AAAI Conference on Artificial Intelligence · 2140 papers
anyECG-chat: A Generalist ECG-MLLM for Flexible ECG Input and Multi-Task Understanding
Haitao Li (Zhejiang University), Zhengxing Huang (Zhejiang University)
CodeSegmentationGenerationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Constructed the anyECG dataset encompassing multiple tasks such as report generation, waveform localization, and multi-ECG comparison, and trained a multilingual large language model named anyECG-chat that supports dynamic length, lead reduction, and multi-ECG input.
AP2O-Coder: Adaptively Progressive Preference Optimization for Reducing Compilation and Runtime Errors in LLM-Generated Code
Jianqing Zhang (Shanghai Jiao Tong University), Jian Cao (Tencent)
CodeOptimizationAI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposed a code generation enhancement method called AP2O-Coder based on offline preference optimization, significantly reducing compilation and runtime errors of LLM-generated code through generating exam-style code, automatic error analysis, step-by-step preference optimization, and adaptive error replay.
Apo2Mol: 3D Molecule Generation via Dynamic Pocket-Aware Diffusion Models
Xinzhe Zheng (University of Florida), Yanjun Li (University of Florida)
CodeDrug DiscoveryProtein Structure PredictionGraph Neural NetworkDiffusion modelBiomedical Data
🎯 What it does: Propose a diffusion-based all-atom model called Apo2Mol, which can simultaneously generate ligands and corresponding holo pockets from apo protein structures.
Approximation Algorithm for Constrained k-Center Clustering: A Local Search Approach
Chaoqi Jia (Rmit University), Jason Xue (Deakin University)
CodeOptimizationTabular
🎯 What it does: Studied the k-center clustering problem with instance-level cannot-link (CL) and must-link (ML) constraints, proposed a threshold and local search based dominating matching set (DMS) framework, and proved it can achieve optimal 2-approximation;
🎯 What it does: Propose a lightweight Affine Prototype-Timestamp (APT) plugin to inject global distribution features into time series prediction, mitigating errors caused by distribution shifts.
🎯 What it does: Propose AquaSplatting dual-branch hybrid 3D representation for real-time reconstruction of underwater scenes and recovery of non-water images.
ARBench: Algorithmic Reasoner or API Alchemist? Evaluating LLMs Beyond API Calls
Ren-Biao Liu (Nanjing University), Ming Li (Nanjing University)
CodeLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the ARBench benchmark to evaluate the algorithmic reasoning capabilities of large language models in implementing machine learning algorithms from scratch without using high-level APIs.
🎯 What it does: Propose an arbitrary-scale high-resolution view rendering framework based on a single 3D Gaussian model, capable of generating high-quality images at both integer and non-integer upscaling ratios.
ARCHE: A Novel Task to Evaluate LLMs on Latent Reasoning Chain Extraction
Pengze Li (Artificial Intelligence Innovation and Incubation Institute of Fudan University), Xi Chen (Artificial Intelligence Innovation and Incubation Institute of Fudan University)
CodeTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the Latent Reasoning Chain Extraction (ARCHE) task, requiring large language models to extract and structure implicit reasoning chains from scientific paper introductions, presenting them in the form of Reasoning Logic Tree (RLT), with nodes as viewpoints and edges as refined forms of Peirce's three types of reasoning (deductive, inductive, abductive);
Shu Wang (Chinese University of Hong Kong), Yuchi Ma (Huawei Cloud Computing Technologies Company Limited)
CodeGenerationRetrievalComputational EfficiencyLarge Language ModelTextGraphRetrieval-Augmented Generation
🎯 What it does: Proposed a hierarchical retrieval-enhanced generation method called ArchRAG based on attribute communities, which efficiently retrieves and injects graph-structured knowledge into large language models to complete question-answering tasks.
🎯 What it does: Proposed and implemented ARDiff, a residual diffusion framework for heterogeneous graph learning, which can progressively denoise and extract relational semantics on the semantic chain CoS.
