ICLR 2026 Papers — Page 38
International Conference on Learning Representations · 5356 papers
R1-Code-Interpreter: LLMs Reason with Code via Supervised and Multi-stage Reinforcement Learning
Yongchao Chen (MIT), Chuchu Fan (MIT)
OptimizationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Trained and evaluated a Code Interpreter LLM for general tasks using a multi-round text/code generation framework combined with supervised learning and multi-stage Group Relative Policy Optimization (GRPO) reinforcement learning.
R1-Reward: Training Multimodal Reward Model Through Stable Reinforcement Learning
YiFan Zhang, Liang Wang (Chinese Academy of Sciences)
Reinforcement Learning from Human FeedbackReinforcement LearningMultimodality
🎯 What it does: Proposed a new multimodal reward model R1-Reward, trained using the StableReinforce algorithm to improve multimodal reward evaluation performance.
R2-Dreamer: Redundancy-Reduced World Models without Decoders or Augmentation
Naoki Morihira (Honda R&D Co., Ltd.), Tatsuya Harada (University of Tokyo)
Recurrent Neural NetworkReinforcement LearningContrastive LearningWorld ModelImage
🎯 What it does: Propose R2-Dreamer, a fully decoder-free and data-augmentation-free image reinforcement learning framework that employs internal redundancy reduction (Barlow Twins) regularization to learn high-quality representations.
R2PS: Worst-Case Robust Real-Time Pursuit Strategies under Partial Observability
Runyu Lu (University of Chinese Academy of Sciences), Dongbin Zhao (Chinese Academy of Sciences)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: Proposed a worst-case robust real-time pursuit strategy (R2PS) under partial observability and dynamically changing graph structures, and implemented its cross-graph reinforcement learning training framework;
R4: Nested Reasoning-Retrieval for Reward Modeling in Role-Playing Agents
Renzhi Wang (Nanjing University of Aeronautics and Astronautics), Piji Li (Tencent)
Reinforcement Learning from Human FeedbackReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a unified framework R4, endowing the reward model and role-playing agent with reasoning and retrieval capabilities, achieving context-based multi-dimensional assessment and generation.
RACE Attention: A Strictly Linear-Time Attention for Long-Sequence Training
Sahil Joshi (Rice University), Anshumali Shrivastava (Rice University)
ClassificationComputational EfficiencyTransformerImageText
🎯 What it does: Proposes RACE Attention, a linear-time, memory-efficient alternative to Softmax Attention that supports training with ultra-long contexts;
RADAR: Learning to Route with Asymmetry-aware Distance Representations
Hang Yi (Singapore Management University), Zhiguang Cao (Singapore Management University)
Autonomous DrivingOptimizationTransformerGraph
🎯 What it does: Proposes the RADAR framework to address the modeling and reasoning of asymmetric distance matrices in neural vehicle routing planning
RADAR: Reasoning-Ability and Difficulty-Aware Routing for Reasoning LLMs
Nigel Fernandez (University of Massachusetts Amherst), Zichao Wang (University of Massachusetts Amherst)
OptimizationExplainability and InterpretabilityComputational EfficiencyLarge Language ModelTextBenchmark
🎯 What it does: Proposed the RADAR framework for intelligent routing between large language model (LLM) configurations (model size + inference budget) to achieve the optimal trade-off between performance and cost.
Radiometrically Consistent Gaussian Surfels for Inverse Rendering
Kyu Beom Han (Korea Advanced Institute of Science and Technology), Sung-eui Yoon
GenerationComputational EfficiencySupervised Fine-TuningNeural Radiance FieldGaussian SplattingImage
🎯 What it does: This paper proposes an inverse rendering framework called RadioGS based on Gaussian surface radiance consistency, which can accurately model indirect illumination in unobserved directions and achieve efficient relighting.
RAEE: A Robust Retrieval-Augmented Early Exit Framework for Efficient Inference
LIANMING HUANG, Chun Jason Xue (National Taiwan University)
RetrievalComputational EfficiencyKnowledge DistillationTextRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval-based early exit framework RAEE, which dynamically determines the exit layer during inference by leveraging an offline-built exit database, using exit information from similar samples to achieve inference acceleration and improve model accuracy.
RAG4DMC: Retrieval-Augmented Generation for Data-Level Modality Completion
Ningxin He (Nankai University), Tiegang Gao (Nankai University)
GenerationData SynthesisMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a RAG4DMC framework based on retrieval-augmented generation to complete missing modalities in multi-modal data at the data level.
RAIN-Merging: A Gradient-Free Method to Enhance Instruction Following in Large Reasoning Models with Preserved Thinking Format
Zhehao Huang (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a gradient-free training method called RAIN-Merging, which can merge instruction tuning models (ITM) with large reasoning models (LRM) while preserving the LRM's 'thinking' format and reasoning quality.
Rainbow Padding: Mitigating Early Termination in Instruction-Tuned Diffusion LLMs
BumJun Kim, Albert No (Yonsei University)
GenerationTransformerSupervised Fine-TuningDiffusion modelTextBenchmark
🎯 What it does: Proposed and studied the 'Rainbow Padding' solution to address the early termination problem caused by <eos> overload in instruction-tuned diffusion-based large language models.
