arXivSub Start free trial

NeurIPS 2025 Papers with Code β€” Page 16

Conference on Neural Information Processing Systems Β· 2283 papers

Praxis-VLM: Vision-Grounded Decision Making via Text-Driven Reinforcement Learning

Zhe Hu (Hong Kong Polytechnic University), Yu Yin (Case Western Reserve University)

CodeAutonomous DrivingRobotic IntelligenceTransformerReinforcement LearningVision Language ModelVideoText

🎯 What it does: This paper proposes a text-driven reinforcement learning method called Praxis-VLM, which enables visual language models to possess transferable reasoning and decision-making capabilities in visual decision-making tasks.

Pre-trained Large Language Models Learn to Predict Hidden Markov Models In-context

Yijia Dai (Cornell University), Jennifer J. Sun (Cornell University)

CodeTransformerLarge Language ModelTextSequential

🎯 What it does: The study investigates the prediction of sequences generated by hidden Markov models (HMM) using pre-trained large language models (LLM) through in-context learning (ICL), finding that their performance can approach the theoretically optimal Viterbi algorithm.

Pre-Trained Policy Discriminators are General Reward Models

Shihan Dou (Shanghai AI Laboratory), Kai Chen (Shanghai AI Laboratory)

CodeReinforcement Learning from Human FeedbackTransformerReinforcement LearningContrastive LearningText

🎯 What it does: Proposes the POLAR framework, redefining the reward model as a policy discriminator, utilizing unsupervised pre-training and a small amount of manually labeled fine-tuning to achieve relative evaluation of different policies;

Predictability Enables Parallelization of Nonlinear State Space Models

Xavier Gonzalez (Stanford University), Scott Linderman (Stanford University)

CodeOptimizationRecurrent Neural NetworkTime SeriesStochastic Differential Equation

🎯 What it does: This paper reveals that the predictability of the model (measured by the maximum Lyapunov exponent) determines the condition number of the optimization problem by transforming the trajectory solution of the nonlinear state space model into a fast-converging parallel optimizable problem, thus judging whether efficient parallelization can be achieved on a GPU.

Predictable Scale (Part II) --- Farseer: A Refined Scaling Law in LLMs

Houyi Li (Fudan University), Daxin Jiang (StepFun)

CodeTransformerLarge Language ModelText

🎯 What it does: This paper proposes and validates a new LLM scaling lawβ€”Farseerβ€”through large-scale experiments, which can accurately predict the loss of models of different scales;

Prediction-Powered Semi-Supervised Learning with Online Power Tuning

Noa Shoham (Technion Institute of Technology), Yaniv Romano (Technion Institute of Technology)

CodeClassificationOptimizationKnowledge DistillationConvolutional Neural NetworkSupervised Fine-TuningReinforcement LearningImageTabular

🎯 What it does: A semi-supervised learning framework called PP-SSL based on Prediction-Powered Inference (PPI) is proposed, which utilizes a teacher model to generate pseudo-labels and construct unbiased gradient estimates, and adaptively adjusts the pseudo-label weight λ through online learning.

Preference Learning with Lie Detectors can Induce Honesty or Evasion

Chris Cundy (FAR AI), Adam Gleave (FAR AI)

CodeRecommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: This paper studies the process of introducing a lie detector in preference learning (post-training of LLMs), evaluating whether it can effectively encourage the model to be honest or lead the model to evade detection and continue lying.

Preference Optimization by Estimating the Ratio of the Data Distribution

Yeongmin Kim (Korea Advanced Institute of Science and Technology), Il-chul Moon

CodeRecommendation SystemOptimizationLarge Language ModelSupervised Fine-TuningReinforcement LearningText

🎯 What it does: A preference optimization framework BPO based on Bregman divergence is proposed, treating DPO as a proportional matching problem, achieving optimality and simplicity without the need for a reward model.

Preference-Based Dynamic Ranking Structure Recognition

Nan Lu (Chinese Academy of Sciences), Xinyu Tian

CodeRecommendation SystemOptimizationTime Series

🎯 What it does: A dynamic framework for simultaneous grouping and ranking is proposed, capable of identifying the group structure of items and its changes in time series.

Preference-driven Knowledge Distillation for Few-shot Node Classification

Xing Wei (Tongji University), Wei Ye (Tongji University)

CodeKnowledge DistillationGraph Neural NetworkLarge Language ModelSupervised Fine-TuningReinforcement LearningTextGraph

🎯 What it does: A Preference-driven Knowledge Distillation (PKD) framework has been developed, combining large language models (LLMs) with various graph neural networks (GNNs) for few-shot node classification in Text Attribute Graphs (TAGs).

