arXivSub Start free trial

AAAI 2026 Papers with Code β€” Page 16

AAAI Conference on Artificial Intelligence Β· 2140 papers

Prompting Adversarial Transferability via Path Flatness Attack

Zeze Tao (Hebei University), Huibing Wang (Dalian Maritime University)

CodeAdversarial AttackImage

🎯 What it does: This paper proposes a path-based flatness attack method called PFA, which significantly enhances the transfer attack effectiveness against black-box models by constraining the flatness of the loss surface along the entire path from the current point to the local minimum.

Promptus: Can Prompt Streaming Replace Video Streaming

Jiangkai Wu (Wangxuan Institute of Computer Technology, Peking University), Xinggong Zhang (Wangxuan Institute of Computer Technology, Peking University)

CodeGenerationCompressionPrompt EngineeringDiffusion modelVideo

🎯 What it does: Convert video frames into prompts for Stable Diffusion, generating video at the receiver end to achieve ultra-low bitrate video streaming.

ProPL: Universal Semi-Supervised Ultrasound Image Segmentation via Prompt-Guided Pseudo-Labeling

Yaxiong Chen (Wuhan University of Technology), Lichao Mou (Technical University of Munich)

CodeSegmentationConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelImageBiomedical DataUltrasound

🎯 What it does: This paper proposes the ProPL framework, achieving multi-organ, multi-task unsupervised semi-supervised ultrasound image segmentation.

Prototype Entropy Alignment: Reinforcing Structured Uncertainty in LLM Reasoning

Zhengyuan Pan (Xiamen University), Qingqiang Wu (University of Chinese Academy of Sciences)

CodeExplainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: Proposed Prototype Entropy Alignment (PEA), a reinforcement learning framework that clusters high-entropy token distributions from expert reasoning trajectories into dynamically updatable prototypes, using these prototypes as process-level rewards to guide large language models' reasoning processes.

Prototype-Driven Active Domain Adaptation with Density Consideration

Zeyu Zhang (Jilin University), Shuai LΓΌ

CodeDomain AdaptationGenerative Adversarial NetworkImageBenchmark

🎯 What it does: This paper proposes an active domain adaptation framework called PDADA, which enhances target domain performance by leveraging prototype-driven density-aware sample selection and adversarial training.

ProtSAE: Disentangling and Interpreting Protein Language Models via Semantically-Guided Sparse Autoencoders

Xiangyu Liu (Nanjing University), Wei Hu (Nanjing University)

CodeExplainability and InterpretabilityTransformerLarge Language ModelAuto EncoderBiomedical Data

🎯 What it does: Semantic-guided training of sparse autoencoders on protein language models to remove semantic contamination, enhance interpretability, and support generation control.

Proxy Zero-Shot Hashing with Multimodal Fusion via Stable Diffusion

Hui Zhang (Ocean University of China), Yuan Cao (Ocean University of China)

CodeRetrievalDiffusion modelContrastive LearningMultimodalityBenchmark

🎯 What it does: Propose a zero-shot hashing framework called PZSH based on Stable Diffusion for multi-modal pseudo image generation and alignment;

Prune&Comp: Free Lunch for Layer-Pruned LLMs via Iterative Pruning with Magnitude Compensation

Xinrui Chen (Tsinghua University), Chun Yuan (Tsinghua University)

CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark

🎯 What it does: This paper proposes a training-agnostic iterative layer pruning method called PRUNE&COMP, which utilizes amplitude compensation techniques to recover the hidden state amplitude gaps caused by removed layers during pruning, thereby maintaining model performance;

PSEO: Optimizing Post-hoc Stacking Ensemble Through Hyperparameter Tuning

Beicheng Xu (Peking University), Bin Cui (Peking University)

CodeOptimizationHyperparameter SearchTabular

🎯 What it does: Propose the PSEO framework, which enhances AutoML's predictive performance through hyperparameter optimization of post-stacking ensembles.

PSPO: Prompt-Level Prioritization and Experience-Weighted Smoothing for Efficient Policy Optimization

Xinxin Zhu (Shenzhen University), F. Richard Yu (Guangdong Laboratory of Artificial Intelligence and Digital Economy)

CodeOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark

🎯 What it does: Proposed a lightweight PSPO algorithm that improves data efficiency and training stability in LLM alignment through experience-weighted reward smoothing and prompt-level prioritized sampling.

PsyPARSE: Retrieval-Augmented Slow Thinking for Personalized Empathetic Counseling

Longxiang Wang (Chongqing University), Ronghao Chen (Peking University)

CodePrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Proposes a training-free psychological counseling framework, PsyPARSE, that simulates the deep thinking and personalized responses of human psychotherapists.

PUFM: Efficient Point Cloud Upsampling via Flow Matching

Zhi-Song Liu (Lappeenranta-Lahti University Of Technology), Lei Li (University Of Leicester)

CodeGenerationComputational EfficiencyGraph Neural NetworkFlow-based ModelPoint Cloud

🎯 What it does: Propose a point cloud up-sampling framework PUFM based on flow matching, which directly learns the optimal transport path from sparse point clouds to dense point clouds, achieving efficient and fine-grained point cloud up-sampling.

PulseMind: A Multi-Modal Medical Model for Real-World Clinical Diagnosis

Jiao Xu (Dalian University of Technology), Ping Wang (Peking University)

CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBiomedical DataBenchmark

🎯 What it does: This paper proposes PulseMind, a multimodal medical large model designed for real-world clinical diagnosis, encompassing the newly constructed MediScope dialogue dataset, the PulseMind Benchmark evaluation framework, and the CRPO reinforcement learning training framework.