🎯 What it does: Proposes a fractal nodes mechanism, which divides the graph into subgraphs and introduces nodes representing subgraphs on the original graph, directly aggregating information at the subgraph level and mutually updating with original nodes.
Argumentative Debates for Transparent Bias Detection
Hamed Ayoobi (University of Groningen), Francesca Toni (Imperial College London)
CodeExplainability and InterpretabilityLarge Language ModelTextTabular
🎯 What it does: Propose a transparent bias detection framework ABIDE based on argumentative debate, which judges individual and group bias situations by constructing a quantized bipolar argumentation framework (QBAF) that leverages neighborhood information.
Shaowu Wu (Sun Yat-sen University), Wei Lu (Sun Yat-sen University)
CodeSafty and PrivacyImage
🎯 What it does: Proposed an Adaptive Robust Iterative Watermarking Framework (ARIW-Framework), achieving dual goals of high visual quality and robustness through residual iterative optimization, parallel noise layers, and gradient-adaptive embedding.
ARQUSUMM: Argument-aware Quantitative Summarization of Online Conversations
An Quang Tang (RMIT University), Zhuang Li (RMIT University)
CodeGenerationTransformerLarge Language ModelText
🎯 What it does: An adaptive argument-aware quantitative summarization framework, ARQUSUMM, is proposed for dialogues on online discussion platforms, achieving structured extraction and quantification of argument claims and reasons.
ARTEM: Enhancing Large Language Model Agents with Spatial-Temporal Episodic Memory
Cassandra Hui-Ming Tan (Singapore Management University), Ah-Hwee Tan (Singapore Management University)
CodeRetrievalRepresentation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the ARTEM architecture, integrating large language models with a self-organizing spatiotemporal episodic memory network (STEM) to achieve more precise event encoding and retrieval.
ASCD: Attention-Steerable Contrastive Decoding for Reducing Hallucination in MLLM
Yujun Wang, Yunpu Ma (Christian Albrechts University Of Kiel)
CodeTransformerVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: This paper proposes an attention scheduling-based contrastive decoding method (ASCD), which significantly reduces visual hallucinations and enhances visual-semantic consistency by directly adjusting the attention distribution of multi-modal large language models during inference.
AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basin
Shuo Yang (Peking University), Li Yuan (Peking University)
CodeOptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a safety-guided fine-tuning method called AsFT, which maintains model safety during fine-tuning by projecting parameter updates into a subspace aligned with the alignment direction (safety direction) and suppressing updates orthogonal to it;
Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation
Shiyuan Li (Griffith University), Shirui Pan (Griffith University)
CodeGenerationOptimizationRecurrent Neural NetworkGraph Neural NetworkLarge Language ModelSupervised Fine-TuningAgentic AITextGraphBenchmark
🎯 What it does: Propose ARG-DESIGNER, a self-recursive graph generation model from scratch, to automatically design communication topologies for multi-agent systems (MAS), generating the required number of agents, roles, and communication links based on natural language task queries.
CodeGenerationData SynthesisTransformerLarge Language ModelImageText
🎯 What it does: The study constructs a multi-dimensional humor generation and evaluation dataset based on the Japanese improvisational comedy game 'Oogiri,' systematically assessing the performance of LLMs in generating humorous content and evaluating humor, while exploring differences in humor perception between humans and LLMs.
Guidio Sewa (MIAT, INRAE), Thomas Schiex (MIAT, INRAE)
CodeOptimizationBenchmark
🎯 What it does: This paper proposes and implements a soft local consistency propagator based on the linear assignment problem for efficiently handling ALLDIFFERENT constraints in weighted constraint satisfaction problems.
🎯 What it does: In remote sensing scenarios where multispectral and RGB images lack strong semantic correspondence, the heterogeneous cross-modal knowledge distillation (ACKD) method is proposed, along with the SemBridge framework to enable knowledge transfer.
