Conference on Neural Information Processing Systems Β· 2283 papers
Reasoning Models Hallucinate More: Factuality-Aware Reinforcement Learning for Large Reasoning Models
Junyi Li (National University of Singapore), Hwee Tou Ng (National University of Singapore)
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper studies the hallucination problem that arises during the fine-tuning of large models in reinforcement learning due to the focus on the final answer, and proposes a reward mechanism that embeds factual verification at each step of reasoningβFactuality-aware Step-wise Policy Optimization (FSPO)βto reduce the occurrence of hallucinations and improve reasoning accuracy.
Bao Nguyen (Chinese University of Hong Kong), Viet Anh Nguyen (Chinese University of Hong Kong)
CodeOptimizationComputational EfficiencyLarge Language ModelContrastive LearningText
π― What it does: EPIC is proposed, an ensemble planning framework based on contrastive learning, designed to dynamically match the most suitable reasoning methods for language model queries, thereby reducing computational costs while maintaining accuracy.
Rebalancing Contrastive Alignment with Bottlenecked Semantic Increments in Text-Video Retrieval
Jian Xiao (Hefei University of Technology), Richang Hong (Hefei University of Technology)
CodeRetrievalContrastive LearningVideoText
π― What it does: This paper proposes the GARE framework, which alleviates inter-modal tension and noisy negative samples in contrastive learning through a learnable incremental approach to text-video pairs.
Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models
Ling Li (Hong Kong University of Science and Technology), Jiaheng Wei (Hong Kong University of Science and Technology)
CodeRecognitionKnowledge DistillationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageMultimodality
π― What it does: The GLOBE system is proposed, utilizing a large visual-language model (LVLM) combined with multi-model distillation and reinforcement learning to geolocate social media images and generate interpretable reasoning paths.
π― What it does: The ReCon-GS framework is proposed, achieving fast and compact reconstruction and real-time rendering of online multi-view dynamic scenes.
π― What it does: For the object detection task, a ReCon framework is proposed, which incorporates region-guided correction and region-aligned cross-attention into a frozen structure of controllable diffusion models (such as ControlNet + Stable Diffusion) in real-time, enhancing the consistency and quality of synthetic samples and annotations without training.
Recurrent Memory for Online Interdomain Gaussian Processes
Wenlong Chen (Imperial College London), Yingzhen Li (Imperial College London)
CodeTime SeriesOrdinary Differential Equation
π― What it does: This paper proposes an online high-order polynomial projection (HiPPO) driven sparse variational Gaussian process model (OHSVGP) to maintain long-term memory in online learning scenarios.
Red-Teaming Text-to-Image Systems by Rule-based Preference Modeling
Yichuan Cao (Chinese Academy of Sciences), Yinpeng Dong (Tsinghua University)
CodeGenerationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringImageText
π― What it does: Designed and implemented a red team framework RPG-RT based on rule preference modeling, utilizing large language models to iteratively modify prompts and fine-tune through rule scoring and Direct Preference Optimization (DPO) to break through the diverse defense mechanisms of closed-source text-to-image systems;
Redefining Experts: Interpretable Decomposition of Language Models for Toxicity Mitigation
Zuhair Hasan Shaik (Mohamed bin Zayed University of Artificial Intelligence), Md Shad Akhtar (Indraprastha Institute of Information Technology Delhi)
CodeClassificationGenerationExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: This study investigates toxicity detection and generation in large language models, proposing a new interpretable intervention method called EigenShift, which precisely suppresses toxic outputs by utilizing the eigenvalue decomposition of the final linear layer.
π― What it does: By iteratively correcting the coupling in the discrete flow model, the total correlation caused by decomposition approximations is reduced, enabling high-quality generation in just a few steps or even a single step.
REDOUBT: Duo Safety Validation for Autonomous Vehicle Motion Planning
Shuguang Wang (City University of Hong Kong), Jianping Wang (City University of Hong Kong)
CodeAutonomous DrivingSafty and PrivacyFlow-based ModelTabular
π― What it does: The REDOUBT framework is proposed for dual safety verification of autonomous driving motion planning, which detects whether the input scene is out of training distribution (OOV) and assesses the uncertainty of planning decisions.