Preference-Driven Multi-Objective Combinatorial Optimization with Conditional Computation

Mingfeng Fan (National University of Singapore), Guillaume Adrien Sartoretti (National University of Singapore)

CodeOptimizationTransformerReinforcement LearningMixture of ExpertsContrastive LearningTabularBenchmark

🎯 What it does: Proposed and implemented the POCCO framework, which addresses multi-objective combinatorial optimization problems using conditional computation blocks and preference-driven training.

Preference-Guided Diffusion for Multi-Objective Offline Optimization

Yashas Annadani (TU Munich), Barbara E Engelhardt

CodeOptimizationNeural Architecture SearchDiffusion modelTabular

🎯 What it does: A preference-guided diffusion model is proposed for generating Pareto fronts in offline multi-objective optimization (MOO).

PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generation

Ao Wang (Tsinghua University), Guiguang Ding (Tsinghua University)

CodeGenerationCompressionComputational EfficiencyTransformerVision Language ModelTextMultimodality

🎯 What it does: Proposes PrefixKV, which achieves efficient generation of visual language models through adaptive prefix KV cache compression.

PreFM: Online Audio-Visual Event Parsing via Predictive Future Modeling

Xiao Yu (Beijing Jiaotong University), Yunchao Wei (Beijing Jiaotong University)

CodeRecognitionComputational EfficiencyKnowledge DistillationVideoMultimodalityAudio

🎯 What it does: This paper proposes an online audio-video event parsing framework called PreFM, which can parse audio, visual, and audio-video events frame by frame in real-time video streams, balancing accuracy and real-time performance.

PRESCRIBE: Predicting Single-Cell Responses with Bayesian Estimation

Jiabei Cheng (Shanghai Jiao Tong University), Jun Xia (Hong Kong University of Science and Technology)

CodeDrug DiscoveryFlow-based ModelBiomedical Data

🎯 What it does: The PRESCRIBE framework is proposed for predicting single-cell gene interference and providing reliable uncertainty estimates.

PRESTO: Preimage-Informed Instruction Optimization for Prompting Black-Box LLMs

Jaewon Chu (Korea University), Hyunwoo J. Kim (KAIST)

CodeOptimizationTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: We propose PRESTO, a framework that accelerates black-box LLM instruction optimization using a soft prompt many-to-one mapping (preimage) structure;

Pretraining a Shared Q-Network for Data-Efficient Offline Reinforcement Learning

Jongchan Park (Hyundai Motor Company), Donghwan Lee (Korea Advanced Institute of Science and Technology)

CodeReinforcement Learning

🎯 What it does: This paper proposes a simple pre-training framework that first pre-trains a shared Q-network's feature extractor using a transition prediction task in offline reinforcement learning, and then uses it as the initialization for the Q-network, directly applied to the training of any offline RL algorithm;

Principled Long-Tailed Generative Modeling via Diffusion Models

Pranoy Das (Purdue University), Vijay Gupta (Purdue University)

CodeGenerationData SynthesisOptimizationDiffusion modelImageStochastic Differential Equation

🎯 What it does: This study investigates the learning methods of diffusion generative models under long-tail distributions, proposing to construct a multi-player Nash game through deep mutual learning to balance the generation quality across different categories.

Prior-Guided Diffusion Planning for Offline Reinforcement Learning

Donghyeon Ki (Korea University), Byung-Jun Lee (Korea University)

CodeRecurrent Neural NetworkReinforcement LearningDiffusion modelTabularBenchmark

🎯 What it does: Proposes Prior Guidance (PG), a learnable prior distribution for training diffusion planners in offline reinforcement learning, allowing for the direct generation of high-value trajectories during inference, thus avoiding multi-trajectory sampling and reward optimization.

Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number

Jingyuan Zhou (Shanghai Jiao Tong University), Lei Xu (Shanghai Jiao Tong University)

CodeDrug DiscoveryGraph Neural NetworkFlow-based ModelGraphOrdinary Differential Equation

🎯 What it does: This paper proposes a target-aware molecular generation framework PAFlow based on flow matching, which can directly generate high-affinity 3D small molecules under the condition of a given protein pocket;

Privacy Reasoning in Ambiguous Contexts

Ren Yi (Google Research), Marco Gruteser (Google Research)

CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: This paper studies the performance of large language models in privacy decision-making, constructing the PrivacyLens+ and ConfAIde+ datasets that include both positive and negative samples, and proposes the Camber framework to enhance privacy judgment through automated context disambiguation.

Private Continual Counting of Unbounded Streams

Ben Jacobsen (University of Wisconsin - Madison), Kassem Fawaz (University of Wisconsin - Madison)

CodeSafty and PrivacyComputational Efficiency

🎯 What it does: A new algorithm for achieving differentially private continuous counting in unbounded streams with unknown input size is proposed.