Pushing Rendering Boundaries: Hard Gaussian Splatting

Qingshan Xu, Hanwang Zhang (Nanyang Technological University)

CodeGenerationNeural Radiance FieldGaussian SplattingImage

🎯 What it does: Propose the Hard Gaussian Splatting (HGS) framework, which leverages multi-view significant location gradients and rendering errors to mine and grow hard Gaussians, significantly improving the quality of Neural View Synthesis (NVS) for 3D scenes;

Put the Space of LoRA Initialization to the Extreme to Preserve Pre-trained Knowledge

Pengwei Tang (Renmin University of China), Debing Zhang (Xiaohongshu Inc)

CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: Designed a new LoRA initialization method called LoRA-Null, which utilizes the null space of pre-trained knowledge activation to initialize the LoRA adapter, thereby better preserving the knowledge of the pre-trained model during fine-tuning.

Putting Fair Division on the Map

Paula BΓΆhm (Institut fΓΌr Informatik TU Clausthal), StanisΕ‚aw Szufa (University of Chicago)

CodeOptimizationExplainability and InterpretabilityTabular

🎯 What it does: This paper constructs two methods to map fair division instances of indivisible goods to a two-dimensional plane, aiming to reveal structural differences between instances and analyze the distributions and properties of real and synthetic instances.

QAPNet: A Quantum-Attentive Patchwise Network for Robust Medical Image Classification Under Noisy Inputs

Maqsudur Rahman (Boise State University), Jun Zhuang (Boise State University)

CodeClassificationContrastive LearningImageBiomedical DataMagnetic Resonance Imaging

🎯 What it does: Designed and implemented QAPNet, which encodes image patches through quantum variational circuits, reweights instances using additive attention, and incorporates supervised contrastive learning with prototype regularization to achieve robust medical image classification under Gaussian noise input.

QiMeng-CRUX: Narrowing the Gap Between Natural Language and Verilog via Core Refined Understanding eXpression

Lei Huang (Chinese Academy of Sciences), Qi Guo (Chinese Academy of Sciences)

CodeGenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText

🎯 What it does: Propose a structured intermediate representation space called CRUX to bridge the gap between natural language descriptions and Verilog code, and build a two-phase training framework (Joint Expression Modeling and Dual-Space Optimization) to achieve more accurate and synthesizable Verilog code generation.

QRShield: Exploiting Vulnerabilities of Latent Diffusion Models for Preventing AI Art Plagiarism

Xunyue Mo (Sun Yat-sen University), Zibin Zheng (Sun Yat-sen University)

CodeGenerationSafty and PrivacyAdversarial AttackDiffusion modelAuto EncoderImage

🎯 What it does: This paper proposes QRShield, a lightweight adversarial artwork protection method targeting Latent Diffusion Models.

Quality-Aware Language-Conditioned Local Auto-Regressive Anomaly Synthesis and Detection

Long Qian (Chinese Academy of Sciences), Jinqiao Wang (Chinese Academy of Sciences)

CodeGenerationData SynthesisAnomaly DetectionTransformerVision Language ModelAuto EncoderImage

🎯 What it does: Proposes ARASβ€”a language-conditioned local autoregressive anomaly synthesis method, embedded into the QARAD quality-aware weighted detection framework, achieving high-resolution, text-controllable anomaly generation and detection.

QuanTaxo: A Quantum Approach to Self-Supervised Taxonomy Expansion

Sahil Mishra (Indian Institute of Technology Delhi), Tanmoy Chakraborty (Indian Institute of Technology Delhi)

CodeRepresentation LearningText

🎯 What it does: Proposes QuanTaxo, a self-supervised hierarchical classification expansion framework based on quantum embeddings, designed to insert new concepts into existing taxonomies.

Quantifying and Improving Adaptivity in Conformal Prediction Through Input Transformations

Sooyong Jang (University of Pennsylvania), Insup Lee (University of Pennsylvania)

CodeClassificationRecognitionImage

🎯 What it does: This paper proposes a transformation ranking and balanced binning method based on input perturbation for constructing adaptive prediction sets in conformal prediction, and introduces two new evaluation metrics, T-CV and T-SS, to more accurately measure the adaptability of prediction sets to sample difficulty. Based on the binning results, it also introduces grouped conditional conformal prediction, leading to improved adaptive prediction set algorithms such as O-LAC and O-SAPS.

Quantifying the Potential to Escape Filter Bubbles: A Behavior-Aware Measure via Contrastive Simulation

Difu Feng (Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)

CodeRecommendation SystemTransformerLarge Language ModelContrastive LearningTabular

🎯 What it does: Conduct a quantitative evaluation of filter bubbles, proposing a behavior-based contrast metric called Bubble Escape Potential (BEP).

Quantum Non-Linear Bandit Optimization

Zakaria Shams Siam (University at Albany State University of New York), Chong Liu (University at Albany State University of New York)

CodeOptimizationHyperparameter SearchTabular

🎯 What it does: Proposed a quantum nonlinear bandit optimization algorithm, Q-NLB-UCB, which achieves logarithmic cumulative regret independent of input dimensions in high-dimensional spaces;

QuantVSR: Low-Bit Post-Training Quantization for Real-World Video Super-Resolution

Bowen Chai (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)

CodeSuper ResolutionComputational EfficiencyVideo

🎯 What it does: This paper proposes QuantVSR, a low-bit post-training quantization framework for real-world video super-resolution.

QuEPT: Quantized Elastic Precision Transformers with One-Shot Calibration for Multi-Bit Switching

Ke Xu (Anhui University), Xingyi Zhang (Anhui University)

CodeComputational EfficiencyTransformerLarge Language ModelImageTextMultimodalityBenchmark

🎯 What it does: Propose QuEPT, a Transformer-based elastic multi-bit post-training quantization framework that supports real-time switching between multi-bit widths and quantization of weights and activations with a single calibration.