🎯 What it does: Propose an attention retention framework on Vision Transformers, utilizing gradient masking to suppress attention drift, thereby mitigating catastrophic forgetting in continual learning.
Attention to Threat-Relevant Objects: Reasoning Detection in Autonomous Driving via Multimodal Large Language Models
Yulin He (National University of Defense Technology), Yusong Tan (National University of Defense Technology)
CodeAutonomous DrivingTransformerReinforcement LearningVision Language ModelImageVideoMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the threat-oriented reasoning detection task and built the corresponding benchmark; designed two methods, ThreatCoT (untrained CoT + visual toolchain) and ThreatReasoner (end-to-end reasoning based on GRPO), to enhance threat object detection and threat level assessment.
🎯 What it does: Generate dynamic prompt vectors using node attributes to enhance model robustness under structural perturbations while keeping the original GNN parameters unchanged.
Attribution Analysis-based Concept Alignment: A Human-in-the-loop Data Debugging Framework
Lei Chai (Beihang University), Jingxuan Xu (Beihang University)
CodeExplainability and InterpretabilityData-Centric LearningDiffusion modelImageText
🎯 What it does: Proposed the AACA framework, which performs human-computer interactive image data debugging through attribute analysis and concept alignment;
Augmentation-invariant Learning Strategy via Data Augmentation for Improving Model Generalization
Yu Miao, Yan Qiang (Taiyuan University Of Technology)
CodeClassificationBiomedical Data
🎯 What it does: Proposed the AILS implicit incremental learning strategy, generating semantically invariant data augmentation in the feature space through variational Bayesian inference to enhance the generalization ability of medical image classification.
Augmenting Intra-Modal Understanding in MLLMs for Robust Multimodal Keyphrase Generation
Jiajun Cao (Xiamen University), Jinsong Su (Xiamen University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the AimKP framework to improve multimodal key phrase generation tasks, significantly enhancing the model's cross-modal and single-modal understanding capabilities.
🎯 What it does: Proposes a new framework called AURORA, aiming to enhance language understanding and reasoning capabilities in reference audio-visual segmentation (Ref-AVS) tasks through structured reasoning and reinforcement learning.
Auto-PRE: An Automatic and Cost-Efficient Peer-Review Framework for Language Generation Evaluation
Junjie Chen (Tsinghua University), Qingyao Ai (Ant Group)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose Auto-PRE, an LLM evaluation framework based on automated qualification exams, which screens LLMs using three key metrics: consistency, relevance, and confidence;
AutoLink: Autonomous Schema Exploration and Expansion for Scalable Schema Linking in Text-to-SQL at Scale
Ziyang Wang (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
CodeAI Code AssistantTransformerLarge Language ModelAgentic AITextTabularRetrieval-Augmented Generation
🎯 What it does: Propose AutoLink, an autonomous agent framework based on large language models (LLMs), transforming schema linking into an iterative, interactive exploration process;
AutoMalDesc: Large-Scale Script Analysis for Cyber Threat Research
Alexandru-Mihai Apostu (CrowdStrike), Mihaela Gaman (University of Bucharest)
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose AutoMalDesc, an automated framework that generates large-scale script security descriptions and detects malicious scripts by leveraging a small number of expert-labeled samples through iterative self-supervised learning.
Automated Human Strategic Behavior Modeling via Large Language Models
Xiaohan Xie (Chinese University of Hong Kong), Jianwei Huang (Chinese University of Hong Kong)
CodeOptimizationExplainability and InterpretabilityTransformerLarge Language Model
🎯 What it does: Proposes the AutoBM framework, which leverages large language models to automatically generate, evaluate, and refine interpretable behavioral models;
Automatic Paper Reviewing with Heterogeneous Graph Reasoning over LLM-Simulated Reviewer-Author Debates
Shuaimin Li (Chinese Academy of Sciences), Min Yang (Chinese Academy of Sciences)
CodeClassificationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: Propose the ReViewGraph framework, which uses multi-agent LLMs to simulate multi-round debates between reviewers and authors, extracts viewpoints and debate relationships, constructs a heterogeneous debate graph, and predicts paper acceptance/rejection decisions using graph neural network reasoning.