Private Evolution Converges

TomΓ‘s GonzΓ‘lez, Aaditya Ramdas

CodeOptimizationSafty and PrivacyDiffusion modelImage

🎯 What it does: This paper proposes a new theoretical framework and conducts a rigorous analysis of the convergence of the Private Evolution (PE) algorithm, providing upper bounds for the 1-Wasserstein distance convergence in Euclidean space and general Banach spaces.

Private Training Large-scale Models with Efficient DP-SGD

Liangyu Wang (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)

CodeSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: A cache-friendly per-layer DP-SGD algorithm called FlashDP is designed to significantly improve memory utilization and throughput in training large language models while ensuring differential privacy.

Probabilistic Stability Guarantees for Feature Attributions

Helen Jin (University of Pennsylvania), Eric Wong (University of Pennsylvania)

CodeExplainability and InterpretabilityTransformerImageText

🎯 What it does: This paper introduces the concept of Soft Stability and presents a model-agnostic, sample-efficient stability certification algorithm (SCA) that can provide a probabilistically rigorous assessment of the robustness of feature importance explanations. It also demonstrates through Boolean function analysis that mild smoothing (MuS) can enhance stability without significantly sacrificing accuracy.

Probing Neural Combinatorial Optimization Models

Zhiqin Zhang (Singapore Management University), Hoong Chuin Lau (Singapore Management University)

CodeOptimizationTransformerTabular

🎯 What it does: This paper is the first to apply probing techniques to Neural Combinatorial Optimization (NCO) models, systematically designing low-order and high-order probing tasks, and proposing a new Coefficient Significance Probing (CS-Probing) method;

Process vs. Outcome Reward: Which is Better for Agentic RAG Reinforcement Learning

wenlin zhang, Xiangyu Zhao (City University of Hong Kong)

CodeTransformerReinforcement LearningAgentic AITextRetrieval-Augmented Generation

🎯 What it does: A process-supervised reinforcement learning-based agentic RAG framework called ReasonRAG is proposed, enabling LLMs to autonomously perform retrieval, generate queries, extract evidence, and answer questions.

Program Synthesis via Test-Time Transduction

Kang-il Lee (Seoul National University), Kyomin Jung (Seoul National University)

CodeAI Code AssistantTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought

🎯 What it does: Proposed and implemented the SYNTRA framework, which generates candidate programs through LLM and actively performs transductive prediction and hypothesis elimination on visible test inputs to enhance the robustness of program synthesis.

Progress Reward Model for Reinforcement Learning via Large Language Models

Xiuhui Zhang (Beihang University), Yue Deng (Beihang University)

CodeRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningMultimodalityChain-of-Thought

🎯 What it does: The PRM4RL framework is proposed, which utilizes large language models to decompose long tasks into subtasks and construct a progress reward model, integrating high-level planning with low-level rewards.

Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities

Tara Akhound-Sadegh (Mila - Quebec AI Institute), Alexander Tong (AITHYRA)

CodeDiffusion model

🎯 What it does: A stepwise inference time annealing framework named PITA is designed for training diffusion models to sample Boltzmann densities at low temperatures.

Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry

Sai Sumedh R. Hindupur (Harvard University), Demba E. Ba (Harvard University)

CodeOptimizationRepresentation LearningAuto EncoderImage

🎯 What it does: This paper reveals the implicit bias of Sparse Autoencoders (SAE) in discovering model concepts through theoretical analysis and experimental validation, and proposes a new geometry-based SAE (SpaDE) to overcome these biases.

Projection-Manifold Regularized Latent Diffusion for Robust General Image Fusion

Lei Cao (Wuhan University), Jiayi Ma (Wuhan University)

CodeRestorationSegmentationDiffusion modelImage

🎯 What it does: This paper proposes a training-free image fusion framework based on a pre-trained latent diffusion model, called PDFuse, which utilizes projection-manifold regularization to achieve multi-source information fusion.

Prompt Tuning Decision Transformers with Structured and Scalable Bandits

Finn Rietz (Γ–rebro University), Lele Cao (King AI Labs Microsoft Gaming)

CodeTransformerReinforcement LearningPrompt EngineeringSequential

🎯 What it does: A method for prompt tuning based on multi-armed bandits is proposed to generate optimal trajectory prompts for the pre-trained Prompt Decision Transformer (PDT) in offline multi-task reinforcement learning, enhancing performance and adapting to out-of-vocabulary (OOV) tasks.

Prompt-Guided Alignment with Information Bottleneck Makes Image Compression Also a Restorer

Xuelin Shen (Guangdong Laboratory of Artificial Intelligence and Digital Economy), Wenhan Yang (Peng Cheng Laboratory)

CodeRestorationCompressionPrompt EngineeringAuto EncoderImage

🎯 What it does: A lightweight prompt-guided information bottleneck compression scheme is proposed, unifying denoising and compression potential representations.