π― What it does: This paper proposes a method that utilizes 'reduction-based pseudo-labels' generated by a multi-branch auxiliary model to address the instance-dependent partial label learning (ID-PLL) problem, and trains the main classifier using these pseudo-labels.
π― What it does: We propose an unsupervised framework for detecting graphical anomaly distributions at test time, called RedOUT, which can extract key structural information from test graphs without modifying the parameters of the pre-trained model, thereby better distinguishing between ID and OOD graphs.
RefLoRA: Refactored Low-Rank Adaptation for Efficient Fine-Tuning of Large Models
Yilang Zhang (University of Minnesota), Georgios B. Giannakis (University of Minnesota)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningImageTextBenchmark
π― What it does: This paper proposes an improved low-rank adaptation method called RefLoRA, which enhances the fine-tuning efficiency and stability of large models by optimally reconstructing low-rank factors at each step.
Refusal Direction is Universal Across Safety-Aligned Languages
Xinpeng Wang (Ludwig Maximilian University of Munich), Barbara Plank (Ludwig Maximilian University of Munich)
CodeSafty and PrivacyTransformerLarge Language ModelTextMultimodality
π― What it does: This study investigates the generalizability of refusal mechanisms in large language models within multilingual environments, proposing and validating the transferability of the Refusal Direction across different languages.
Registration is a Powerful Rotation-Invariance Learner for 3D Anomaly Detection
Yuyang Yu (South China University of Technology), Shengfeng He (Singapore Management University)
CodeAnomaly DetectionPoint CloudBenchmark
π― What it does: The Reg2Inv framework is proposed, which jointly learns the point cloud registration process and anomaly detection tasks to obtain rotation-invariant and locally discriminative features; during the inference phase, the registration matrix is used to align the test point cloud with the prototype, and then feature comparison is performed to achieve object-level and point-level anomaly detection.
ReID5o: Achieving Omni Multi-modal Person Re-identification in a Single Model
Jialong Zuo (Huazhong University of Science and Technology), Changxin Gao (Huazhong University of Science and Technology)
CodeRecognitionRetrievalTransformerMixture of ExpertsImageTextMultimodality
π― What it does: Proposes the OM-ReID problem, constructs a high-quality dataset ORBench across five modalities (RGB, infrared, colored pencil, sketch, text), and designs a unified model ReID5o that can handle any combination of modalities;
π― What it does: The GLARE framework is proposed, which transforms the large-scale virtual screening (VS) problem into a Markov Decision Process (MDP), adaptively selecting candidate molecules for experimental labeling through reinforcement learning (RL).
Reinforcement Learning Finetunes Small Subnetworks in Large Language Models
Sagnik Mukherjee (University of Illinois Urbana-Champaign), Hao Peng (University of Illinois Urbana-Champaign)
CodeLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper studies the fine-tuning process of reinforcement learning in large language models, finding that RL only updates a very small proportion (5%-30%) of the parameter subnetwork, and the fine-tuning of this subnetwork can restore the performance and parameters of the fully fine-tuned model. It further analyzes the consistency of the subnetwork and the factors leading to sparse updates.
Reinforcement Learning for Out-of-Distribution Reasoning in LLMs: An Empirical Study on Diagnosis-Related Group Coding
Hanyin Wang (Mayo Clinic), Jimeng Sun (University of Illinois Urbana-Champaign)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningBiomedical DataElectronic Health Records
π― What it does: This paper proposes and trains the DRG-SAPPHIRE model for the medical coding task of DRG coding, utilizing large-scale reinforcement learning to achieve automated coding and generate interpretable clinical reasoning.
Reinforcement Learning for Reasoning in Large Language Models with One Training Example
Yiping Wang (University of Washington), yelong shen
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: In the paper, the authors propose the '1-shot RLVR' method, which uses only one training sample to stimulate the mathematical reasoning ability of large language models through reinforcement learning, significantly improving performance on multiple mathematical benchmarks.