Query-Routed Activation Editing with Truth-hierarchical Preference Optimization

Kewei Liao (Beihang University), Xianglong Liu (Beihang University)

CodeTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: Proposes the Query Routing and Activation Editing (QRAE) framework to dynamically route queries to the most relevant attention heads and customize activation edits to reduce hallucinations in large language models (LLMs).

QuMAB: Query-based Multi-annotator Behavior Pattern Learning

Liyun Zhang (University of Tokyo), Yuta Nakashima (Tongji University)

CodeExplainability and InterpretabilityRepresentation LearningTransformerVision Language ModelImageVideo

🎯 What it does: Proposed QuMAB, a query-based multi-annotator behavior pattern learning framework, focusing on capturing individual discriminative behaviors of each annotator rather than simply seeking sample-level consistency.

QuoTA: Query-oriented Token Assignment via CoT Query Decouple for Long Video Comprehension

Yongdong Luo (Xiamen University), Jiebo Luo (University of Rochester)

CodeTransformerLarge Language ModelVision Language ModelVideoTextBenchmarkChain-of-Thought

🎯 What it does: Proposes a training-free query-oriented visual token allocation framework named QuoTA, enhancing the efficiency and effectiveness of long video understanding.

R-Tuning: Wavelet-Decomposed Replay and Semantic Alignment for Continual Adaptation of Pretrained Time-Series Models

Tianyi Yin (Tongji University), Weiming Shen (Zhongguancun Academy)

CodeDomain AdaptationKnowledge DistillationTransformerTime Series

🎯 What it does: Proposes the Replay Tuning (R-Tuning) framework, which enables continuous adaptation of pre-trained time series models without original data by utilizing frequency-aware replay samples and semantic alignment.

RAA: Achieving Interactive Remove/Add Anything via Fully Synthetic Data

Delong Liu (Beijing University of Posts and Telecommunications), Zhicheng Zhao (Beijing University of Posts and Telecommunications)

CodeImage HarmonizationGenerationData SynthesisTransformerLarge Language ModelDiffusion modelFlow-based ModelImageMultimodality

🎯 What it does: Propose a fully automated, self-improving synthetic data generation pipeline that generates over 514,510 pairs of high-quality image editing samples, and build the RAA framework to achieve precise object addition/removal editing based on this;

RAC-DMVC: Reliability-Aware Contrastive Deep Multi-View Clustering Under Multi-Source Noise

Shihao Dong (Nanjing University of Information Science and Technology), Xinzhong Zhu (Zhejiang Normal University)

CodeRepresentation LearningAuto EncoderContrastive LearningMultimodalityBenchmark

🎯 What it does: Proposes a Reliability-Aware Contrastive Deep Multi-View Clustering Framework (RAC-DMVC), capable of achieving unsupervised clustering in multi-source noisy environments where observation noise and missing noise coexist.

Radar-APLANC: Unsupervised Radar-based Heartbeat Sensing via Augmented Pseudo-Label and Noise Contrast

Ying Wang (Nanjing University of Information Science & Technology), Xiaobai Li (Zhejiang University)

CodeData-Centric LearningContrastive LearningTime SeriesBiomedical Data

🎯 What it does: Proposed and implemented an unsupervised radar heartbeat detection framework called Radar-APLANC, which extracts heartbeat signals using enhanced pseudo labels and noise contrastive learning.

RadarMP: Motion Perception for 4D mmWave Radar in Autonomous Driving

Ruiqi Cheng, Wei Liang (Beijing Racobit Electronic Information Technology Co Ltd)

CodeObject DetectionAutonomous DrivingTransformerOptical FlowMultimodalityPoint Cloud

🎯 What it does: Propose the RadarMP framework, which utilizes 4D mmWave radar tensors from consecutive frames to simultaneously perform radar target detection and 3D scene flow estimation, achieving motion perception;

RAG-Enhanced Collaborative LLM Agents for Drug Discovery

Namkyeong Lee (Korea Advanced Institute of Science and Technology), Gabriele Scalia (Genentech)

CodeDrug DiscoveryTransformerAgentic AITextGraphBiomedical DataRetrieval-Augmented Generation

🎯 What it does: Proposed a multi-agent, retrieval-augmented generation (RAG)-driven CLADD framework for zero-shot drug discovery question-answering tasks, capable of integrating external knowledge graphs and annotations without domain-specific fine-tuning.

RAG-R1:Incentivizing the Search and Reasoning Capabilities of LLMs Through Multi-Query Parallelism

Zhiwen Tan, Jinjie Gu (Ant Group)

CodeRetrievalComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Propose the RAG-R1 framework, which enables LLMs to adaptively leverage internal and external knowledge during inference and support think-then-search interaction through two-phase training (Format Learning SFT + Retrieval-Enhanced RL);

RAGFort: Dual-Path Defense Against Proprietary Knowledge Base Extraction in Retrieval-Augmented Generation

Qinfeng Li (Zhejiang University), Xuhong Zhang (Zhejiang University)

CodeSafty and PrivacyTransformerContrastive LearningTextRetrieval-Augmented Generation

🎯 What it does: To address knowledge base reconstruction attacks in RAG systems, this paper proposes RAGFort, a dual-path defense framework that combines mutual class isolation and intra-class suppression. It systematically evaluates the contributions of two attack paths and achieves comprehensive protection.