RainPro-8: An Efficient Deep Learning Model to Estimate Rainfall Probabilities Over 8 Hours
Rafael Pablos Sarabia (Aarhus University), Ira Assent (Aarhus University)
Computational EfficiencyConvolutional Neural NetworkTransformerMultimodalityTime Series
🎯 What it does: Propose an efficient deep learning model, RainPro-8, for predicting the probability distribution of precipitation over the next 8 hours within the European region, integrating multi-source data fusion and probabilistic output;
Random Anchors with Low-rank Decorrelated Learning: A Minimalist Pipeline for Class-Incremental Medical Image Classification
Xinyao Wu (Chinese University of Hong Kong), Raymond Kai-yu Tong (Chinese University of Hong Kong)
ClassificationTransformerImageBiomedical DataBenchmark
🎯 What it does: Propose an extremely simple RA-LDL framework that achieves class-incremental learning for medical image classification by leveraging pre-trained ViT features through random anchor projection, low-rank residual calibration, and closed-form Ridge regression.
Random Controlled Differential Equations
Francesco Piatti (Imperial College London), William F. Turner (Imperial College London)
ClassificationTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a time series learning framework based on stochastic control differential equations, constructing two stochastic dynamical models RF-CDE and R-RDE, which correspond to RBF-lifted signature kernels and rough signature kernels in the infinite width limit;
Random Label Prediction Heads for Studying Memorization in Deep Neural Networks
Marlon Becker (University of Muenster), Benjamin Risse (University of Muenster)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: Measure and regularize memorization in deep networks by inserting random label prediction heads (RLP-heads) at various layers.
Random Policy Valuation is Enough for LLM Reasoning with Verifiable Rewards
Haoran He (Hong Kong University Of Science And Technology), Ling Pan (Hong Kong University Of Science And Technology)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a RL method called ROVER based on uniform random strategy value assessment, which directly uses the Q-values of the random strategy for greedy or Softmax-based action selection, thereby enhancing mathematical reasoning and diversity in LLMs.
Random Spiking Neural Networks are Stable and Spectrally Simple
Ernesto Araya (Ludwig-Maximilians-Universität München), Gitta Kutyniok (Ludwig-Maximilians-Universität München)
ClassificationSpiking Neural NetworkImage
🎯 What it does: This paper studies the noise sensitivity and stability of discrete-time LIF pulse neural network classifiers through Boolean function analysis, and provides corresponding probability upper bounds.
Random-projection ensemble dimension reduction
Wenxing Zhou (University of Edinburgh), Timothy Ivor Cannings (University of Edinburgh)
Representation Learning
🎯 What it does: Propose a dimensionality reduction method based on random projection ensemble (RPEDR), which selects the best-performing projection within multiple projection groups, aggregates their outer products, and obtains a low-dimensional subspace through singular value decomposition;
Randomization Boosts KV Caching, Learning Balances Query Load: A Joint Perspective
Fangzhou Wu (University of Wisconsin-Madison), Qiuyi Zhang (Google DeepMind)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a unified mathematical model for KV cache and query routing, and designs two algorithms: randomized leaf token eviction (RLT) and learning-based greedy routing (LBGR), significantly improving inference efficiency in a multi-LLM server environment.
Randomized Antipodal Search Done Right for Data Pareto Improvement of LLM Unlearning
Ziwen Liu (Rice University), Zhaozhuo Xu (Stevens Institute of Technology)
RetrievalOptimizationData-Centric LearningLarge Language ModelText
🎯 What it does: Proposed a retrieval method called RASLIK based on random reverse search, which efficiently and with low variance selects the forget set and retain set in unlearning of large language models, thereby enhancing the Pareto performance of forgetting-keeping.
Rank-GRPO: Training LLM-based Conversational Recommender Systems with Reinforcement Learning
Yaochen Zhu (University of Virginia), Nathan Kallus (University of Virginia)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the ConvRec-R1 framework, which employs a two-phase training approach: first, generating high-quality, catalog-based demonstration data through the Remap-Reflect-Adjust pipeline for behavior cloning, and then aligning the conversational recommendation model using Rank-GRPO for reinforcement learning.
RankFlow: Property-aware Transport for Protein Optimization
Lu Yu (La Trobe University), Ramana Rao Kompella (La Trobe University)
OptimizationDrug DiscoveryTransformerLarge Language ModelFlow-based ModelBiomedical Data
🎯 What it does: This paper proposes RankFlow, a conditional flow model that predicts functional assay values of protein variants by refining the embeddings of pre-trained protein language models (PLM).
RAP: 3D Rasterization Augmented End-to-End Planning
Lan Feng (EPFL), Alexandre Alahi (EPFL)
Data SynthesisAutonomous DrivingAdversarial AttackTransformerContrastive Learning
🎯 What it does: Proposes the RAP framework, which leverages lightweight 3D rasterization and Raster-to-Real feature alignment to perform large-scale data augmentation and training for end-to-end driving models.
Rapid Training of Hamiltonian Graph Networks Using Random Features
Atamert Rahma (Technical University of Munich), Felix Dietrich (Technical University of Munich)
Computational EfficiencyGraph Neural NetworkGraphBenchmarkPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposed a Hamiltonian graph network based on random features (RF-HGN), achieving fast training without gradient descent.