Reinforcement Learning Teachers of Test Time Scaling
Edoardo Cetin (Sakana AI), Yujin Tang (Sakana AI)
CodeKnowledge DistillationLarge Language ModelReinforcement LearningText
π― What it does: A new Reinforcement Learning Teacher (RLT) framework is proposed, which trains language models to provide explanations for known answers, used for student distillation and RL cold start.
Reinforcing Spatial Reasoning in Vision-Language Models with Interwoven Thinking and Visual Drawing
Junfei Wu (Institute of Automation, Chinese Academy of Sciences), Tieniu Tan (Institute of Automation, Chinese Academy of Sciences)
CodeTransformerReinforcement LearningVision Language ModelImage
π― What it does: This paper proposes a reasoning paradigm for drawing operations in visual space, enabling visual-language models to perform spatial reasoning directly on images through drawing boxes and auxiliary lines.
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningDiffusion modelText
π― What it does: A new reasoning framework called Diffusion Chain of Lateral Thought (DCoLT) is proposed, which treats the intermediate steps in the reverse diffusion process of diffusion language models (DLM) as 'thinking' actions and employs reinforcement learning based on the correctness of the final answer as a reward.
π― What it does: A visual prompt-based image editing framework called RelationAdapter is proposed, which can efficiently extract and transfer editing intentions on the Diffusion Transformer;
Reliably detecting model failures in deployment without labels
Viet Nguyen (University of Toronto), Rahul Krishnan
CodeDomain AdaptationAnomaly DetectionImageTabularBiomedical DataElectronic Health Records
π― What it does: A post-deployment model performance degradation detection method called D3M is proposed, which requires no labels and no training data after training, and can automatically identify model failure during distribution drift.
Relieving the Over-Aggregating Effect in Graph Transformers
Junshu Sun (Institute of Computing Technology), Shuhui Wang (Institute of Computing Technology)
CodeGraph Neural NetworkTransformerGraph
π― What it does: The Wideformer method is proposed to address the over-aggregation problem in graph Transformers, significantly improving the model's efficiency in utilizing global information.
CodeMeta LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: Proposed and implemented the ReMA framework, which utilizes multi-agent reinforcement learning to decompose the reasoning process of large language models into two stages: meta-thinking and fine-grained reasoning, and further enhances reasoning capabilities through multi-round interactions.
REMI: Reconstructing Episodic Memory During Internally Driven Path Planning
Zhaoze Wang (University of Pennsylvania), Vijay Balasubramanian (University of Pennsylvania)
CodeRecurrent Neural NetworkImage
π― What it does: A framework for internal path planning based on MEC-Hippocampal networks, called REMI, is proposed, which can recall target locations through sensory cues and generate planned paths.
ReMindRAG: Low-Cost LLM-Guided Knowledge Graph Traversal for Efficient RAG
Yikuan Hu (Sichuan University), Chen Huang (National University of Singapore)
CodeRetrievalComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
π― What it does: This paper proposes REMINDRAG, a KG-RAG system that utilizes LLM-guided knowledge graph traversal, combining node exploration, node utilization, and memory replay.
π― What it does: Proposes the Region Encoder Network (REN), which directly extracts high-quality region vectors from a frozen patch encoder using point prompts, eliminating the expensive segmentation step.
RePIC: Reinforced Post-Training for Personalizing Multi-Modal Language Models
Yeongtak Oh (Seoul National University), Sungroh Yoon (Seoul National University)
CodeRecognitionGenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
π― What it does: A post-training framework based on reinforcement learning, RePIC, is designed to enhance the visual recognition and personalized generation capabilities of multimodal large language models in personalized image captioning.
CodeCompressionOptimizationTransformerLarge Language ModelText
π― What it does: A training-independent deep pruning method called ReplaceMe is proposed, which compresses the model by replacing consecutive Transformer blocks with linear transformations.