Random Amalgamation of Adapters for Flatter Loss Landscapes: Towards Class-Incremental Learning with Better Stability

Yao Deng (Huazhong University of Science and Technology), Jiaqi Gui (Huazhong University of Science and Technology)

CodeClassificationOptimizationRepresentation LearningTransformerImage

🎯 What it does: This paper proposes a category-incremental learning method for pre-trained models, significantly enhancing the model's stability and performance during the incremental learning process through the combination of Randomly Mixed Adapters (RAA) and an improved Sharpness-Aware Minimization (SAM+).

Random is Faster than Systematic in Multi-Objective Local Search

Zimin Liang (University of Birmingham), Miqing Li (University of Birmingham)

CodeOptimizationBenchmark

🎯 What it does: Investigate the efficiency differences between random exploration of neighborhoods and systematic exploration in multi-objective local search.

RAPTOR: Real-Time High-Resolution UAV Video Prediction with Efficient Video Attention

Zhan Chen (Aerospace Information Research Institute Chinese Academy of Sciences), Yidan Zhang (Aerospace Information Research Institute Chinese Academy of Sciences)

CodeGenerationAutonomous DrivingComputational EfficiencyVideo

🎯 What it does: Proposes the RAPTOR framework for single-channel real-time generation of high-resolution future video frames, addressing the speed-resolution-quality trilemma in UAV navigation.

RatioSketch: Towards More Accurate Frequency Estimation in Data Streams via a Lightweight Neural Network

Mengbo Wang (Tsinghua University), Mingwei Xu (Peking University)

CodeComputational EfficiencySupervised Fine-TuningTime Series

🎯 What it does: Proposed a lightweight neural network correction framework called RatioSketch to enhance the frequency estimation accuracy of traditional hand-crafted sketches (e.g., CMS, TowerSketch) and existing neural sketches (MetaSketch, LegoSketch).

RAW-Flow: Advancing RGB-to-RAW Image Reconstruction with Deterministic Latent Flow Matching

Zhen Liu (University of Electronic Science and Technology of China), Shuaicheng Liu (Independent Researcher)

CodeImage TranslationRestorationFlow-based ModelAuto EncoderImage

🎯 What it does: Propose a RGBβ†’RAW image reconstruction framework (RAW-Flow) based on potential flow matching, treating inverse ISP as a deterministic latent space transmission problem.

RCP-LO: A Relative Coordinate Prediction Framework for Generalizable Deep LiDAR Odometry

Chen Liu (Xiamen University), Cheng Wang (Xiamen University)

CodePose EstimationAutonomous DrivingConvolutional Neural NetworkDiffusion modelSimultaneous Localization and MappingPoint Cloud

🎯 What it does: This paper proposes a framework called RCP-LO, which rewrites LiDAR pose estimation as relative coordinate prediction. It generates robust coordinates using diffusion models and solves the pose in one step through differentiable geometric weighted singular value decomposition, achieving end-to-end training and efficient inference.

RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior

Junyao Yang (South China University of Technology), Ziqian Zeng (South China University of Technology)

CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelTextTabularBiomedical DataFinance RelatedChain-of-Thought

🎯 What it does: This paper proposes a new model merging framework called RCP-Merging, which treats the parameters of long-chain reasoning models as priors, preserving reasoning capabilities during merging while incorporating domain knowledge sensitivity and reasoning retention metrics to achieve a balance between domain knowledge and reasoning capabilities.

Re-SpS: A Reinforcement Learning Approach to Speculative Sampling

Chenan Wang (William & Mary), Haipeng Chen (William & Mary)

CodeComputational EfficiencyRepresentation LearningLarge Language ModelReinforcement LearningText

🎯 What it does: To address the inference latency issue in large language models, we propose the Re-SpS framework, which dynamically adjusts the hyperparameters of speculative sampling using reinforcement learning.

REACT-LLM: A Benchmark for Evaluating LLM Integration with Causal Features in Clinical Prognostic Tasks

Linna Wang (Sichuan University), Li Lu (Sichuan University)

CodeClassificationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringBiomedical DataElectronic Health RecordsBenchmarkChain-of-Thought

🎯 What it does: Built a benchmark framework named REACT-LLM to evaluate the synergistic effects of large language models (LLMs) and causal features in clinical prognosis tasks.

Real-time 3D Object Detection with Inference-Aligned Learning

Chenyu Zhao (Wuhan University), Nan Xue (Ecole Polytechnique Federale De Lausanne)

CodeObject DetectionKnowledge DistillationConvolutional Neural NetworkPoint Cloud

🎯 What it does: Propose the SR3D framework to achieve real-time indoor point cloud 3D object detection, bridging the training-inference gap through spatial-prioritized label assignment and self-distillation mechanisms.

Realistic Face Reconstruction from Facial Embeddings via Diffusion Models

Dong Han (Huawei Technologies), Joachim Denzler (Friedrich Schiller University)

CodeGenerationSafty and PrivacyDiffusion modelImage

🎯 What it does: Propose a privacy-preserving facial reconstruction framework called FEM, which maps arbitrary facial embeddings to a pre-trained IPA-FaceID diffusion model to generate high-resolution, realistic facial images, and verifies its attack effectiveness in traditional FR and PPFR systems.

REAP: Enhancing RAG with Recursive Evaluation and Adaptive Planning for Multi-Hop Question Answering

Yijie Zhu (Jiangnan University), Ning Wang (Jiangnan University)

CodeTransformerTextRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Propose the REAP framework, combining Sub-Task Planner and Fact Extractor to iteratively enhance RAG's reasoning and fact extraction in multi-hop question answering.