Prompt-guided Disentangled Representation for Action Recognition

tianci wu, zhang liang

CodeRecognitionRepresentation LearningGraph Neural NetworkPrompt EngineeringVideo

🎯 What it does: This paper proposes a prompt-guided decoupled representation learning framework called ProDA, which can extract and separate subgraph representations of any specified action from spatial-temporal scene graphs in multi-action videos, thereby enhancing action recognition and localization performance.

ProSpero: Active Learning for Robust Protein Design Beyond Wild-Type Neighborhoods

Michal Kmicikiewicz (Helmholtz Munich), Ewa Szczurek (Helmholtz Munich)

CodeOptimizationProtein Structure PredictionConvolutional Neural NetworkReinforcement LearningBiomedical Data

🎯 What it does: A proactive learning framework PROSPERO based on frozen pre-trained generative models is proposed, utilizing target masking and biologically constrained sequence Monte Carlo sampling to iteratively update the surrogate model and guide the generation of high-fitness and novel protein sequences.

Prot2Text-V2: Protein Function Prediction with Multimodal Contrastive Alignment

Xiao Fei (Γ‰cole Polytechnique), Michalis Vazirgiannis (Mohamed bin Zayed University of Artificial Intelligence)

CodeGenerationProtein Structure PredictionTransformerLarge Language ModelContrastive LearningTextMultimodality

🎯 What it does: This paper proposes the Prot2Text-V2 model, which can directly generate free-text protein function descriptions from amino acid sequences.

Protein Design with Dynamic Protein Vocabulary

Nuowei Liu (East China Normal University), Yuanbin Wu

CodeGenerationProtein Structure PredictionTransformerLarge Language ModelTextBiomedical DataRetrieval-Augmented Generation

🎯 What it does: A functional-driven protein design framework named PRODVA is proposed, which can dynamically retrieve and integrate text function descriptions with natural protein fragments into the generation process, outputting structurally feasible and functionally aligned protein sequences.

Protein Inverse Folding From Structure Feedback

Junde Xu (Chinese University of Hong Kong), Jiezhong Qiu

CodeOptimizationProtein Structure PredictionSupervised Fine-TuningGraph

🎯 What it does: By combining the structural feedback of protein folding models with Direct Preference Optimization (DPO), the inverse folding (protein sequence design) model is optimized to generate sequences that are closer to the target three-dimensional structure.

ProtInvTree: Deliberate Protein Inverse Folding with Reward-guided Tree Search

Mengdi Liu (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)

CodeProtein Structure PredictionReinforcement LearningDiffusion modelBiomedical Data

🎯 What it does: This paper proposes ProtInvTree, a protein inverse folding framework based on reward-guided tree search, which achieves stepwise sequence generation through phased focus and baselining actions.

Provable Scaling Laws for the Test-Time Compute of Large Language Models

Yanxi Chen (Alibaba Group), Jingren Zhou (Alibaba Group)

CodeTransformerLarge Language ModelTextChain-of-Thought

🎯 What it does: Two types of black-box LLM-based two-stage elimination and league-style aggregation algorithms are proposed, along with a provable extension law for computational performance during testing.

Provably Efficient Online RLHF with One-Pass Reward Modeling

Long-Fei Li (Nanjing University), Zhi-Hua Zhou (Nanjing University)

CodeOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningText

🎯 What it does: Proposes a one-pass reward modeling algorithm for online RLHF that eliminates the need for historical data storage and achieves constant time updates per round;

Proxy Target: Bridging the Gap Between Discrete Spiking Neural Networks and Continuous Control

Zijie Xu (Peking University), Zhaofei Yu (Peking University)

CodeSpiking Neural NetworkReinforcement LearningSequential

🎯 What it does: A proxy target framework is proposed to address the mismatch between the discrete SNN and the soft update mechanism of the continuous target network, stabilizing the reinforcement learning training of SNN in continuous control.

Proxy-SPEX: Sample-Efficient Interpretability via Sparse Feature Interactions in LLMs

Landon Butler (University of California Berkeley), Kannan Ramchandran (University of California Berkeley)

CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: PROXYSPEX is proposed, an efficient explanation method for discovering sparse hierarchical interactions in LLMs through Gradient Boosting Trees (GBT).

PRSformer: Disease Prediction from Million-Scale Individual Genotypes

Payam Dibaeinia (23andMe), Aly A Khan

CodeClassificationOptimizationTransformerBiomedical Data

🎯 What it does: A multi-task Transformer model named PRSformer has been constructed to directly predict the risk of 18 autoimmune and inflammatory diseases from millions of individual genotypes.

Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization

Tianjun Yao (Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Shen (Mohamed bin Zayed University of Artificial Intelligence)

CodeGraph Neural NetworkGraph

🎯 What it does: A pruning-based graph OOD method called PrunE is proposed, which enhances the generalization ability of graph neural networks on out-of-distribution data by removing spurious edges to maintain invariant subgraphs.

Pseudo-Riemannian Graph Transformer

Le Viet Quan, Viet Cuong Ta (VNU University of Engineering and Technology)

CodeClassificationRepresentation LearningGraph Neural NetworkTransformerGraph

🎯 What it does: A differential isometric framework for graph embedding is proposed, which decomposes pseudo-Riemannian manifolds into a product of spheres and hyperbolic spaces, and based on this, a pseudo-Riemannian graph Transformer (Q-GT) is constructed to effectively represent complex graph structures.

PseuZO: Pseudo-Zeroth-Order Algorithm for Training Deep Neural Networks

Pengyun Yue (Peking University), Zhouchen Lin (Peking University)

CodeOptimizationLarge Language ModelSupervised Fine-TuningText

🎯 What it does: A Pseudo-Zeroth-Order (PseuZO) optimization framework is proposed, which first estimates the Jacobian matrix of the model output, then combines it with the gradient of the external loss, and employs exponential moving average and sliding window techniques to achieve efficient fine-tuning of large-scale LLMs.

PT-MoE: An Efficient Finetuning Framework for Integrating Mixture-of-Experts into Prompt Tuning

Zongqian Li (University of Cambridge), Nigel Collier (University of Cambridge)

CodeTransformerPrompt EngineeringMixture of ExpertsText

🎯 What it does: PT-MoE is proposed, an efficient prompt tuning framework that combines matrix factorization with mixture of experts (MoE) routing.

Purity Law for Neural Routing Problem Solvers with Enhanced Generalizability

Wenzhao Liu (University of Chinese Academy of Sciences), Tiande Guo (University of Chinese Academy of Sciences)

CodeOptimizationReinforcement LearningTabular

🎯 What it does: Proposes the Purity Law and designs the PUPO training framework based on this law to enhance the generalization ability of neural routing solvers.

Put CASH on Bandits: A Max K-Armed Problem for Automated Machine Learning

Amir Rezaei Balef (University of TΓΌbingen), Katharina Eggensperger (University of TΓΌbingen)

CodeOptimizationHyperparameter SearchTabular

🎯 What it does: This paper proposes a maximum k-armed bandit methodβ€”MaxUCBβ€”for the joint algorithm selection and hyperparameter optimization (CASH) problem in automated machine learning (AutoML). The method decides on the model and its hyperparameters by calling a single HPO iteration during each round of budget allocation, aiming to maximize the observed best performance.

Puzzles: Unbounded Video-Depth Augmentation for Scalable End-to-End 3D Reconstruction

Jiahao Ma (Australian National University), Chuong Nguyen (Data61/CSIRO)

CodeData SynthesisDepth EstimationImageVideo

🎯 What it does: A data augmentation framework called Puzzles is proposed, which slices and rearranges a single image or short video, simulating camera motion to generate unbounded virtual videos with pose and depth for training end-to-end 3D reconstruction models.

PyraMotion: Attentional Pyramid-Structured Motion Integration for Co-Speech 3D Gesture Synthesis

Zhizhuo Yin (Hong Kong University of Science and Technology), Pan Hui (Hong Kong University of Science and Technology)

CodeGenerationData SynthesisTransformerAuto EncoderMultimodalityAudio

🎯 What it does: A framework for audio-driven full-body gesture synthesis based on multi-scale discrete encoding, called PyraMotion, is proposed, and an Attentional Pyramidal VQ-VAE (APVQ-VAE) is designed for multi-scale gesture discretization and reconstruction.

Q-Insight: Understanding Image Quality via Visual Reinforcement Learning

Weiqi Li (Peking University), Jian Zhang (Peking University)

CodeRecognitionOptimizationExplainability and InterpretabilityReinforcement LearningVision Language ModelImage

🎯 What it does: Q-Insight is proposed, a multi-task visual language model based on GRPO for score regression of image quality and degradation perception, capable of generating interpretable reasoning steps.

Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment

Deokjae Lee (Seoul National University), Hyun Oh Song (Seoul National University)

CodeCompressionOptimizationComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: This paper addresses the problem of post-training quantization (PTQ) for large language models (LLMs) and proposes a customizable set of score bit quantizers called Q-Palette, which is embedded in the mixed scheme quantization (MSQ) framework to achieve better compression and inference speed under memory or latency constraints.