RAPID$^3$: Tri-Level Reinforced Acceleration Policies for Diffusion Transformer
Wangbo Zhao (National University of Singapore), Yang You (National University of Singapore)
Computational EfficiencyTransformerReinforcement LearningDiffusion modelImage
🎯 What it does: This paper proposes RAPID 3, a three-layer reinforcement acceleration strategy tailored for Diffusion Transformers, achieving adaptive acceleration per image through three lightweight policy heads;
RAR: Reversing Visual Attention Re-Sinking for Unlocking Potential in Multimodal Large Language Models
Zhehan Kan (Tsinghua University), Wenming Yang (Tsinghua University)
Explainability and InterpretabilityComputational EfficiencySupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Investigate the root causes of poor performance in the output layer of multimodal large language models and propose a parameter-agnostic Sink Attention Dynamic Sparsification (SADS) framework;
RAS: Retrieval-And-Structuring for Knowledge-Intensive LLM Generation
Pengcheng Jiang (University of Illinois Urbana Champaign), Jiawei Han (University of Illinois Urbana Champaign)
RetrievalExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Construct and utilize a problem-specific knowledge graph for multi-round retrieval and reasoning to enhance the inference accuracy of large language models in knowledge-intensive tasks.
RATE-DISTORTION OPTIMIZED PRAGMATIC COMMUNICATION FOR COLLABORATIVE PERCEPTION
Genjia Liu (Shanghai Jiao Tong University), Siheng Chen (Shanghai Jiao Tong University)
CompressionAutonomous DrivingOptimizationImageVideo
🎯 What it does: Propose a pragmatic rate-distortion theory based on information theory and implement the RDcomm framework, significantly enhancing task performance and communication efficiency in multi-agent collaborative perception.
Rating Quality of Diverse Time Series Data by Meta-learning from LLM Judgment
Shunyu Wu (Sun Yat-sen University), See-Kiong Ng (National University of Singapore)
Data-Centric LearningMeta LearningLarge Language ModelPrompt EngineeringTime Series
🎯 What it does: This paper proposes the TSRating framework, which uses LLM to comparatively evaluate the quality of multi-domain time series data and trains a Meta learning TSRater model for efficient quality scoring.
RAVEN: End-to-end Equivariant Robot Learning with RGB Cameras
David Klee (Northeastern University), Robin Walters (Northeastern University)
Robotic IntelligenceConvolutional Neural NetworkTransformerFlow-based ModelRectified FlowImage
🎯 What it does: Proposes RAVEN, an end-to-end SE(3) equivariant robotic learning framework capable of performing closed-loop manipulation tasks using only RGB camera inputs.
RAVENEA: A Benchmark for Multimodal Retrieval-Augmented Visual Culture Understanding
Jiaang Li (University of Copenhagen), Serge Belongie (University of Copenhagen)
RetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed and released the RAVENEA benchmark, integrating cultural center visual question answering (cVQA) and cultural information image captioning (cIC) tasks, and achieving retrieval-augmented visual cultural understanding evaluation through manually annotated Wikipedia documents.
RayI2P: Learning Rays for Image-to-Point Cloud Registration
Xinjun Li (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
Pose EstimationAutonomous DrivingTransformerImagePoint Cloud
🎯 What it does: Proposed a light-based image-point cloud registration framework that predicts 3D ray bundles corresponding to each image patch and directly estimates camera pose using a differentiable ray-guided regression module.
RCPU: Rotation-Constrained Error Compensation for Structured Pruning of Large Language Models
Shuichiro Haruta (KDDI Research), Mori Kurokawa (KDDI Research)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed a method called Rotational Constraint-based Error Compensation (RCPU) for large language models after structured pruning
RD-HRL: Generating Reliable Sub-Goals for Long-Horizon Sparse-Reward Tasks
Yixiang Shan (Jilin University), Yi Chang (Jilin University)
Reinforcement LearningBenchmark
🎯 What it does: Propose RD-HRL, which extracts transition regions in offline goal-conditioned reinforcement learning through a reliability-driven decision mechanism, providing high-level planners with decision-level goals to achieve more reliable subgoal selection and improve performance in long-horizon sparse reward tasks.
RE-PO: Robust Enhanced Policy Optimization as a General Framework for LLM Alignment
Xiaoyang Cao (Massachusetts Institute of Technology), Chao Yu (Tsinghua University)
Reinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: Propose a robust reinforcement learning framework called RE-PO, which utilizes EM iterations to simultaneously estimate the reliability of each preference label and adaptively weight the training loss, thereby improving the alignment effectiveness of large language models in noisy label environments.
REA-RL: Reflection-Aware Online Reinforcement Learning for Efficient Reasoning
Hexuan Deng (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes REA-RL, a reflection-aware online reinforcement learning framework that utilizes small reflection models for online sequential revision and incorporates reflection rewards to reduce overthinking in large-scale reasoning models (LRM), while maintaining or even improving reasoning accuracy.