Automating Complex Document Workflows via Stepwise and Rollback-Enabled Operation Orchestration
Yanbin Zhang (East China Normal University), Renjun Hu (East China Normal University)
CodeClassificationTransformerLarge Language ModelTextMultimodalitySequentialBenchmark
🎯 What it does: Proposed the AutoDW framework, achieving multi-step workflow automation for document processing, supporting incremental planning and rollback;
🎯 What it does: Proposed and implemented the Dynamic Threshold Determination (DTD) algorithm to automatically adjust the threshold of concept drift detectors, thereby improving the overall predictive performance of the model.
AutoPP: Towards Automated Product Poster Generation and Optimization
Jiahao Fan (JD.COM), Ching Law (JD.COM)
CodeGenerationOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelImageText
🎯 What it does: Built the AutoPP automated pipeline to achieve end-to-end generation of high-quality product posters from basic product information, with iterative optimization through online CTR feedback.
AutoTool: Efficient Tool Selection for Large Language Model Agents
Jingyi Jia (Huazhong University of Science and Technology), Qinbin Li (Huazhong University of Science and Technology)
CodeComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelAgentic AIContrastive LearningTextGraph
🎯 What it does: This paper proposes the AutoTool framework, which constructs a graph structure using tool usage inertia to enable LLM-call-free selection of tools and parameters within an LLM agent.
🎯 What it does: By treating neglected low-expressed genes as auxiliary tasks, joint training is employed to enhance the accuracy of predicting spatial gene expression from pathological images.
🎯 What it does: Propose AV-Edit, a fine-grained, semantically aligned sound effect editing framework that leverages visual, audio, and text multimodal information to add, delete, or replace audio elements in videos.
Aware First, Think Less: Dynamic Boundary Self-Awareness Drives Significant Gains in Reasoning Efficiency in Large Language Models
Qiguang Chen (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
CodeComputational EfficiencyLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose a Dynamic Reasoning Boundary Self-Aware Framework (DR.SAF), enabling large language models to assess problem difficulty in real-time during reasoning based on their current reasoning capabilities, dynamically adjusting reasoning depth and output length, thereby significantly reducing redundant reasoning steps.
Chaoyun Wang (Xi'an Jiaotong University), Caigui Jiang (University of Tokyo)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a lightweight convolutional network based on axis-aligned geometric constraints for de-warping single document images, and introduces an axis-aligned preprocessing strategy along with a new evaluation metric called AAD.
Balanced Knowledge Distillation for Large Language Models with Mix-of-Experts
Jiajun Liu (Southeast University), Zijie Xu (Southeast University)
CodeKnowledge DistillationLarge Language ModelMixture of ExpertsText
🎯 What it does: Proposes a new knowledge distillation framework called B-Distill under the MoE structure to compress large language models while maintaining the MoE architecture.
🎯 What it does: Propose an adaptive offline reinforcement learning fine-tuning method, ARFM, which utilizes energy-weighted flow matching to fine-tune Vision-Language-Action (VLA) flow models, thereby improving the action accuracy in complex downstream tasks.
BAT: Learning Event-based Optical Flow with Bidirectional Adaptive Temporal Correlation
Gangwei Xu (Huazhong University of Science and Technology), Xin Yang (Zhejiang University)
CodeConvolutional Neural NetworkRecurrent Neural NetworkOptical FlowTime Series
🎯 What it does: This paper proposes the BAT framework, which estimates optical flow for event cameras by leveraging bidirectional adaptive temporal correlation.