π― What it does: Reprogramming existing latent diffusion models (LDM) without training to achieve high-quality, high-efficiency high-resolution image generation.
RePO: Understanding Preference Learning Through ReLU-Based Optimization
Junkang Wu (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
CodeRecommendation SystemOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelText
π― What it does: A ReLU-based preference learning method called RePO is proposed, which uses a threshold Ξ³ to filter training samples, thereby achieving alignment with large language models without using Sigmoid weights or regularization.
Repo2Run: Automated Building Executable Environment for Code Repository at Scale
Ruida Hu (Harbin Institute of Technology), Cuiyun Gao (Harbin Institute of Technology)
CodeAI Code AssistantTransformerLarge Language ModelText
π― What it does: Proposes Repo2Run, an LLM-based agent for automating the construction of executable code environments, capable of generating runnable Dockerfiles for any Python repository and successfully executing unit tests.
RepoMaster: Autonomous Exploration and Understanding of GitHub Repositories for Complex Task Solving
Huacan Wang (Midea Group), Pin Lyu (Institute of Automation, Chinese Academy of Sciences)
CodeOptimizationAI Code AssistantTransformerLarge Language ModelAgentic AITextGraphBenchmark
π― What it does: The RepoMaster framework is proposed, which automatically searches, analyzes, and extracts core components of GitHub repositories, and autonomously explores and executes tasks with the assistance of LLMs, thereby completing complex real-world tasks.
π― What it does: Proposes the REG method, which synchronously injects noise and concatenates low-level image latent variables with high-level category tokens in diffusion training, achieving joint denoising of images and semantics;
Repurposing AlphaFold3-like Protein Folding Models for Antibody Sequence and Structure Co-design
Nianzu Yang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
CodeProtein Structure PredictionDiffusion modelBiomedical Data
π― What it does: Transform the AlphaFold3-like protein folding model into an antibody sequence-structure co-design model, incorporating a sequence diffusion head to achieve co-diffusion of CDR sequences and structures;
Repurposing Marigold for Zero-Shot Metric Depth Estimation via Defocus Blur Cues
Chinmay Talegaonkar (University of California San Diego), Nicholas Antipa (University of California San Diego)
CodeDepth EstimationDiffusion modelImage
π― What it does: In monocular scene settings, panoramic focused images taken with different apertures and defocused images are used to infer optimized pre-trained relative depth generated by Marigold, combined with a physical degradation model to obtain zero-shot, measurable depth estimation.
CodeTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
π― What it does: Train large language models to actively use search engines during the reasoning process, forming a complete 'think-search-retrieve-think-answer' cycle;
π― What it does: A plugin framework named ReservoirTTA is proposed for long-term adaptation during testing, which detects and routes samples from different domains by maintaining a set of domain-specific models (model pool) and performing style-based online clustering, achieving domain adaptation and preventing catastrophic forgetting.
Restoring Pruned Large Language Models via Lost Component Compensation
Zijian Feng (Nanyang Technological University), Kezhi Mao (Nanyang Technological University)
CodeRestorationTransformerLarge Language ModelText
π― What it does: A recovery method for pruned LLMs called RestoreLCC is proposed, which restores performance by compensating for lost components through comparative detection of key attention heads.
π― What it does: This paper proposes a global perception graph filtering framework GΒ²Former based on Transformer, which utilizes multi-channel bandpass filters to constrain global attention, thereby stabilizing the attention mechanism and enhancing the generalization ability of Graph Neural Networks (GNNs);
Rethinking Fair Federated Learning from Parameter and Client View
Kaiqi Guan (Wuhan University), Mang Ye (Wuhan University)
CodeFederated LearningImage
π― What it does: Designed and implemented FedPW, a federated learning framework that enhances fairness and overall performance through parameter pruning and adaptive weighted aggregation.
π― What it does: A new unsupervised Hebbian learning framework SPHeRe is designed and implemented, utilizing structural projection and orthogonal constraints to learn robust features, and achieving hierarchical pre-training of deep networks through a lightweight auxiliary module.