ReactDance: Hierarchical Representation for High-Fidelity and Coherent Long-Form Reactive Dance Generation
Jingzhong Lin (East China Normal University), Gaoqi He (East China Normal University)
GenerationDiffusion modelVideoAudio
🎯 What it does: Propose a diffusion model based on a hierarchical latent space that can generate high-fidelity, long-duration reaction dance sequences under the condition of given lead dancer actions and music.
ReactID: Synchronizing Realistic Actions and Identity in Personalized Video Generation
Wei Li (University of Science and Technology of China), Tao Mei (HiDream.ai Inc)
GenerationTransformerLarge Language ModelDiffusion modelVideoText
🎯 What it does: Propose the ReactID framework to address the trade-off between identity consistency and action realism in personalized video generation.
Read the Room: Video Social Reasoning with Mental-Physical Causal Chains
Lixing Niu (Peking University), Lifeng Fan (BIGAI)
Large Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: This paper constructs a high-quality video social reasoning benchmark, R3-Bench, and generates a large-scale training set, R3-FDT, through an automated pipeline to evaluate and enhance the psychophysical causal reasoning capabilities of multimodal models.
Reading Images Like Texts: Sequential Image Understanding in Vision-Language Models
Yueyan Li (Beijing University of Posts and Telecommunications), Xiaojie Wang (Beijing University of Posts and Telecommunications)
CompressionKnowledge DistillationRepresentation LearningTransformerMultimodality
🎯 What it does: Analyze the processing of VLM visual encoders layer by layer through logit lens, study the two-stage attribute recognition and semantic disambiguation in the 'what' path, and the spatial geometric structure of 2D RoPE in the 'where' path, and propose instruction-agnostic Run-Length Encoding token compression and RoPE Scaling position enhancement methods based on this.
Readout Representation: Redefining Neural Codes by Input Recovery
Shunsuke Onoo (Kyoto University), Yukiyasu Kamitani (Kyoto University)
Representation LearningConvolutional Neural NetworkTransformerImageText
🎯 What it does: Propose a readout representation framework, defining representation as a set of information recoverable from features, and verifying the recoverable range of a single input in the feature space through feature inversion.
Real-Time Motion-Controllable Autoregressive Video Diffusion
Kesen Zhao (University of Science and Technology of China), Hanwang Zhang (University of Science and Technology of China)
GenerationReinforcement LearningDiffusion modelFlow-based ModelVideoTextMultimodalityStochastic Differential Equation
🎯 What it does: Proposed a few-step autoregressive video diffusion model called AR-Drag based on reinforcement learning, capable of achieving real-time image-to-video generation and motion control
Real-Time Reasoning Agents in Evolving Environments
Yule Wen (Tsinghua University), Hao Zhu (Stanford University)
Computational EfficiencyTransformerLarge Language ModelAgentic AIBenchmark
🎯 What it does: Propose the real-time reasoning problem, develop the Real-Time Reasoning Gym, and introduce the AgileThinker model that combines reactive and planning-based reasoning
Real-Time Robot Execution with Masked Action Chunking
Haoxuan Wang (University of Illinois Chicago), Gaowen Liu (Cisco Research)
Computational EfficiencyRobotic IntelligenceReinforcement LearningFlow-based Model
🎯 What it does: This paper proposes the REMAC method, which improves the reliability of real-time robotic execution under asynchronous inference through masked action blocking and prefix preservation sampling.
REAL: Reading Out Transformer Activations for Precise Localization in Language Model Steering
Li-Ming Zhan (Hong Kong Polytechnic University), Xiao-Ming Wu (Hong Kong Polytechnic University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningText
🎯 What it does: This paper proposes a framework called REAL for behavior regulation of large language models during inference. It first performs nonlinear decomposition of activations for each attention head or layer using a vector quantized autoencoder (VQ-AE), and learns behavior-related latent subspaces through supervised contrastive loss. Subsequently, it models discrete code sequences with a self-regressive prior, calculates behavior relevance scores for each module, and selects and activates them with importance-weighted injection.
RealPDEBench: A Benchmark for Complex Physical Systems with Real-World Data
Peiyan Hu (Westlake University), Tailin Wu (Westlake University)
Convolutional Neural NetworkTransformerTabularTime SeriesBenchmarkPhysics Related
🎯 What it does: Propose and release the RealPDEBench benchmark, integrating real experimental measurements with numerical simulation data, defining three categories of training tasks, nine evaluation metrics, and benchmarking ten scientific ML models.
Realtime Video Frame Interpolation using One-Step Diffusion Sampling
Yongrui Ma (ByteDance Inc.), Tianfan Xue (ByteDance Inc.)
GenerationComputational EfficiencyConvolutional Neural NetworkTransformerDiffusion modelOptical FlowVideo
🎯 What it does: This paper proposes a real-time video frame interpolation framework called RDVFI, which generates sparse potential keyframes using one-time diffusion sampling, solves high-order continuous pixel trajectories, and interpolates input frames through optical flow, ultimately synthesizing high-quality intermediate frames at high resolution.