Reasoning Transfer for an Extremely Low-Resource and Endangered Language: Bridging Languages Through Sample-Efficient Language Understanding

Khanh-Tung Tran (University College Cork), Hoang D. Nguyen (University College Cork)

CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought

🎯 What it does: This paper proposes an English-Pivoted Chain-of-Thought (CoT) training method, which enhances the mathematical reasoning ability of extremely low-resource languages (e.g., Irish) by enforcing the reasoning steps to remain in English internally within the model.

ReBoot: Encrypted Training of Deep Neural Networks with CKKS Bootstrapping

Alberto Pirillo (Politecnico di Milano), Luca Colombo (Politecnico di Milano)

CodeSafty and PrivacyImageTabular

🎯 What it does: Propose the ReBoot framework to realize fully encrypted, non-interactive multi-layer perceptron (MLP) training.

RecCocktail: A Generalizable and Efficient Framework for LLM-Based Recommendation

Min Hou (Hefei University Of Technology), Meng Wang (Shanghai Jiao Tong University)

CodeRecommendation SystemLarge Language ModelSupervised Fine-TuningTextTabular

🎯 What it does: Propose the RecCocktail framework, integrating global and domain-specific LLM recommendation paradigms through LoRA weight mixing to achieve model generality and domain specialization;

ReCode: Updating Code API Knowledge with Reinforcement Learning

Haoze Wu (Zhejiang University), Ningyu Zhang (Nanjing University)

CodeAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmark

🎯 What it does: This paper proposes the ReCode framework, which utilizes reinforcement learning fine-tuning to enable large language models to migrate code when facing updates to external library APIs.

RECoRD: A Multi-Agent LLM Framework for Reverse Engineering Codebase to Relational Diagram

Yuan Xue (Ohio State University), Hoiyi Ng

CodeExplainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraphBenchmark

🎯 What it does: Construct the RECoRD multi-agent framework, utilizing LLM reverse parsing of code to generate causal graphs.

Recovering Coherent Affective Patterns: Addressing Modality Missing in Multimodal Sentiment Analysis

Huiting Huang (Xi'an Jiaotong University), Mengling Feng (National University of Singapore)

CodeClassificationRestorationExplainability and InterpretabilityTransformerGenerative Adversarial NetworkMultimodality

🎯 What it does: Propose a two-stage multimodal sentiment analysis framework RECAP, addressing the common modality missing problem in real scenarios. First, a generative model is used to recover missing modalities, then self-supervised mutual information decomposition ensures semantic completeness, followed by information-guided ranking attention for modality fusion and sentiment prediction.

Rectification Reimagined: A Unified Mamba Model for Image Correction and Rectangling with Prompts

Linwei Qiu (Beihang University), Fengying Xie (Chinese University of Hong Kong)

CodeRestorationPrompt EngineeringMixture of ExpertsOptical FlowImage

🎯 What it does: Proposed a unified image correction and rectangularization framework called UniRect, which can complete four common camera distortion correction tasks (portrait correction, wide-angle rectangularization, stitching rectangularization, and rotation correction) within the same model.

RecToM: A Benchmark for Evaluating Machine Theory of Mind in LLM-based Conversational Recommender Systems

Mengfan Li (Huazhong University of Science and Technology), Yang Deng (Singapore Management University)

CodeRecommendation SystemTransformerLarge Language ModelTextBenchmarkChain-of-Thought

🎯 What it does: Propose the RECTOM benchmark to evaluate the theory of mind (ToM) capabilities of large language models in conversational recommendation systems;

Reducing Goal State Divergence with Environment Design

Kelsey Sikes (Colorado State University), Sarath Sreedharan (Colorado State University)

CodeRobotic IntelligenceReinforcement Learning from Human FeedbackBenchmark

🎯 What it does: Propose the Goal State Divergence (GSD) metric to measure the discrepancy between the robot's final state after completing a human-specified task and the human's expected state, and define the Human-Robot Goal State Alignment Design (HRGAD) problem, which aims to find the minimal environmental modifications to reduce GSD.

Reducing the Scope of Language Models

David Yunis (Toyota Technological Institute at Chicago), Danish Contractor (IBM Research AI)

CodeDomain AdaptationTransformerSupervised Fine-TuningText

🎯 What it does: Investigated and evaluated multiple methods to restrict large language models to a specified domain (i.e., 'scoping'), making them generate answers only for relevant queries and reject irrelevant ones.

Redundant Queries in DETR-Based 3D Detection Methods: Unnecessary and Prunable

Lizhen Xu (Xi'an Jiaotong University), Jianru Xue (Xi'an Jiaotong University)

CodeObject DetectionComputational EfficiencyTransformerImagePoint Cloud

🎯 What it does: Propose the GPQ (Gradually Pruning Queries) method, which gradually prunes redundant queries in the DETR-based 3D detection model to reduce computational and memory costs.

RefAdGen: High-Fidelity Advertising Image Generation

Yiyun Chen (Hong Kong University of Science and Technology (Guangzhou)), Weikai Yang (Hong Kong University of Science and Technology (Guangzhou))

CodeGenerationData SynthesisConvolutional Neural NetworkDiffusion modelImage

🎯 What it does: Proposed a high-fidelity advertising image generation framework called RefAdGen that does not require fine-tuning for each product individually.

RefiDiff: Progressive Refinement Diffusion for Efficient Missing Data Imputation

Md Atik Ahamed (University of Kentucky), Qiang Cheng (University of Kentucky)

CodeComputational EfficiencyData-Centric LearningDiffusion modelTabularBenchmarkStochastic Differential Equation

🎯 What it does: Propose a hybrid predictive and generative missing value imputation method called RefiDiff, adopting a progressive refinement strategy in pre- and post-stages;

Refine-IQA: Multi-Stage Reinforcement Finetuning for Perceptual Image Quality Assessment

Ziheng Jia (Shanghai Jiao Tong University), Xiongkuo Min (Shanghai Jiao Tong University)

CodeLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality

🎯 What it does: Propose a multi-stage reinforcement learning fine-tuning framework called Refine-IQA to enhance the perception and scoring capabilities of multimodal large models in image quality assessment (IQA).