π― What it does: This paper studies the loss function of the diffusion bridge sampler, proposing the use of the reverse Kullback-Leibler loss with the log-derivative trick (rKL-LD) as a replacement for the Log Variance loss, and proving its theoretical advantages and experimental performance;
Rethinking Neural Combinatorial Optimization for Vehicle Routing Problems with Different Constraint Tightness Degrees
Fu Luo (Southern University of Science and Technology), Zhenkun Wang (National University of Singapore)
CodeOptimizationTransformerReinforcement LearningMixture of ExpertsTabular
π― What it does: This study investigates the generalization ability of neural combinatorial optimization models in vehicle routing problems with varying constraint tightness and proposes variable constraint training and a multi-expert module to enhance performance.
Rethinking Nighttime Image Deraining via Learnable Color Space Transformation
Qiyuan Guan (Dalian Polytechnic University), Jinshan Pan (Nanjing University of Science and Technology)
CodeRestorationTransformerImage
π― What it does: This paper proposes a high-quality nighttime rainy image de-raining dataset, HQ-NightRain, and designs a two-stage color space transformation network, CST-Net, based on a learnable color space converter (CSC) and implicit illumination guidance (IIG), for de-raining in the Y channel of the YCbCr color space.
π― What it does: Treating high-level features as charged particles, the Vlasov-Poisson system is used to model their collective behavior under a self-consistent electric field, and OOD detection and generalization are achieved by solving the boundary of the steady-state potential field and particle distribution.
Jan Quan (KU Leuven), Panagiotis Patrinos (KU Leuven)
CodeOptimizationTabular
π― What it does: This paper re-examines Principal Component Analysis (PCA) through the framework of Differential Convex (DC) duality, proposing three new DC dual formulations. It proves that if one function of the original problem is unitary invariant, then the corresponding dual is kernelizable and possesses out-of-sample generalization capability. Furthermore, it reveals the equivalence of the DC Algorithm (DCA) under the variance maximization objective with the simultaneous iteration/QR algorithm, and provides the corresponding kernelized robust PCA dual and its algorithm.
Rethinking Residual Distribution in Locate-then-Edit Model Editing
Xiaopeng Li (National University of Defence Technology), Jie Yu (National University of Defence Technology)
CodeTransformerLarge Language ModelText
π― What it does: The BLUE strategy is proposed, which enhances the effectiveness of the localization-reediting model editing method by directly calculating the residuals through updating only the first and last layers of the key layers, avoiding weight errors caused by residual distribution.
π― What it does: A CNN-RNN hybrid network is proposed for real-time object detection in event cameras, emphasizing fine-grained temporal modeling at low resolution stages and fusing features through multi-scale branches.
π― What it does: This paper proposes SwapGT, a node classification model based on Tokenized Graph Transformer, which primarily generates diverse token sequences through a novel token swapping operation and combines a Transformer backbone network with a center alignment loss for node representation learning.
Rethinking Verification for LLM Code Generation: From Generation to Testing
Zihan Ma (Xi'an Jiaotong University), Kai Chen (Shanghai AI Laboratory)
CodeGenerationAI Code AssistantLarge Language ModelTextBenchmark
π― What it does: Proposed and implemented SAGA (a human-LLM collaborative test case generation framework), and based on this, constructed the TCGBench and CodeCompass evaluation benchmarks;
Wei Liu (Shanghai Jiao Tong University), Hao Zhou (Institute for AI Industry Research at Tsinghua University)
CodeDrug DiscoveryTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: Designed and trained an agent RETRO-R1 based on a large language model (LLM), which interacts with a single-step model through reinforcement learning in multi-step retro-synthesis planning to dynamically construct synthesis pathways.
π― What it does: A worst-path optimization framework based on tree-structured MDP is proposed, and a search-free multi-step inverse synthesis planning method called InterRetro is implemented.
Luhuan Wu (Columbia University), John Patrick Cunningham
CodeDiffusion modelScore-based ModelTabular
π― What it does: A training-free reverse diffusion sequence Monte Carlo (RDSMC) sampler is proposed for sampling from unnormalized target distributions.