REAP the Experts: Why Pruning Prevails for One-Shot MoE compression
Mike Lasby (Cerebras Systems Inc.), Vithursan Thangarasa (University of Calgary)
CompressionMixture of ExpertsText
🎯 What it does: Proposed a novel expert pruning method (REAP) based on routing weights and expert activation norms, achieving compression and evaluation on various scales of SMoE language models.
reAR: Rethinking Visual Autoregressive Models via Token-wise Consistency Regularization
Qiyuan He (National University of Singapore), Angela Yao (National University of Singapore)
GenerationComputational EfficiencyRepresentation LearningTransformerImage
🎯 What it does: Propose a training regularization method for visual autoregressive models called reAR, addressing performance bottlenecks caused by inconsistency between the generator and tokenizer
Reasoned Safety Alignment: Ensuring Jailbreak Defense via Answer-Then-Check
Chentao Cao (ByteDance Seed), Hang Li (ByteDance Seed)
Safty and PrivacyLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposed the 'answer-first-then-verify' safety alignment strategy and constructed the ReSA dataset with 80K samples to train LLMs to plan answers and check safety before generating responses, thereby defending against jailbreak attacks.
Reasoning as Representation: Rethinking Visual Reinforcement Learning in Image Quality Assessment
Shijie Zhao (ByteDance Inc.), Jian Zhang (Peking University)
Domain AdaptationRepresentation LearningReinforcement LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: This paper re-examines the role of reasoning-based multimodal large language models in image quality assessment (IQA), proposing two efficient frameworks: Reasoning-Aligned Cross-Domain Training (RACT) via reinforcement learning and inference-free Reasoning-Aligned Lightweight IQA (RALI), significantly enhancing cross-domain generalization capabilities;
Reasoning Boosts Opinion Alignment in LLMs
Frédéric Berdoz (ETH Zurich), Roger Wattenhofer (ETH Zurich)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTabularBenchmarkChain-of-Thought
🎯 What it does: A reinforcement learning-driven inference framework that directly learns individual political views from survey data and generates structured reasoning and answers.
Reasoning in Space via Grounding in the World
Yiming Chen (Westlake University), Peidong Liu (Westlake University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityPoint CloudMeshBenchmarkChain-of-Thought
🎯 What it does: Propose the GS-Reasoner framework, which utilizes a semantic-geometric hybrid 3D scene representation to achieve autoregressive 3D visual localization and integrates localization into a chain-of-thought process for spatial reasoning.
Reasoning Language Model Inference Serving Unveiled: An Empirical Study
Qi Li (Hong Kong University of Science and Technology), Xiaowen Chu (Hong Kong University of Science and Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper conducts a systematic empirical study on reasoning-phase large language models (RLLMs) during inference, proposing the ASU evaluation framework and ASU-Perf benchmark suite. It analyzes differences between RLLMs and traditional LLMs in inference services across dimensions such as memory consumption, straggler effects, adaptive runtime, and domain preferences. Subsequently, it experimentally validates the effectiveness of mainstream LLM optimization techniques (model weight quantization, KV cache quantization, prefix caching, and speculative decoding) on RLLMs, and finally verifies the universality of the results under real-world workloads following a Gamma distribution.
Reasoning Models Can be Accurately Pruned Via Chain-of-Thought Reconstruction
Ryan Lucas (LinkedIn), Rahul Mazumder (LinkedIn)
CompressionComputational EfficiencyTextChain-of-Thought
🎯 What it does: This paper proposes the 'Reasoning-Aware Compression (RAC)' method, which uses chain-of-thought (CoT) activations generated during inference for calibration in model pruning, thereby compressing inference models in one step while maintaining high accuracy.
Reasoning on Time-Series for Financial Technical Analysis
Kelvin J.L. Koa (National University of Singapore), Tat-Seng Chua (National University of Singapore)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningTime SeriesFinance RelatedChain-of-Thought
🎯 What it does: Propose the Verbal Technical Analysis (VTA) framework, integrating natural language reasoning with time series models to achieve interpretable multi-step stock price prediction.
Reasoning or Retrieval? A Study of Answer Attribution on Large Reasoning Models
Yuhui Wang (Stony Brook University), Ting Wang (Stony Brook University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Investigate the answer generation mechanism of large-scale reasoning models (LRM), demonstrating that answers are generated both through chain-of-thought (CoT) reasoning and internal retrieval memory, and verifying the coexistence of these two mechanisms via joint perturbation experiments.
Reasoning Scaffolding: Distilling the Flow of Thought from LLMs
Xiangyu Wen (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
Knowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the Reasoning Scaffolding framework, which uses discrete interpretable semantic signals rather than directly copying text for knowledge distillation in LLMs.
Reasoning with Sampling: Your Base Model is Smarter Than You Think
Aayush Karan (Harvard University), Yilun Du (Harvard University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a sampling algorithm called 'Power Sampling' based on the probability distribution of a foundational large language model (LLM). By applying power weighting to the probabilities of the foundational model and utilizing Metropolis-Hastings sampling, the algorithm enhances inference performance without training or a validator during inference.
Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning
Yunhao Gou (Southern University of Science and Technology), Yu Zhang (Southern University of Science and Technology)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed the RAPID framework, which separates perception and reasoning in multi-modal large language models (MLLMs). The framework uses MLLM to generate visual descriptions and temporary answers, which are then passed as text context to a powerful text LLM for reasoning. Visual Perception Optimization (VPO) aligns the MLLM's visual descriptions with the final reasoning results, improving reasoning quality.