Q3R: Quadratic Reweighted Rank Regularizer for Effective Low-Rank Training

Ipsita Ghosh (University of Central Florida), Christian KΓΌmmerle (University of Central Florida)

CodeOptimizationTransformerSupervised Fine-TuningImageText

🎯 What it does: This paper proposes a quadratic reweighted rank regularization (Q3R) based on the Iteratively Reweighted Least Squares (IRLS) framework, which induces a low-rank structure in the network weight matrix through regularization during training. It can be used for low-rank pre-training of large-scale models as well as for parameter-efficient fine-tuning.

QFFT, Question-Free Fine-Tuning for Adaptive Reasoning

Wanlong Liu (University of Electronic Science and Technology of China), Benyou Wang (Huawei Noah's Ark Lab)

CodeTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought

🎯 What it does: A question-free fine-tuning (QFFT) method is proposed, enabling large language models to use concise short chain thinking (Short CoT) for simple questions and automatically switch to reflective long chain thinking (Long CoT) for difficult questions, thus achieving efficient and accurate reasoning.

QiMeng-MuPa: Mutual-Supervised Learning for Sequential-to-Parallel Code Translation

Changxin Ke (State Key Laboratory of Processors, Institute of Computing Technology, Chinese Academy of Sciences), Yunji Chen (State Key Laboratory of Processors, Institute of Computing Technology, Chinese Academy of Sciences)

CodeAI Code AssistantTransformerLarge Language ModelSequential

🎯 What it does: A mutual supervision learning framework named QiMeng-MuPa is proposed to automatically translate sequential code into parallel code and ensure functional equivalence through unit testing.

QiMeng-SALV: Signal-Aware Learning for Verilog Code Generation

Yang Zhang (Institute of Computing Technology), Yunji Chen (Institute of Computing Technology)

CodeGenerationAI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningText

🎯 What it does: A Verilog code generation method based on signal-level reinforcement learning, QiMeng-SALV, is proposed, which utilizes signal implementations in partially correct modules to provide functional rewards, improving RL training.

QSVD: Efficient Low-rank Approximation for Unified Query-Key-Value Weight Compression in Low-Precision Vision-Language Models

Yutong Wang (New York University), Sai Qian Zhang (New York University)

CodeCompressionComputational EfficiencyTransformerVision Language ModelMultimodality

🎯 What it does: Designed and implemented the QSVD method, which jointly compresses the query-key-value (QKV) weights of the visual-language model (VLM) using SVD, and combines low-precision quantization to significantly reduce model parameters, KV cache, and computational load while maintaining or even improving inference accuracy.

QuadEnhancer: Leveraging Quadratic Transformations to Enhance Deep Neural Networks

Qian Chen (Chinese University of Hong Kong), Yin Zhang (Chinese University of Hong Kong)

CodeClassificationOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningImageText

🎯 What it does: Insert a lightweight quadratic enhancement module into each linear layer of the neural network to capture the quadratic interactions between features.

QuadricFormer: Scene as Superquadrics for 3D Semantic Occupancy Prediction

Sicheng Zuo (Tsinghua University), Jiwen Lu (Tsinghua University)

CodeSegmentationAutonomous DrivingComputational EfficiencyTransformerMixture of ExpertsPoint Cloud

🎯 What it does: Proposes a sparse object center representation based on superquadrics, achieving efficient 3D semantic occupancy prediction.

Quantifying Elicitation of Latent Capabilities in Language Models

Elizabeth Donoway (University of California), Jan Leike (Anthropic)

CodeClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: The study investigates the minimum number of trainable parameters required to activate the potential capabilities of large language models and constructs an 'activation frontier' to describe the relationship between parameters and performance; fine-tuning experiments are conducted using a randomly selected small number of parameters.

Quantifying Statistical Significance of Deep Nearest Neighbor Anomaly Detection via Selective Inference

Mizuki Niihori (Nagoya University), Ichiro Takeuchi (Nagoya University)

CodeAnomaly DetectionConvolutional Neural NetworkImageTabular

🎯 What it does: A statistical significance quantification method for deep k-nearest neighbor anomaly detection (Deep k-NN AD) based on selective inference (SI) is proposed, providing p-values for detection results and enabling controllable false positive rates.

Quantization-Free Autoregressive Action Transformer

Ziyad Sheebaelhamd (University of Tuebingen), Claire Vernade (University of Tuebingen)

CodeAutonomous DrivingRobotic IntelligenceTransformerReinforcement LearningMultimodality

🎯 What it does: A Quantization-Free Autoregressive Action Transformer (Q-FAT) is proposed, which directly models actions in continuous action spaces and predicts action distributions using GMM, achieving behavior cloning for robotic simulation tasks.

QuARI: Query Adaptive Retrieval Improvement

Eric Xing (Washington University in St. Louis), Nathan Jacobs (Washington University in St. Louis)

CodeRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality

🎯 What it does: A query adaptive retrieval framework QuARI is proposed, which utilizes transformers to generate query-specific linear projections to improve image-to-image and text-to-image retrieval.