Revising and Falsifying Sparse Autoencoder Feature Explanations
George Ma (University of California), Somayeh Sojoudi (University of California)
CodeExplainability and InterpretabilityLarge Language ModelAuto EncoderTextChain-of-Thought
π― What it does: This study investigates how to improve the automatic interpretation of features from Sparse Autoencoders (SAE), proposing more refined structured explanations and a tree-based iterative generation method.
π― What it does: This paper re-examines the 1-peer index graph, proposes the k-peer index graph and null-cascade graph, and constructs a dynamic communication topology that can achieve finite-time convergence under any number of nodes while maintaining the exchangeability of the mixing matrix, aimed at enhancing the efficiency of decentralized learning.
π― What it does: A novel infrared-visible image fusion method called HCLFuse is proposed based on human cognitive principles, utilizing a multi-scale variational bottleneck encoder and a physics-guided diffusion model.
Revisiting Logit Distributions for Reliable Out-of-Distribution Detection
Jiachen Liang (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
π― What it does: A post-processing OOD detection method based on logit gap (LogitGap) is proposed, which distinguishes ID and OOD samples by utilizing the difference between the maximum logit and the average of the other logits, and further enhances performance by automatically selecting top-N logits.
Revisiting LRP: Positional Attribution as the Missing Ingredient for Transformer Explainability
Yarden Bakish, Lior Wolf
CodeExplainability and InterpretabilityTransformerImageText
π― What it does: A Layer-wise Relevance Propagation method considering position encoding (PA-LRP) is proposed for the interpretability of Transformers.
π― What it does: This paper proposes a multi-agent world model DIMA based on diffusion models, using sequential agent modeling to reduce complexity and improve the accuracy of global state transition predictions.
π― What it does: Proposes an orthogonal residual update mechanism that only retains the part of the module output that is orthogonal to the input stream to update the network.
π― What it does: In the era of Visual Foundation Models (VFM), systematic experiments on semi-supervised learning (SSL) are conducted, establishing a multi-task benchmark based on VTAB, and proposing a self-training framework V-PET achieved through efficient fine-tuning of multi-model parameters (PEFT) and the integration of VFM's pseudo-labels.
π― What it does: A novel linear GCN model called BoostGCN is proposed, which replaces the traditional Laplacian suppression aggregation by amplifying the important interactions of first-order neighbors to enhance recommendation effectiveness and training efficiency.
π― What it does: This paper proposes Reward Reasoning Models (RRM), transforming the reward model into a reasoning task that first performs chain-of-thought before providing rewards; it evolves reasoning capabilities through reinforcement learning and employs strategies such as ELO scoring or knockout tournaments in multi-response scenarios.
π― What it does: This paper proposes the framework of 'Reward-Oriented Causal Representation Learning (RO-CRL)' and designs a corresponding adaptive exploration algorithm to learn low-dimensional causal variables and graph structures from incomplete observational data through limited samples, ultimately optimizing downstream rewards within the available intervention space.
RF-Agent: Automated Reward Function Design via Language Agent Tree Search
Ning Gao (Beihang University), Yue Deng (Beihang University)
CodeRobotic IntelligenceLarge Language ModelReinforcement LearningAgentic AI
π― What it does: This paper proposes the RF-Agent framework, which combines language models with Monte Carlo Tree Search to automatically generate reward functions for high-performance low-level control tasks.
π― What it does: This paper proposes RFMPose, a category-level 6D object pose estimation framework based on Riemannian Flow Matching, which generates pose distributions on SE(3) using continuous probability flows.
RHYTHM: Reasoning with Hierarchical Temporal Tokenization for Human Mobility
Haoyu He (Northeastern University), Qi Wang
CodeTransformerLarge Language ModelPrompt EngineeringTime SeriesSequential
π― What it does: Proposed and implemented the RHYTHM framework, which combines a frozen pre-trained large language model (LLM) with hierarchical temporal tokenization and semantic prompts to predict human mobility trajectories.