Reasoning-Driven Multimodal LLM for Domain Generalization
Zhipeng Xu (Xidian University), Nannan Wang (Xidian University)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed a reasoning-driven domain generalization framework RD-MLDG based on a multimodal large language model, leveraging class-related reasoning chains to enhance cross-domain performance.
ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory
Siru Ouyang (UIUC), Tomas Pfister (Google Cloud AI Research)
Reinforcement Learning from Human FeedbackLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the REASONINGBANK memory framework, which can distill transferable reasoning strategies from both successful and failed experiences, and combines with MATTS to achieve self-evolution during testing
Reassessing Layer Pruning in LLMs: New Insights and Methods
Yao Lu (Zhejiang University of Technology), Zhaowei Zhu (D5 Data)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically evaluates hierarchical pruning methods for large language models (LLMs), proposes and verifies an efficient pruning strategy combining simple reverse-order pruning (Reverse-Order) and partial fine-tuning of residual layers (Partial-Layer Fine-Tuning), provides theoretical gradient flow analysis, and ultimately achieves significant pruning effects on Llama-3.1-8B-Instruct, Llama-3-8B, and Llama-3-70B.
ReCAPA: Hierarchical Predictive Correction to Mitigate Cascading Failures
Xiyin Zeng (Hong Kong University of Science and Technology (Guangzhou)), Hao Wang (South China Normal University)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelContrastive LearningImageVideoTextMultimodality
🎯 What it does: Proposes the ReCAPA framework, which utilizes hierarchical prediction correction (HPCC) and cross-level alignment mechanisms to perform multi-level correction on actions, subgoals, and trajectories, thereby reducing error accumulation in long-term tasks.
RECAST: Expanding the Boundaries of LLMs' Complex Instruction Following with Multi-Constraint Data
Zhengkang Guo (Fudan University), Xiaoqing Zheng (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Designed the RECAST framework and constructed the RECAST-30K dataset to train LLMs to accurately follow instructions under multiple constraints, and proposed the RLVC reinforcement learning method to further improve constraint satisfaction rates.
RECODE-H: A Benchmark for Research Code Development with Interactive Human Feedback
Chunyu Miao (University Of Illinois Chicago), Philip S. Yu
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose the RECODE-H benchmark to evaluate the performance of LLMs in interactive research code generation.
ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving
Yongkang Li (Huazhong University of Science and Technology), Xinggang Wang (Xiaomi EV)
Autonomous DrivingReinforcement LearningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposes the ReCogDrive framework, integrating Vision-Language Models with diffusion planners to achieve end-to-end autonomous driving;
RECON: Robust symmetry discovery via Explicit Canonical Orientation Normalization
Alonso Urbano (Zuse Institute Berlin), Sebastian Pokutta (Zuse Institute Berlin)
ClassificationPose EstimationAuto EncoderImage
🎯 What it does: Propose the RECON method, which utilizes unsupervised class pose decomposition and explicit normalization to discover instance-specific symmetric distributions and achieve data alignment canonicalization.
Reconciling Visual Perception and Generation in Diffusion Models
Liulei Li (Zhejiang University), Wenguan Wang (Zhejiang University)
ClassificationSegmentationGenerationDepth EstimationTransformerDiffusion modelImage
🎯 What it does: Propose the GENREP model, achieving simultaneous visual understanding and image generation in a single training process.
Reconstruct Anything Model a lightweight general model for computational imaging
Matthieu Terris (Université Paris-Saclay), Julián Tachella (ENSL)
RestorationSuper ResolutionConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingComputed TomographyPhysics Related
🎯 What it does: Propose a lightweight general-purpose model RAM that can solve multiple computational imaging inverse problems (e.g., deblurring, CT/MRI reconstruction, super-resolution) in a single framework.
Reconstructing KV Caches with Cross-Layer Fusion for Enhanced Transformers
Hongzhan Lin (Taobao & Tmall Group of Alibaba), Bo Zheng (Taobao & Tmall Group of Alibaba)
Computational EfficiencyRepresentation LearningTransformerText
🎯 What it does: Propose a cross-layer KV cache reconstruction method called FusedKV and its lightweight version FusedKV-Lite, which learns to fuse KV caches from lower and middle layers to reduce memory consumption in higher-layer caches.
Reconstruction Alignment Improves Unified Multimodal Models
Ji Xie (University Of California Berkeley), XuDong Wang (University Of Washington)
GenerationComputational EfficiencyRepresentation LearningVision Language ModelAuto EncoderImageMultimodality
🎯 What it does: Propose the RECA post-training method, which leverages the visual understanding encoder's embeddings from the unified multimodal model (UMM) as dense text prompts for self-supervised image reconstruction to align understanding and generation.