Bayes-Optimal Fair Classification with Multiple Sensitive Features
Yi Yang, Xiangyu Chang (Xi'an Jiaotong University)
CodeClassificationTabular
🎯 What it does: This paper theoretically derives the Bayes-optimal fair classifier under scenarios with multiple sensitive features, and provides Bayes-optimal decision rules that satisfy approximate fairness metrics (mean difference MD and mean ratio MR). Subsequently, it proposes an in-processing method based on regularization during training and a post-processing method based on calibrated plug-in approaches, which can learn these Bayes-optimal classifiers from finite samples.
BayesAgent: Bayesian Agentic Reasoning Under Uncertainty via Verbalized Probabilistic Graphical Modeling
Hengguan Huang (University of Copenhagen), Samir Bhatt (University of Copenhagen)
CodeExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringTextMultimodalityChain-of-Thought
🎯 What it does: Propose a method that uses natural language prompts to enable large language models (LLMs) to autonomously construct latent variables and their dependencies, leveraging numerical Bayesian inference and differentiable calibration loss to form the Verbalized Probabilistic Graphical Modeling (vPGM) framework, achieving agency reasoning and uncertainty quantification without requiring expert-designed models.
Bayesian Network Structural Consensus via Greedy Min-Cut Analysis
Pablo Torrijos (Universidad de Castilla-La Mancha), José A. Gámez (Universidad de Castilla-La Mancha)
CodeFederated LearningExplainability and InterpretabilityTabularBenchmark
🎯 What it does: Propose a greedy structural consensus algorithm (MCBNC) based on minimum cut analysis to obtain sparse, interpretable consensus DAGs from multiple Bayesian Network (BN) structures in federated learning or multi-model aggregation scenarios.
🎯 What it does: Propose the Bayesian Enhancement Model (BEM), which addresses the one-to-many mapping in image enhancement through Bayesian Neural Networks (BNN) and achieves efficient inference via a two-stage BNN-DNN framework.
🎯 What it does: Propose a three-branch collaborative learning framework based on binary cross-entropy (BCE3S) to address the long-tailed recognition problem;
🎯 What it does: Propose a backdoor defense framework called BeDKD based on the direction mapping module and adversarial knowledge distillation, which balances defense effectiveness and model performance using a small number of clean samples and contaminated data.
Behavior Tokens Speak Louder: Disentangled Explainable Recommendation with Behavior Vocabulary
Xinshun Feng (Beihang University), Shuai Wang (Beihang University)
CodeRecommendation SystemExplainability and InterpretabilityGraph Neural NetworkLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningTextSequential
🎯 What it does: Propose the BEAT framework, converting user-item interaction behaviors into discrete behavioral vocabulary interpretable by large language models (LLMs), enabling zero-shot explainable recommendations;
Alex W. Goodall, Francesco Belardinelli (Imperial College London)
CodeReinforcement LearningBenchmark
🎯 What it does: Propose a Behavior Policy Optimization (BPO) framework that designs low-variance behavior policies to collect offline data, achieving more stable return estimation and higher sample efficiency in online reinforcement learning.
CodeTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose a rule knowledge-driven reasoning task, construct an automated large-scale benchmark RULER, and design the RAMPS process reward framework to enhance multi-rule reasoning.
Benchmarking LLMs for Political Science: A United Nations Perspective
Yueqing Liang (Illinois Institute Of Technology), Kai Shu (Illinois Institute Of Technology)
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the UNBench benchmark, systematically organizing and annotating draft proposals, voting records, and discussion texts from the UN Security Council between 1994-2024, covering four political decision-making tasks;
Benchmarking LLMs’ Mathematical Reasoning with Unseen Random Variables Questions
Zijin Hong (Hong Kong Polytechnic University), Xiao Huang (Beihang University)
CodeLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper introduces the RV-BENCH benchmark, which evaluates the true mathematical reasoning ability of LLMs through generated random variable questions (RVQs).
Benchmarking Multimodal Knowledge Conflict for Large Multimodal Models
Yifan Jia (Shandong University), Dongrui Liu (Shanghai AI Laboratory)
CodeLarge Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the MMKC-Bench benchmark to evaluate the behavior and detection capabilities of multimodal models when facing multimodal knowledge conflicts in retrieval-augmented generation.
