shengnan Hu (Central China Normal University), Yahong Chen (Central China Normal University)
CodePose EstimationTransformerImage
🎯 What it does: To address occlusion, scale variation, and complex interactions in multi-person human pose estimation, the authors propose the TopoDA framework.
Learning Topology-Driven Multi-Subspace Fusion for Grassmannian Deep Networks
Xuan Yu (Jiangnan University), Tianyang Xu (Jiangnan University)
CodeClassificationRecognitionVideoGraphBiomedical Data
🎯 What it does: Proposed a topology-driven multi-subspace fusion framework aimed at capturing complex geometric structures by dynamically selecting and weighting task-related subspaces.
Learning Underwater Image Enhancement Iteratively Without Reference Images
Yi Tang (Kitami Institute of Technology), Hiroshi Masui (Kitami Institute of Technology)
CodeRestorationTransformerVision Language ModelDiffusion modelImage
🎯 What it does: This paper proposes an unsupervised iterative diffusion model framework for underwater image enhancement, decomposing the task into colorization and color compensation, and enhancing warm color information through a quantization mechanism.
Learning Vision-Based Neural Network Controllers with Semi-Probabilistic Safety Guarantees
Xinhang Ma (Washington University in St. Louis), Yevgeniy Vorobeychik (Washington University in St. Louis)
CodeAutonomous DrivingSafty and PrivacyGenerative Adversarial NetworkImage
🎯 What it does: Proposes a semi-probabilistic safety verification framework (SPV) that combines reachability analysis with conditional generative networks and distribution-agnostic tail bounds to achieve scalable verification and training for visual controllers;
🎯 What it does: Proposes a buffer-free continual multi-task learning framework LwP, which avoids catastrophic forgetting by preserving the geometric structure of the shared feature space.
🎯 What it does: Proposed the Length-Adaptive Interest Network (LAIN), which explicitly incorporates sequence length information into CTR prediction models to balance long- and short-sequence modeling, significantly enhancing CTR prediction performance.
LENS: Learning to Segment Anything with Unified Reinforced Reasoning
Lianghui Zhu (Huazhong University of Science & Technology), Xinggang Wang (vivo Mobile Communication Co., Ltd)
CodeSegmentationTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose the LENS framework to achieve segmentation tasks under text prompts, using a unified reinforcement learning strategy for chained reasoning and segmentation optimization during testing.
🎯 What it does: In symbolic music emotion recognition, this paper evaluates MIDIBERT's insufficient perception of musical modes (major/minor), proposes a mode-oriented diagnostic method MoGE, and designs the MoFi framework, which leverages FiLM to inject mode knowledge into the Transformer's lower layers to enhance emotion recognition performance.
Let’s Think with Images Efficiently! An Interleaved-Modal Chain-of-Thought Reasoning Framework with Dynamic and Precise Visual Thoughts
Xu Liu (Harbin Institute of Technology), Libo Qin (Central South University)
CodeSegmentationComputational EfficiencyLarge Language ModelVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Propose a new Interleaved-Modal Chain-of-Thought framework called DAP-ICOT, which achieves more efficient multimodal reasoning through dynamic visual thinking integration and precise visual thinking guidance.
Leveraging Failed Samples: A Few-Shot and Training-Free Framework for Generalized Deepfake Detection
Shibo Yao (Beijing Jiaotong University), Chunjie Zhang (Chinese Academy of Sciences)
CodeAnomaly DetectionTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Propose FTNet, a training-free framework that achieves deepfake detection with only a minimal number of samples (just one synthetic sample), leveraging CLIP intermediate layer features to construct a Key-Value cache and discriminates test samples via nearest neighbor classification;
Leveraging Visual Blur Perception Characteristics for EEG Decoding
Wenchao Liu (Harbin Institute of Technology), Haifeng Li (Harbin Institute of Technology)
CodeRetrievalRepresentation LearningConvolutional Neural NetworkContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Propose a visual decoding framework for EEG based on visual blur perception features, leveraging multi-level Gaussian blur and feature selection to construct personalized visual representations, and employing a forward-constrained spatial convolution EEG encoder with CLIP for contrastive learning.
LexChain: Modeling Legal Reasoning Chains for Chinese Tort Case Analysis
Huiyuan Xie (Tsinghua University), Zhiyuan Liu (Northeastern University)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the LexChain framework and evaluation benchmark to systematically model legal reasoning chains in Chinese civil tort cases, and evaluate and improve various large language models.