Refinement Contrastive Learning of Cell–Gene Associations for Unsupervised Cell Type Identification

Liang Peng (Shantou University), Hau-San Wong (Shantou University)

CodeClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraphBiomedical Data

🎯 What it does: Developed an unsupervised cell type identification framework named scRCL, which utilizes cell-gene associations to improve cell embedding representations.

Reflect Then Learn: Active Prompting for Information Extraction Guided by Introspective Confusion

Dong Zhao (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)

CodeTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: Propose an active prompting framework APIE based on LLMs, which actively selects the most informative examples by evaluating model uncertainty through 'introspective confusion,' thereby improving the accuracy and robustness of information extraction with few samples.

ReflexDiffusion: Reflection-Enhanced Trajectory Planning for High-lateral-acceleration Scenarios in Autonomous Driving

Xuemei Yao (National University of Defense Technology), Kewei Yang (National University of Defense Technology)

CodeAutonomous DrivingDiffusion model

🎯 What it does: Proposes improving diffusion model-based trajectory planning through a reflection mechanism during the inference phase, specifically enhancing vehicle driving safety in high lateral acceleration scenarios.

REFO: Reinforced Evolutionary Faithfulness Optimization for Large Language Models

Yi Wang (Hong Kong University of Science and Technology), Sihong Xie (Hong Kong University of Science and Technology)

CodeRetrievalOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation

🎯 What it does: Proposed the REFO framework, which leverages a self-evolving iterative process to automatically generate high-quality data and employs DPO for iterative training, significantly enhancing the faithfulness of retrieval-augmented generation (RAG) systems.

RefSTAR: Blind Face Image Restoration with Reference Selection, Transfer, and Reconstruction

Zhicun Yin (Harbin Institute of Technology), Wangmeng Zuo (Nanyang Technological University)

CodeRestorationConvolutional Neural NetworkDiffusion modelImage

🎯 What it does: Proposed the RefSTAR framework for blind face image restoration tasks, significantly enhancing facial identity preservation and texture detail recovery through three steps: reference image selection, transfer, and reconstruction.

Region-Point Joint Representation for Effective Trajectory Similarity Learning

Hao Long (University of Electronic Science and Technology of China), Shuo Shang (University of Electronic Science and Technology of China)

CodeRetrievalRepresentation LearningConvolutional Neural NetworkGraph Neural NetworkTransformerMixture of ExpertsContrastive LearningGraphTime SeriesStochastic Differential Equation

🎯 What it does: Proposes a trajectory similarity learning framework called RePo that integrates regional and point-level features for more accurate trajectory similarity computation.

RegionRAG: Region-level Retrieval-Augmented Generation for Visual Document Understanding

Yinglu Li (University of Science and Technology of China), Hongtao Xie (University of Science and Technology of China)

CodeRetrievalLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought

🎯 What it does: Propose the RegionRAG framework, shifting retrieval granularity from document-level to region-level, leveraging region-based retrieval to provide precise context for large language models, thereby enhancing visual document understanding and question-answering performance.

Regression over Classification: Assessing Image Aesthetics via Multimodal Large Language Models

Xingyuan Ma (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)

CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodality

🎯 What it does: Propose a new method, ROC4MLLM, to improve score prediction of multimodal large language models in image aesthetics assessment (IAA);

Regular Games -- an Automata-Based General Game Playing Language

RadosΕ‚aw Miernik (University of WrocΕ‚aw), Wojciech Pawlik (University of WrocΕ‚aw)

CodeOptimizationComputational EfficiencyBenchmark

🎯 What it does: Proposed a new general-purpose game play system called Regular Games (RG), which implements rule descriptions through a low-level finite automaton language and provides high-level languages and tool ecosystems, making game design both concise and efficient.

Reimagining Anomalies: What If Anomalies Were Normal?

Philipp Liznerski (RPTU University Kaiserslautern-Landau), Marius Kloft (RPTU University Kaiserslautern-Landau)

CodeAnomaly DetectionExplainability and InterpretabilityDiffusion modelGenerative Adversarial NetworkImage

🎯 What it does: Propose a method using adversarial generative networks and diffusion models to generate diverse counterfactual examples to explain the discriminative criteria of image anomaly detectors.

Reinforce Trustworthiness in Multimodal Emotional Support System

Huy M. Le (Mohamed bin Zayed University of Artificial Intelligence), Binh T. Nguyen (Singapore Management University)

CodeClassificationRecognitionReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityAudio

🎯 What it does: Proposed and implemented the MultiMood framework, which integrates multimodal features from video, audio, and text to predict emotional support elements, generates responses aligned with professional psychological therapy standards using a large language model (LLM), and enhances response credibility through reinforcement learning.

Rejoining Precious Artifacts: Efficiently Bone Stick Rejoining Based Massive Fragment Images by Contour, Script, and Texture

Xingyi Wang (Sichuan University), Jian Peng (Sichuan University)

CodeClassificationRecognitionConvolutional Neural NetworkGraph Neural NetworkImage

🎯 What it does: This paper proposes a lightweight visual graph neural network, RejoinViG, which can directly determine the feasibility of joining bone pen fragments and predict the joining direction using fragment images without requiring manual annotations.