CodeGenerationDrug DiscoveryFlow-based ModelSequentialBiomedical Data
π― What it does: The RiboFlow model is proposed, which can generate RNA molecules that satisfy structural and sequence consistency and have high binding affinity under the condition of given small molecule ligands.
RidgeLoRA: Matrix Ridge Enhanced Low-Rank Adaptation of Large Language Models
Junda Zhu (Beihang University), Qun Liu (Huawei Noah's Ark Lab)
CodeTransformerLarge Language ModelTextMultimodality
π― What it does: In the parameter-efficient fine-tuning of large language models, the RidgeLoRA architecture is proposed, which improves the parallel structure of LoRA to a series connection and adds a diagonal ridge term to enhance representational capacity.
π― What it does: A method for matching Lagrangian flows based on global diffeomorphisms, called DIFFEOCFM, is proposed to generate realistic samples on brain connectivity matrices (SPD/correlation matrices).
Right Question is Already Half the Answer: Fully Unsupervised LLM Reasoning Incentivization
Qingyang Zhang (Tianjin University), Yatao Bian (National University of Singapore)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningText
π― What it does: A completely unsupervised LLM inference incentive method called EMPO is proposed, which enhances the model's reasoning ability by minimizing semantic entropy.
CodeGraph Neural NetworkPoint CloudMeshBenchmarkPhysics Related
π― What it does: This study proposes a region interaction operator RIGNO based on graph neural networks for learning PDE solution operators on arbitrary geometric domains.
π― What it does: The FedPhoenix framework is proposed, which disrupts client-specific overfitting features by randomly resetting some parameters of the global model during each communication round, thereby guiding the global model to learn more general features.
π― What it does: This paper proposes a risk-aware constrained reinforcement learning framework using Optimized Certainty Equivalence (OCE), which can simultaneously consider tail risk in rewards and constraints.
Xihong Su (University of New Hampshire), Marek Petrik (University of New Hampshire)
CodeOptimizationReinforcement LearningTabular
π― What it does: This paper proposes two model-free Q-learning algorithms for risk aversion, targeting the exponential risk measure (ERM) and exponential value at risk (EVaR) objectives under the total-reward criterion, and provides rigorous convergence proofs; it also presents experimental results on two classic discrete domains (Cliff-Walking and Gambler's Ruin).
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper studies the introduction of a token-level direct preference optimization method Ra-DPO using nested risk measures during the alignment process of large language models, aiming to suppress the risk of the model deviating from the reference model while maintaining alignment performance.
RL Tango: Reinforcing Generator and Verifier Together for Language Reasoning
Kaiwen Zha (Massachusetts Institute of Technology), Dina Katabi (Massachusetts Institute of Technology)
CodeTransformerLarge Language ModelReinforcement LearningText
π― What it does: Through the alternating training of the generator and validator in the RL framework TANGO, the multi-step reasoning ability of LLM is jointly enhanced, avoiding reliance on fixed or SFT-trained reward models;
RobIA: Robust Instance-aware Continual Test-time Adaptation for Deep Stereo
Jueun Ko (Ewha Womans University), Dongbo Min (Ewha Womans University)
CodeDepth EstimationDomain AdaptationAutonomous DrivingMixture of ExpertsImage
π― What it does: A robust instance-aware continuous testing adaptation framework for stereo depth estimation (RobIA) is proposed, which achieves online adaptation to dynamic domain shifts by introducing an input-aware mixture of experts module and a robust AdaptBN teacher.
π― What it does: A physical information-based world model for RoboScape is proposed, integrating RGB video generation, temporal depth prediction, and adaptive keypoint dynamics to enhance the physical realism and action controllability of robot scene videos.
Robot-R1: Reinforcement Learning for Enhanced Embodied Reasoning in Robotics
Dongyoung Kim (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality
π― What it does: The ROBOT-R1 framework is proposed, which trains large visual language models through reinforcement learning to enhance the embodied reasoning capabilities in robot control.