Libra-MIL: Multimodal Prototypes Stereoscopic Infused with Task-specific Language Priors for Few-shot Whole Slide Image Classification
Zhenfeng Zhuang (Xiamen University), Liansheng Wang (Xiamen University)
CodeClassificationLarge Language ModelMultimodalityBiomedical Data
🎯 What it does: Propose Libra-MIL, a multi-instance learning framework integrating task-specific text priors, bimodal prototype learning, and 3D optimal transport for whole-slide image classification under few annotations.
🎯 What it does: Propose the LiDAR-GS++ method, which incorporates a diffusion prior into Gaussian Splatting to achieve real-time high-quality LiDAR relighting, particularly improving the synthesis effect of extrapolated viewpoints.
LiDARCrafter: Dynamic 4D World Modeling from LiDAR Sequences
Alan Liang (National University of Singapore), Wei Tsang Ooi (National University of Singapore)
CodeGenerationAutonomous DrivingLarge Language ModelDiffusion modelPoint Cloud
🎯 What it does: Proposes LiDARCrafter, a controllable 4D LiDAR generation and editing framework capable of generating dynamic LiDAR sequences containing geometric, motion, and structural priors based on natural language instructions, while supporting scene editing.
Light but Sharp: SlimSTAD for Real-Time Action Detection from Sensor Data
Wei Cui (Institute for Infocomm Research, Agency for Science, Technology and Research), Bing Li (University of Electronic Science and Technology of China)
🎯 What it does: Proposes SlimSTAD, a lightweight framework specifically designed for sensor-based temporal action detection, enabling high-accuracy, low-latency real-time action localization and classification on edge devices.
🎯 What it does: Propose a lightweight framework ATLAS that dynamically constructs web topology graphs and unifies semantic representations for Vision-and-Language web navigation and question answering.
🎯 What it does: Developed a lightweight color harmonization method called MKL-Harmonizer that can achieve real-time AR object color matching on edge devices.
Amit Jena (Texas A&M University), Le Xie (Harvard University)
CodeOptimizationMeta LearningTransformerLarge Language ModelSequentialBenchmarkPhysics Related
🎯 What it does: Proposed the LILAD framework, combining In-Context Learning (ICL) with Lyapunov stability to achieve adaptive and stable system identification.
Listening Between the Frames: Bridging Temporal Gaps in Large Audio-Language Models
Hualei Wang (Beijing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences), Xiangdong Wang (Beijing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelTextMultimodalityAudio
🎯 What it does: Propose TimeAudio, an improved large audio-language model focusing on fine-grained temporal localization and long audio understanding.
🎯 What it does: Propose LiteGE, which uses PCA to compress the Unsigned Distance Field (UDF) and obtain a lightweight shape descriptor, directly predicting geodesic distances on 3D surfaces and achieving shape correspondence.
CodeData SynthesisComputational EfficiencyTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a resource-efficient long-text data synthesis framework called LiteLong, which utilizes the BISAC classification structure and a multi-agent debate mechanism to generate diverse topics, then retrieves and concatenates them into 128K-token training samples using BM25.
LiViBench: An Omnimodal Benchmark for Interactive Livestream Video Understanding
Xiaodong Wang (Peking University), Peixi Peng (Peking University)
CodeLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed LiViBench, an interactive live video full-modal evaluation benchmark covering audio, speech, and real-time comments, constructed with 3,168 videos and 3,175 multiple-choice questions based on a semi-automated annotation workflow.
LLaVA-MS-PIT: Multi-Modal Schema-Guided Progressive Instruction Tuning for Multi-Modal Event Extraction
Hui Zhang (Central China Normal University), Wei Emma Zhang (University of Adelaide)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: Proposes the LLaVA-MS-PIT framework, which achieves multimodal event extraction through progressively instruction-tuned multimodal event patterns.
LLaVA-UHD v2: Exploiting Hierarchical Vision Granularity in MLLMs via Inverse Semantic Pyramid
Yipeng Zhang (Tsinghua University), Maosong Sun (National University Of Singapore)
CodeTransformerSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Propose Hiwin Transformer for enhancing hierarchical visual feature representation in multimodal large language models, constructing an inverse semantic pyramid and compressing features using hierarchical window attention.
CodeRecommendation SystemTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextTabularSequential
🎯 What it does: Proposed the LGSID framework, which first aligns the geographic knowledge of large language models (LLMs) through reinforcement learning, and then achieves hierarchical geographic project tokenization to enhance geographic perception and semantic representation in local life recommendations;
CodeComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed a LLM-free supervised image caption evaluation metric called Pearl, capable of unified assessment in both reference-based and reference-free settings.