Quartet: Native FP4 Training Can Be Optimal for Large Language Models

Roberto L. Castro, Dan Alistarh (IST Austria)

CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: This paper proposes and implements a fully low-precision 4-bit floating-point training framework called Quartet, which can complete the full training of large language models on NVIDIA Blackwell GPUs while maintaining accuracy comparable to or even better than FP16.

R-KV: Redundancy-aware KV Cache Compression for Reasoning Models

Zefan Cai (University of Wisconsin - Madison), Junjie Hu (University of Wisconsin - Madison)

CodeCompressionComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: A KV cache compression method for inference models, R-KV, is proposed, which dynamically eliminates useless tokens during the inference process using a dual evaluation of importance and redundancy, significantly reducing KV cache usage.

R$^2$ec: Towards Large Recommender Models with Reasoning

Runyang You (Hong Kong Polytechnic University), Liqiang Nie (Harbin Institute of Technology)

CodeRecommendation SystemTransformerLarge Language ModelReinforcement LearningContrastive LearningTabular

🎯 What it does: This paper proposes R2ec, a unified large-scale recommendation model that integrates inference and recommendation functions within the same architecture, and designs the RecPO RL training framework to achieve end-to-end optimization.

R1-ShareVL: Incentivizing Reasoning Capabilities of Multimodal Large Language Models via Share-GRPO

Huanjin Yao (Tsinghua University), Jiaxing Huang (Nanyang Technological University)

CodeOptimizationTransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality

🎯 What it does: The Share-GRPO method, based on reinforcement learning, enhances the performance of multimodal large language models in long-chain reasoning tasks.

R2R: Efficiently Navigating Divergent Reasoning Paths with Small-Large Model Token Routing

Tianyu Fu (Tsinghua University), Yu Wang (Tsinghua University)

CodeGenerationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText

🎯 What it does: A token-level router R2R is designed, allowing small language models (SLM) to operate during most generation steps, only invoking the large language model (LLM) on critical 'path divergence' tokens, significantly reducing inference costs while maintaining high-quality reasoning.

RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning

Hao Gao (Huazhong University of Science and Technology), Xinggang Wang (Huazhong University of Science and Technology)

CodeAutonomous DrivingTransformerReinforcement LearningGaussian SplattingPoint Cloud

🎯 What it does: Proposes the RAD framework, utilizing 3D Gaussian Splatting (3DGS) digital twin environments for closed-loop reinforcement learning to train end-to-end autonomous driving strategies.

RadarQA: Multi-modal Quality Analysis of Weather Radar Forecasts

Xuming He (ZheJiang University), LEI BAI

CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodality

🎯 What it does: RadarQA is proposed, a weather radar forecast quality analysis framework based on a multimodal large language model, which includes four tasks (scoring and evaluation of frames/sequences).

Radial Attention: $\mathcal O(n \log n)$ Sparse Attention for Long Video Generation

Xingyang Li (Massachusetts Institute of Technology), Song Han (Massachusetts Institute of Technology)

CodeGenerationData SynthesisComputational EfficiencyTransformerSupervised Fine-TuningVideo

🎯 What it does: Proposes Radial Attention, a sparse attention mechanism with a complexity of O(n log n), designed for efficient generation of long videos;

RadZero: Similarity-Based Cross-Attention for Explainable Vision-Language Alignment in Chest X-ray with Zero-Shot Multi-Task Capability

Jonggwon Park (DEEPNOID Inc), Kyoyun Choi (DEEPNOID Inc)

CodeClassificationObject DetectionSegmentationExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningImageTextMultimodalityBiomedical DataComputed Tomography

🎯 What it does: A visual-language alignment framework called RadZero has been developed for zero-shot classification, localization, and segmentation on chest X-ray images, along with a similarity-based cross-attention mechanism (VL-CABS) and multi-positive contrastive learning.

RAG4GFM: Bridging Knowledge Gaps in Graph Foundation Models through Graph Retrieval Augmented Generation

Xingliang Wang (Zhejiang University), Shuiguang Deng (Zhejiang University)

CodeGenerationRetrievalGraph Neural NetworkGraphRetrieval-Augmented Generation

🎯 What it does: The RAG4GFM framework is proposed, which enhances retrieval-augmented generation for graph foundational models through multi-level graph indexing, task-aware retrieval, and graph fusion, improving knowledge update efficiency and inference credibility.