Reliable-View 2D-3D Key-Part Aligned Transformer with Reinforced Masking for 3D Point Cloud Understanding

Xianglong Jin (Northwestern Polytechnical University), Feiping Nie (Northwestern Polytechnical University)

CodeClassificationSegmentationTransformerReinforcement LearningAuto EncoderContrastive LearningPoint CloudBenchmark

🎯 What it does: Train 3D point cloud representations using self-supervised multi-view 2D-3D alignment and reinforced masking Masked Autoencoder, integrating multi-view images to enhance point cloud understanding.

Relink: Constructing Query-Driven Evidence Graph On-the-Fly for GraphRAG

Manzong Huang (Hefei University of Technology), Xindong Wu (College of William and Mary)

CodeRetrievalRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextGraphRetrieval-Augmented Generation

🎯 What it does: Built a dynamic, query-driven evidence graph framework called Relink to instantly complete missing knowledge and filter noisy facts in open-domain question answering, enhancing the accuracy of multi-hop reasoning.

Remember Me: Bridging the Long-Range Gap in LVLMs with Three-Step Inference-Only Decay Resilience Strategies

Peng Gao (Hong Kong Baptist University), Hui Zhang (University of Wollongong)

CodeComputational EfficiencyVision Language ModelMultimodality

🎯 What it does: Propose a three-step decay tolerance strategy (T-DRS) during the reasoning phase, utilizing semantic-driven, distance control, and remote re-enhancement mechanisms to address the long-range attention decay issue in large vision-language models based on RoPE.

Remodeling Semantic Relationships in Vision-Language Fine-Tuning

Xiangyang Wu, Zhenwei Shi (Nanyang Technological University)

CodeComputational EfficiencyRepresentation LearningSupervised Fine-TuningVision Language ModelMultimodality

🎯 What it does: Propose a parameter-efficient visual-language fine-tuning framework named LSRM, which achieves more precise cross-modal alignment and fusion by leveraging multi-layer visual features, semantic relation projection, and inheritable cross-attention.

RemoteReasoner: Towards Unifying Geospatial Reasoning Workflow

Liang Yao (Hohai University), Pai Peng (Hohai University)

CodeObject DetectionSegmentationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageMultimodalityBenchmark

🎯 What it does: Proposed RemoteReasoner, a unified geospatial reasoning workflow capable of performing pixel-level, region-level, and contour-level reasoning tasks within a single model;

RENEW: Risk- and Energy-Aware Navigation in Dynamic Waterways

Mingi Jeong (Virginia Tech), Alberto Quattrini Li (Dartmouth)

CodeOptimizationPhysics Related

🎯 What it does: Designed a risk and energy consumption joint planner RENEW for Autonomous Surface Vehicles (ASVs), integrating dynamic current environments and safety constraints to generate safe and energy-optimal paths.

Rep Deep & Machine Learning: Exemplar-Free Continual Video Action Recognition via Slow-Fast Collaborative Learning

Xueyi Zhang (National University of Defense Technology), Huiping Zhuang (National University of Defense Technology)

CodeRecognitionKnowledge DistillationConvolutional Neural NetworkVideo

🎯 What it does: Studied sample-free continual video action recognition and proposed the slow-fast collaborative learning framework SFCL;

Res-Bench: Benchmarking the Robustness of Multimodal Large Language Models to Dynamic Resolution Input

Chenxu Li (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)

CodeTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBenchmark

🎯 What it does: Constructed the ResBench benchmark to systematically evaluate the robustness of multimodal large language models (MLLMs) across different image resolutions, and proposed novel continuous error metrics such as ACE/RCE;

ResMAS: Resilience Optimization in LLM-based Multi-agent Systems

Zhilun Zhou (Tsinghua University), Fengli Xu (Huawei)

CodeOptimizationGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextGraph

🎯 What it does: Propose ResMASβ€”a two-phase framework that first uses a reward model and RL to train LLMs to automatically generate robust topologies, then optimizes prompts based on the topology to enhance the robustness of large language model multi-agent systems under random failures.

Resource Efficient Sleep Staging via Multi-Level Masking and Prompt Learning

Lejun Ai (South China University of Technology), Rui Wang (Huazhong University of Technology)

CodeClassificationComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkTransformerPrompt EngineeringTime SeriesBiomedical Data

🎯 What it does: Studied the sleep staging task in resource-constrained environments, proposing a Mask-Aware Sleep Staging (MASS) framework based on multi-level masking and global prompt learning, which can achieve high-precision sleep stage classification using only about 10% of 30-second EEG signals.

Response Attack: Exploiting Contextual Priming to Jailbreak Large Language Models

Miao Ziqi (Shanghai Artificial Intelligence Laboratory), Jing Shao (Shanghai Artificial Intelligence Laboratory)

CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: This paper proposes a new attack framework called Response Attack (RA), which utilizes a mildly harmful response generated in the conversation as a contextual preface to induce large language models to generate non-compliant content.

Retaliatory Attacks Against Federated Unlearning via Data Leakage

Xinyi Sheng (University of Sydney), Sen Fu (University of Sydney)

CodeFederated LearningSafty and PrivacyImageTabular

🎯 What it does: This paper proposes novel retaliatory attacks against Federated Unlearning (FU), including Anti-Unlearning Attack (AUA) and Discrimination-Unlearning Attack (DUA).

Rethinking Crystal Symmetry Prediction: A Decoupled Perspective

Liheng Yu (University of Science and Technology of China), Pengkun Wang (University of Science and Technology of China)

CodeClassificationConvolutional Neural NetworkTransformerPhysics Related

🎯 What it does: Built the XRDecoupler framework, adopting a decoupled perspective to address sub-attribute confusion (SPC) in PXRD prediction.