Benchmarking Visual LLMs Resilience to Unanswerable Questions on Visually Rich Documents
Davide Napolitano (Politecnico di Torino), Fabrizio Battiloro (Politecnico di Torino)
CodeLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This study proposes the VRD-UQA evaluation framework, which automatically detects three types of corruption in multi-page visual rich documents (VRDs): entity, document elements, and layout. It generates answerable but unanswerable questions and evaluates the robustness of visual large language models (VLLMs) in detecting unanswerable questions at both page-level and document-level.
Best-Effort Policies for Robust Markov Decision Processes
Alessandro Abate (University of Oxford), Francesco Fabiano (University of Oxford)
CodeReinforcement LearningTabular
🎯 What it does: Proposes the Best-Effort Robust Strategy (ORBE), providing a systematic method based on dominance and best-effort for resolving multiple optimal robust strategies in RMDP, along with complete theoretical analysis and algorithm implementation.
Better Matching, Less Forgetting: A Quality-Guided Matcher for Transformer-based Incremental Object Detection
Qirui Wu (Northwestern Polytechnical University), Peng Wang (Northwestern Polytechnical University)
CodeObject DetectionTransformerImage
🎯 What it does: In incremental object detection, to address the catastrophic forgetting problem caused by the "background-foreground confusion" in DETR series models, a quality-guided minimum cost maximum flow (Q-MCMF) matcher is proposed. It can retain one-to-one matching while eliminating low IoU erroneous matches, thereby removing incorrect supervision.
🎯 What it does: Proposes a LiDAR-centric multi-modal 3D object detection framework called BEVDilation, which employs image BEV as implicit guidance to alleviate depth estimation errors, and enhances the density and semantic information by sparsifying foreground voxels through the Sparse Voxel Dilation Block (SVDB) and Semantic-Guided BEV Dilation Block (SBDB).
Beware of Reasoning Overconfidence: Pitfalls in the Reasoning Process for Multi-solution Tasks
Jiannan Guan (Harbin Institute of Technology), Wanxiang Che (Du Xiaoman Science Technology Co Ltd)
CodeOptimizationExplainability and InterpretabilityLarge Language ModelBenchmarkChain-of-Thought
🎯 What it does: This paper investigates the overconfidence in reasoning of large language models on multi-solution tasks and proposes the MuSoBench benchmark to systematically evaluate and analyze this issue.
🎯 What it does: This paper proposes a sparse geometry-preserving adapter combination framework based on task potential prototypes, generating new adapters under zero-shot settings by solving ℓ₁-regularized sparse reconstruction in the task prototype space.
CodeRetrievalExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the PDRR framework, which first uses LLM to predict question types and decompose them into structured triplets, then retrieves KB facts through retrieval, and finally completes reasoning and generates answers via LLM; this method achieves training-free, interpretable multi-step reasoning.
Beyond Content: A Comprehensive Speech Toxicity Dataset and Detection Framework Incorporating Paralinguistic Cues
Zhongjie Ba (State Key Laboratory of Blockchain and Data Security, Zhejiang University), Li Lu (State Key Laboratory of Blockchain and Data Security, Zhejiang University)
🎯 What it does: Proposed the ToxiAlert-Bench speech toxicity dataset and the ToxiAlert dual-head model for simultaneous detection of toxicity categories and sources (text, acoustic, or both).
Beyond Cosine Similarity: Magnitude-Aware CLIP for No-Reference Image Quality Assessment
Zhicheng Liao (South China Normal University), Baoliang Chen (South China Normal University)
CodeTransformerVision Language ModelContrastive LearningImageBenchmark
🎯 What it does: This paper proposes a no-reference image quality assessment method, introducing Box-Cox normalization of image embedding amplitude based on the semantic similarity of the CLIP model, and employing a confidence-guided fusion strategy to achieve more accurate quality scoring.