Marshal Arijona Sinaga, Samuel Kaski (Aalto University)
CodeAnomaly DetectionOptimizationComputational EfficiencyTabularTime SeriesFinance Related
π― What it does: A robust computationally aware Gaussian process model (RCaGP) has been developed, which can simultaneously address the uncertainty caused by approximation in large-scale data and the robustness issues brought by outliers.
π― What it does: This paper proposes the Cross-Scene Spatial Reasoning and Localization (CSSRG) task and introduces the matching-localization two-stage CoRe framework to address the challenges of large-scale scene retrieval and fine-grained text-object alignment.
π― What it does: A lossless watermarking framework ALIGNED-IS for autoregressive audio generation models is proposed, addressing the issue of label mismatch caused by re-encoding;
Robust Ego-Exo Correspondence with Long-Term Memory
Yijun Hu (University of Chinese Academy of Sciences), Libo Zhang (Rochester Institute of Technology)
CodeObject DetectionSegmentationTransformerMixture of ExpertsVideoBenchmark
π― What it does: A robust perspective correspondence framework LMβEEC based on SAMβ―2 is proposed, specifically addressing the object-level correspondence and segmentation issues in self-perspective (ego) and external perspective (exo) videos synchronized over time.
Robust Explanations of Graph Neural Networks via Graph Curvatures
Yazheng Liu (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
CodeOptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
π― What it does: A method is proposed to enhance the robustness of graph neural network (GNN) explanations through graph curvature, aiming to improve the credibility of GNNs in high-risk applications.
π― What it does: This paper studies and proposes a robust framework for Graph Condensation (MRGC) to mitigate attacks on node features, structure, and labels.
Robust Hallucination Detection in LLMs via Adaptive Token Selection
Mengjia Niu (Imperial College London), Guansong Pang (Singapore Management University)
CodeAnomaly DetectionTransformerLarge Language ModelText
π― What it does: This paper proposes a framework named HaMI, which adaptively selects the tokens that best represent hallucinations using multi-instance learning (MIL) within the internal representations of LLMs, and trains a hallucination detector based on this; it also incorporates predictive uncertainty information to enhance the representation.
Robust Minimax Boosting with Performance Guarantees
Santiago Mazuelas (Basque Center of Applied Mathematics), Veronica Alvarez
CodeOptimizationTabular
π― What it does: A Robust Maximum Value Minimization Boosting (RMBoost) method is proposed, which directly minimizes the error probability and provides finite sample performance guarantees.
Robust Policy Expansion for Offline-to-Online RL under Diverse Data Corruption
Longxiang He (Tsinghua University), Li Shen (Shenzhen Campus of Sun Yat-sen University)
CodeReinforcement Learning
π― What it does: A robust strategy expansion method RPEX has been developed to combat data corruption in offline-to-online reinforcement learning (O2O RL).
RobustMerge: Parameter-Efficient Model Merging for MLLMs with Direction Robustness
Fanhu Zeng (Institute of Automation, Chinese Academy of Sciences), Hao Tang (Peking University)
CodeGenerationOptimizationTransformerLarge Language ModelMixture of ExpertsVision Language ModelMultimodality
π― What it does: This paper proposes a training-free, data-free parameter-efficient model merging method called RobustMerge, specifically designed for multi-modal large language models (MLLMs), which can merge multi-task expert models into a single multi-task model.
π― What it does: A robust optimization-based diffusion sampling framework RODS is proposed, which can actively detect and correct hallucinated outputs during the generation process.
Role-aware Multi-agent Reinforcement Learning for Coordinated Emergency Traffic Control
Ming Cheng (Central South University), Senzhang Wang (Central South University)
CodeAutonomous DrivingRecurrent Neural NetworkGraph Neural NetworkReinforcement LearningGraphTime Series
π― What it does: A Role-aware Multi-agent Traffic Control (RMTC) framework is proposed, which jointly learns the collaborative decision-making of traffic lights, emergency vehicles (EMV), and regular vehicles (REV) to achieve rapid passage for EMVs while maintaining overall traffic smoothness.