Rethinking Explanation Evaluation Under the Retraining Scheme

Yi Cai (Freie UniversitΓ€t Berlin), Gerhard Wunder (Freie UniversitΓ€t Berlin)

CodeClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImage

🎯 What it does: Proposed and improved a retraining-based feature importance explanation evaluation method, identified and corrected the Sign problem, and introduced efficient evaluation schemes such as KAFT-C.

Rethinking Flow and Diffusion Bridge Models for Speech Enhancement

Dahan Wang (Nanjing University), Jing Lu (Nanjing University)

CodeRestorationDiffusion modelFlow-based ModelAudio

🎯 What it does: This paper unifies the theoretical frameworks of flow models and diffusion bridge models, and proposes an improved bridge model;

Rethinking Irregular Time Series Forecasting: A Simple Yet Effective Baseline

Xvyuan Liu (East China Normal University), Bin Yang (East China Normal University)

CodeComputational EfficiencyRepresentation LearningTime SeriesBiomedical Data

🎯 What it does: Proposes the APN framework, which adaptively generates high-quality patch representations through the Time-Aware Patch Aggregation (TAPA) module, and combines query aggregation with a shallow MLP to achieve efficient prediction for irregular multivariate time series.

Rethinking Label Consistency of In-Context Learning: An Implicit Transductive Label Propagation Perspective

Haoyang Chen (Beihang University), Junfan Chen (Beihang University)

CodeClassificationData SynthesisLarge Language ModelText

🎯 What it does: This paper reconsiders ICL as transductive learning, proposes a label propagation framework based on Bayesian transduction, and designs the TopK-SD data synthesis method to enhance the consistency of demonstration labels.

Rethinking Multi-Instance Learning Through Graph-Driven Fusion: A Dual-Path Approach to Adaptive Representation

Yu-Xuan Zhang (Southwest Jiaotong University), Mingxing Zhang (Southwest Jiaotong University)

CodeClassificationRecommendation SystemGraph Neural NetworkImageTextBiomedical Data

🎯 What it does: This paper proposes a graph-driven multi-instance learning framework called GDF-MIL, which employs an adaptive dual-path fusion strategy to enhance weakly supervised classification performance.

Rethinking Multimodal Point Cloud Completion: A Completion-by-Correction Perspective

Wang Luo (Sun Yat-sen University), Guocong Quan (Sun Yat-sen University)

CodeGenerationGraph Neural NetworkTransformerImageMultimodalityPoint Cloud

🎯 What it does: Generate complete 3D priors from single-view RGB images and perform correction in the feature space, followed by a three-stage process of dual-modal feature encoding, seed generation, and hierarchical refinement to achieve point cloud completion.

Rethinking Rainy 3D Scene Reconstruction via Perspective Transforming and Brightness Tuning

Qianfeng Yang (Dalian Polytechnic University), Jiyu Jin (Dalian Polytechnic University)

CodeRestorationGenerationData SynthesisRecurrent Neural NetworkGaussian SplattingImage

🎯 What it does: This study constructs a new 3D rainy scene reconstruction dataset called OmniRain3D and proposes an end-to-end REVR-GSNet framework, which can simultaneously achieve brightness recovery, rain removal, and 3D Gaussian Splatting reconstruction under rainy images;

Rethinking Saliency Maps: A Cognitive Human Aligned Taxonomy and Evaluation Framework for Explanations

Yehonatan Elisha (Tel Aviv University), Noam Koenigstein (Tel Aviv University)

CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage

🎯 What it does: Proposed the Reference-Frame Γ— Granularity (RFxG) two-axis framework, redefining the reference frame (point-to-point vs. contrastive) and granularity (fine-grained vs. group-level) of saliency maps, and introduced four new contrastive and group-level faithfulness metrics based on this framework.

Rethinking Target Label Conditioning in Adversarial Attacks: A 2D Tensor-Guided Generative Approach

Hangyu Liu (Zhejiang University), Donglin Wang (Westlake University)

CodeAdversarial AttackConvolutional Neural NetworkTransformerDiffusion modelGenerative Adversarial NetworkImage

🎯 What it does: This paper proposes a generative adversarial attack framework called TGAF, based on a two-dimensional semantic tensor, to achieve multi-target transferable adversarial attacks.

Rethinking the Dark Knowledge and Kullback-Leibler Divergence Loss in Knowledge Distillation Under Capacity Mismatching

Yingchao Wang (Beijing Institute of Technology), Weilun Fei (Beijing Institute of Technology)

CodeClassificationKnowledge DistillationConvolutional Neural NetworkImage

🎯 What it does: Proposed a knowledge distillation method called RCKD based on a relative confidence matrix to address the shortcomings of traditional KL distillation in terms of class ranking and gradient competition;

Rethinking the Reliability of Multi-agent System: A Perspective from Byzantine Fault Tolerance

Lifan Zheng (Zhejiang University), Yu Tian (Tsinghua University)

CodeTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: The study investigates the reliability of LLM-driven multi-agent systems (MAS) from a Byzantine fault tolerance perspective, and proposes a confidence-based weighted Byzantine fault tolerance consensus mechanism called CP-WBFT, enhancing system robustness in extreme Byzantine environments.

Rethinking the Sampling Criteria in Reinforcement Learning for LLM Reasoning: A Competence-Difficulty Alignment Perspective

Deyang Kong (Peking University), Wei Ye (Peking University)

CodeAI Code AssistantTransformerLarge Language ModelReinforcement LearningTextBenchmark

🎯 What it does: Proposes the CDAS sampling framework based on matching model capability with problem difficulty, improving sample efficiency and performance in RL training.