ReconViaGen: Towards Accurate Multi-view 3D Object Reconstruction via Generation
Jiahao Chang (Chinese University of Hong Kong), Xiaoguang Han (Chinese University of Hong Kong)
GenerationTransformerDiffusion modelFlow-based ModelImageMesh
🎯 What it does: Integrate reconstruction prior (VGGT) with diffusion generation prior (TRELLIS) to propose a multi-scale coarse-to-fine generation framework (ReconViaGen), achieving high-precision, complete 3D object reconstruction from multi-view without camera parameters.
Recover Cell Tensor: Diffusion-Equivalent Tensor Completion for Fluorescence Microscopy Imaging
Chenwei Wang, Wenqi Zhu (University Of Oxford)
RestorationDiffusion modelScore-based ModelBiomedical Data
🎯 What it does: Propose a tensor completion framework for three-dimensional cell imaging in fluorescence microscopy (FM), treating sparsely and noisily sampled data with uniform Z-axis sampling as a low-rank tensor recovery problem;
Rectified Decoupled Dataset Distillation: A Closer Look for Fair and Comprehensive Evaluation
Xinhao Zhong (Harbin Institute of Technology), Yaowei Wang (Harbin Institute of Technology)
Computational EfficiencyKnowledge DistillationImageBenchmark
🎯 What it does: Proposed a unified evaluation framework RD 3, systematically re-evaluating and correcting the post-evaluation inconsistencies of existing separated dataset distillation methods;
Rectifying LLM Thought from Lens of Optimization
Junnan Liu (Monash University), Kai Chen (Shanghai Artificial Intelligence Laboratory)
OptimizationTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes the REPRO method, which corrects the Chain-of-Thought (CoT) of large language models (LLMs) from an optimization perspective, reducing overthinking and enhancing reasoning efficiency.
Recurrent Action Transformer with Memory
Egor Cherepanov (AXXX), Aleksandr Panov (AXXX)
TransformerReinforcement LearningTime SeriesSequential
🎯 What it does: Proposes an offline reinforcement learning model named Recurrent Action Transformer with Memory (RATE), which employs mechanisms such as segmented recursion, learnable memory embedding, hidden state caching, and Memory Retention Valve (MRV) to achieve superior decision-making in partially observable, long-horizon tasks.
RedacBench: Can AI Erase Your Secrets?
Hyunjun Jeon (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
Safty and PrivacyLarge Language ModelTextBenchmark
🎯 What it does: Developed the RedacBench benchmark for evaluating text redaction safety across different domains and strategies, covering multiple industries and security policies.
RedCodeAgent: Automatic Red-teaming Agent against Diverse Code Agents
Chengquan Guo (University of Chicago), Bo Li (University of Chicago)
Adversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Developed an automated red team proxy, RedCodeAgent, targeting LLM code agents to execute red team attacks, leveraging a memory module, toolbox, and sandbox evaluation to identify security vulnerabilities in code generation and execution.
ReDDiT: Rehashing Noise for Discrete Visual Generation
Tianren Ma (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
GenerationTransformerDiffusion modelImageText
🎯 What it does: Proposes ReDDiT—a discrete diffusion transformer that leverages re-hashed noise, significantly improving the quality of discrete visual generation;
Redirection for Erasing Memory (REM): Towards a universal unlearning method for corrupted data
Stefan Schoepf (University of Cambridge), Eleni Triantafillou (Google DeepMind)
Data-Centric LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed a new two-dimensional (discovery rate and statistical regularity) classification task for corrupted data delearning, and introduced a general delearning method called REM that can balance performance in this two-dimensional space;
RedSage: A Cybersecurity Generalist LLM
Naufal Suryanto (Khalifa University), Ernesto Damiani (University of Milan)
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmark
🎯 What it does: Proposed and open-sourced an 8B-scale cybersecurity-specific LLM named RedSage, encompassing continuous pre-training, post-training fine-tuning, and end-to-end evaluation.
RedTeamCUA: Realistic Adversarial Testing of Computer-Use Agents in Hybrid Web-OS Environments
Zeyi Liao (Ohio State University), Huan Sun (Ohio State University)
Adversarial AttackLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the REDTEAMCUA red team adversarial testing framework and the RTC-BENCH benchmark to evaluate the vulnerability of Computer Use Agents (CUA) to indirect prompt injection within a hybrid Web-OS sandbox.
Reducing Belief Deviation in Reinforcement Learning for Active Reasoning of LLM Agents
Deyu Zou (Chinese University of Hong Kong), Yu Gong (ByteDance)
TransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmark
🎯 What it does: This paper proposes a reward method called T3 based on early truncation, which controls the belief deviation of LLM agents in multi-round active reasoning, thereby improving credit assignment and training stability in reinforcement learning.
Reducing Class-Wise Performance Disparity via Margin Regularization
Beier Zhu (University of Science and Technology of China), Hanwang Zhang (Nanyang Technological University)
ClassificationConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Investigated the issue of inter-class accuracy differences in deep neural networks on balanced data, and proposed a novel regularization method called MR2 to reduce this gap.
Reducing Contextual Stochastic Bilevel Optimization via Structured Function Approximation
Maxime Bouscary (Massachusetts Institute of Technology), Saurabh Amin (Massachusetts Institute of Technology)
OptimizationHyperparameter Search
🎯 What it does: Proposes a method to reduce the complexity of contextual stochastic bilevel optimization (CSBO) problems through structured function approximation.