Beyond Counting: Evaluating Abstract and Emotional Reasoning in Vision-Language Models
Yuan Zhou (Institute of Automation, Chinese Academy of Sciences), Shiming Xiang (Institute of Automation, Chinese Academy of Sciences)
CodeLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Created the EmojiGrid benchmark using procedurally synthesized emoji grid images and over 29,000 QA pairs, covering nine cognitive subtasks including perception, relational and structural reasoning, and abstract and emotional reasoning;
🎯 What it does: This work proposes the Geometry-Aware Adaptive Routing (GAAR) module, which dynamically assigns high-level features to Euclidean or hyperbolic space to enhance the geometric representation capability in remote sensing image semantic segmentation.
Beyond Fixed Depth: Adaptive Graph Neural Networks for Node Classification Under Varying Homophily
Asela Hevapathige (Australian National University), Ahad N. Zehmakan (Australian National University)
CodeClassificationGraph Neural NetworkGraph
🎯 What it does: This paper proposes an Adaptive Depth Graph Neural Network (AD-GNN) based on node-level adaptive aggregation depth, which can simultaneously apply to both homogeneous and heterogeneous graphs within the same model.
🎯 What it does: Investigated a co-adaptive curriculum generation method ATLAS based on unsupervised environment design, aiming to simultaneously optimize tasks and environments to enhance the robustness of general agents in scenarios with scarce solvable task-environment pairs.
🎯 What it does: Propose a two-stage method that first recovers global illumination in the frequency domain through the Residual Fourier-Guided Module (RFGM), and then refines textures and edges in the spatial domain using Patch Mamba and Grad Mamba, focusing on fine-grained detail restoration in extremely dark images.
Beyond Immediate Activation: Temporally Decoupled Backdoor Attacks on Time Series Forecasting
Zhixin Liu (Nankai University), Xiangrui Cai (Nankai University)
CodeAdversarial AttackGraph Neural NetworkTime Series
🎯 What it does: Designed and implemented the TDBA framework to achieve spatiotemporal decoupling backdoor attacks in multivariate time series prediction, enabling triggers to be flexibly activated across different dimensions and time delays;
🎯 What it does: Proposes the SPECTRA framework, jointly modeling missing patterns and class imbalance in irregular time series, providing an information-theoretic perspective on the 'missing-imbalance coupling' theory;
Beyond MSE: Ordinal Cross-Entropy for Probabilistic Time Series Forecasting
Jieting Wang (Shanxi University), Xiaolei Shang (Shanxi University)
CodeTransformerTime Series
🎯 What it does: Convert time series regression into an ordinal classification problem, using ordinal cross-entropy (OCE) to train a probabilistic distribution model, which can quantify uncertainty while maintaining robustness.
Beyond Next Token Probabilities: Learnable, Fast Detection of Hallucinations and Data Contamination on LLM Output Distributions
Guy Bar-Shalom (Technion), Haggai Maron (Technion)
CodeAnomaly DetectionTransformerLarge Language ModelText
🎯 What it does: Proposes utilizing the complete output distribution of LLMs (including the probability distribution of the next token and actual token probabilities) to detect hallucinations and data contamination, and designs a lightweight Transformer architecture, LOS-NET, for learning.
Ante Wang (Xiamen University), Jinsong Su (Baidu Inc.)
CodeLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: Propose a Proactive Critical Thinking paradigm, enabling large language models to actively ask questions to obtain missing information when encountering incomplete or incorrect problems, thereby achieving better collaboration and reasoning;
🎯 What it does: Propose a demo-centric anchoring framework DA4ICL, which improves Alzheimer's disease detection in-context learning by utilizing multi-dimensional retrieval and projection vector anchoring;
🎯 What it does: This study proposes a linear-time remote sensing change detection framework called ChangeRWKV based on RWKV, which employs a hierarchical encoder and a spatial-temporal fusion module to achieve efficient multi-scale feature extraction and difference determination.