Reducing information dependency does not cause training data privacy. Adversarially non-robust features do.
Rasmus Torp (Dartmouth College), Adam Breuer (Dartmouth College)
ClassificationSafty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper challenges the traditional view that the stronger the model's dependence on training data, the more susceptible it is to model inversion attacks (MIA) through a series of experiments. It points out that privacy leakage mainly originates from the 'non-robust features' (generalizable but imperceptible features) learned by the model. Subsequently, the Anti-Adversarial Training (AT-AT) training framework is proposed, which actively encourages the model to learn non-robust features, significantly reducing the reconstruction accuracy under MIA while maintaining or even improving model accuracy.
Reducing Semantic Mismatch in Brain-to-Text Decoding Through Personalized Multimodal Masking
Jiaxuan Chen (Zhejiang University), Gang Pan (Zhejiang University)
GenerationTransformerLarge Language ModelVision Language ModelMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed the Yo'Mind framework, which utilizes Optimal Transport-driven personalized multimodal semantic masking to improve the brain-text decoding process and address semantic matching bias;
Reducing Symmetry Increase in Equivariant Neural Networks
Ning Lin (Renmin University of China), Hao Sun (Renmin University of China)
Drug DiscoveryGraph Neural NetworkPoint CloudGraph
🎯 What it does: Studied the symmetry increase problem that equivariant neural networks (ENNs) encounter when processing symmetric inputs, introduced the concept of symmetry lower bounds, provided a computable algorithm and design guidelines, and verified the theory through visualization and experiments.
Reevaluating Policy Gradient Methods for Imperfect-Information Games
Max Rudolph (University of Texas at Austin), Samuel Sokota (Carnegie Mellon University)
Computational EfficiencyHyperparameter SearchReinforcement LearningTabular
🎯 What it does: Achieved precise exploitability computation for large-scale incomplete information games, conducting over 7,000 training experiments on seven DRL algorithms;
Ref-Adv: Exploring MLLM Visual Reasoning in Referring Expression Tasks
Qihua Dong (Northeastern University), Yun Fu (Northeastern University)
RecognitionLarge Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposes the Ref-Adv dataset to evaluate the visual reasoning and localization capabilities of multimodal large language models in referential expression understanding tasks.
RefAny3D: 3D Asset-Referenced Diffusion Models for Image Generation
Hanzhuo Huang (ShanghaiTech University), Sibei Yang (Sun Yat-sen University)
GenerationTransformerDiffusion modelAuto EncoderImagePoint CloudMesh
🎯 What it does: Proposed RefAny3D, a conditional diffusion model based on 3D assets, capable of simultaneously generating RGB images and corresponding point maps, achieving geometric and texture consistency for 3D models.
Reference Grounded Skill Discovery
Seungeun Rho (Georgia Institute of Technology), Sehoon Ha (Georgia Institute of Technology)
Representation LearningRobotic IntelligenceReinforcement LearningContrastive LearningVideo
🎯 What it does: This paper proposes a framework named RGSD for skill discovery in high-degree-of-freedom agents based on reference data. It first maps reference motions to a unit hypersphere using contrastive learning, then simultaneously achieves imitation of reference skills and discovery of semantically related new skills in this latent space.
Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning
Xuan Li (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
OptimizationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextGraph
🎯 What it does: Studied instruction-driven molecular optimization tasks, proposing a reference-guided policy optimization (RePO) method that enables large language models to perform multi-step reasoning and generate optimized molecules satisfying similarity constraints without trajectory supervision.
References Improve LLM Alignment in Non-Verifiable Domains
Kejian Shi (Yale University), Arman Cohan (Yale University)
Data-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose using high-quality reference answers to guide LLMs as evaluators (RefEval/RefMatch), and apply it to self-improvement of LLM alignment.
Referring Layer Decomposition
Fangyi Chen (ByteDance), Longyin Wen (ByteDance)
GenerationPrompt EngineeringVision Language ModelDiffusion modelAuto EncoderMultimodality
🎯 What it does: Proposed and implemented a Referring Layer Decomposition (RLD) task based on multimodal prompts (text, points, boxes, masks), capable of predicting complete RGBA layers from a single RGB image.
Refine Drugs, Don’t Complete Them: Uniform-Source Discrete Flows for Fragment-Based Drug Discovery
Benno Kaech (In Virtuo Laboratories), Gianvito Grasso (In Virtuo Laboratories)
Drug DiscoveryTransformerReinforcement LearningDiffusion modelFlow-based ModelText
🎯 What it does: Propose the InVirtuoGen model, which utilizes a unified source discrete flow to perform molecule generation, property optimization, and lead optimization on fragmented SMILES.
Refine Now, Query Fast: A Decoupled Refinement Paradigm for Implicit Neural Fields
Tianyu Xiong (Ohio State University), Han Wei Shen
Super ResolutionComputational EfficiencyRepresentation LearningNeural Radiance FieldPhysics Related
🎯 What it does: Propose the Decoupled Representation Refinement (DRR) framework, which separates the deep reconstruction network from the low-cost embedding structure to achieve efficient, high-quality Implicit Neural Representation (INR);