🎯 What it does: This paper proposes and investigates the impact of gradient noise non-uniformity on the generalization of deep networks, demonstrating its predictive power for generalization and proposing an optimization algorithm based on this metric.
Beyond Single-Point Perturbation: A Hierarchical, Manifold-Aware Approach to Diffusion Attacks
Zhijie Wang (Zhejiang University), Cong Wang (Zhejiang University)
CodeAdversarial AttackDiffusion modelImage
🎯 What it does: By introducing hierarchical multi-step perturbations in the diffusion generation process, combined with unprompted internal self-attention guidance and dynamic manifold alignment, high-fidelity and high-transferability adversarial examples are generated.
🎯 What it does: Proposed the SF-RSSM dual-branch world model, combining a fast branch with residual prediction and a slow branch with GRU, improving model-based reinforcement learning performance through cross-scale distillation and a consistency-driven curiosity module.
Beyond Single-Step Updates: Reinforcement Learning of Heuristics with Limited-Horizon Search
Gal Hadar (Ben Gurion University), Shahaf S. Shperberg (Ben Gurion University)
CodeOptimizationReinforcement LearningGraph
🎯 What it does: Proposed the Limited Horizon Bellman-based Learning (LHBL) method, which updates the heuristic function using finite-depth search and improves state sampling.
Beyond Static: Related Questions Retrieval Through Conversations in Community Question Answering
Xiao Ao (University of Electronic Science and Technology of China), Weikang Guo (Southwestern University of Finance and Economics)
CodeRetrievalTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a label-enhanced conversational question retrieval model called TeCQR, which dynamically refines query representations using clarification questions and user feedback, significantly improving the effectiveness of relevant question retrieval on community Q&A platforms.
🎯 What it does: Investigated the height differences in UAV-based multi-view action recognition, proposing the Aero Partial Order Guided Network (Aerorder) to separate action features from viewpoint features and leverage the partial order relationship of height for knowledge transfer.
Beyond the Mean: Fisher-Orthogonal Projection for Natural Gradient Descent in Large Batch Training
Yishun Lu (University of Oxford), Wesley Armour (University of Oxford)
CodeOptimizationImage
🎯 What it does: Propose the Fisher-Orthogonal Projection (FOP) optimizer, incorporating geometry-aware variance correction into natural gradient updates, enabling training with extremely large batches.
🎯 What it does: Proposes two novel deep reinforcement learning backdoor attacks: component-level TrojanentRL and post-training attack InfrectroRL that does not require a training pipeline;
BhashaKritika: Building Synthetic Pretraining Data at Scale for Indic Languages
Guduru Manoj (Krutrim), Shubham Agarwal (Krutrim)
CodeData SynthesisTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Built and released a multilingual Indian synthetic pre-training corpus named BhashaKritika, containing 540B words, and designed an end-to-end pipeline for multi-technique generation and quality assessment.
Bi-level Personalization for Federated Foundation Models: A Task-vector Aggregation Approach
Yiyuan Yang (University of Technology Sydney), Jing Jiang (University of Technology Sydney)
CodeFederated LearningLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: Propose FedBip, a two-layer personalized framework that performs task-specific fine-tuning on clients and uses task vectors for similarity-weighted aggregation on the server side, achieving efficient personalization and collaboration for federated foundation models.
Bias Association Discovery Framework for Open-Ended LLM Generations
Jinhao Pan (George Mason University), Ziwei Zhu (George Mason University)
CodeExplainability and InterpretabilityLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the Bias Association Discovery Framework (BADF), systematically discovering and quantifying associations between different social identities and descriptive concepts (biases) in open-ended generated text from large language models (LLMs).
Bias-Restrained Prefix Representation Finetuning for Mathematical Reasoning
Sirui Liang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences), Kang Liu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences)
CodeRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Proposes the Bias-Restrained Prefix Representation Finetuning (BREP) method, which significantly improves the performance of large language models (LLMs) on mathematical reasoning tasks by restricting bias magnitude, training prefixes, and intervening only during the early inference phase.