RAGRouter: Learning to Route Queries to Multiple Retrieval-Augmented Language Models

Jiarui Zhang (Shanghai Jiao Tong University), Guihai Chen (Shanghai Jiao Tong University)

CodeRetrievalTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation

🎯 What it does: This paper proposes a multi-model routing method called RAGRouter for retrieval-augmented generation (RAG) scenarios, which can dynamically select the most suitable large language model (LLM) to generate answers based on queries and retrieved documents.

Rainbow Delay Compensation: A Multi-Agent Reinforcement Learning Framework for Mitigating Observation Delays

Songchen Fu (Institute of Acoustics, Chinese Academy of Sciences), Yonghong Yan (Institute of Acoustics, Chinese Academy of Sciences)

CodeKnowledge DistillationRecurrent Neural NetworkTransformerReinforcement LearningSequential

🎯 What it does: To address the common issue of observation delay in multi-agent reinforcement learning, the Rainbow Delay Compensation (RDC) framework is proposed, which achieves adaptive compensation for delays through modules such as a compensator that reconstructs delay-free observations, delay-aligned critics, curriculum learning strategies, and knowledge distillation.

Random Forest Autoencoders for Guided Representation Learning

Adrien Aumon (UniversitΓ© de MontrΓ©al), Jake Slater Rhodes

CodeRepresentation LearningAuto EncoderImageTabularBiomedical Data

🎯 What it does: This paper proposes Random Forest Autoencoders (RF‑AE), a supervised visualization method that combines neighborhood information generated by random forests with the structure of autoencoders, capable of efficiently generating low-dimensional embeddings for unlabeled new samples.

Random Search Neural Networks for Efficient and Expressive Graph Learning

Michael Ito (University of Michigan), Jenna Wiens (University of Michigan)

CodeDrug DiscoveryRecurrent Neural NetworkGraph Neural NetworkGraphBiomedical Data

🎯 What it does: This paper proposes a Random Search Neural Network (RSNN), which replaces random walks with depth-first search (DFS) to address the insufficient expressiveness of RWNN when sampling is limited. It is proven that RSNN can achieve full graph coverage with logarithmic search times on sparse graphs, making it a universal approximator.

Randomized-MLP Regularization Improves Domain Adaptation and Interpretability in DINOv2

Joel Valdivia Ortega (Helmholtz Munich), Tingying Peng (Helmholtz Munich)

CodeSegmentationDomain AdaptationExplainability and InterpretabilityTransformerContrastive LearningImageBiomedical DataComputed Tomography

🎯 What it does: This paper proposes Randomized-MLP (RMLP) regularization, which replaces the trainable MLP head in ViT (such as DINOv2) to better align features with semantics during the refinement process and enhance the interpretability of attention and feature maps.

Ranking-based Preference Optimization for Diffusion Models from Implicit User Feedback

Yi-Lun Wu (National Yang Ming Chiao Tung University), Hong-Han Shuai (National Yang Ming Chiao Tung University)

CodeGenerationRecommendation SystemOptimizationReinforcement LearningDiffusion modelImage

🎯 What it does: This paper proposes a preference optimization framework for text-to-image diffusion models based on inverse reinforcement learning, called Diffusion-DRO.

RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation

Zixun Wang (Chinese University of Hong Kong), Ben Dai (Chinese University of Hong Kong)

CodeSegmentationComputational EfficiencyImage

🎯 What it does: This paper proposes RankSEG-RMA, an algorithm that achieves efficient and compatible non-overlapping multi-class semantic segmentation through recursive matrix approximation.

Rao-Blackwell Gradient Estimators for Equivariant Denoising Diffusion

Vinh Tong (University of Stuttgart), Mathias Niepert

CodeGenerationData SynthesisProtein Structure PredictionDiffusion modelTabularSequential

🎯 What it does: This paper proposes Orbit Diffusion, which utilizes a Rao-Blackwellized gradient estimator to train diffusion models under symmetry constraints, significantly reducing gradient variance and accelerating convergence.

RAPTR: Radar-based 3D Pose Estimation using Transformer

Sorachi Kato (Mitsubishi Electric Research Laboratories), Petros Boufounos (Mitsubishi Electric Research Laboratories)

CodePose EstimationTransformerPoint Cloud

🎯 What it does: A radar-based 3D human pose estimation framework called RAPTR is proposed, which can directly estimate complete 3D poses from multi-view radar heatmaps under weak supervision conditions using only 3D bounding boxes (BBox) and 2D keypoint labels.

RAST: Reasoning Activation in LLMs via Small-model Transfer

Siru Ouyang (University of Illinois Urbana-Champaign), Jiawei Han (University of Illinois Urbana-Champaign)

CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark

🎯 What it does: Developed the RAST method, which injects the logit differences obtained from RL training of a small model into a large model during inference, thereby activating the inference capability of the large model and avoiding the high-cost RL training on the large model.

RAT: Bridging RNN Efficiency and Attention Accuracy via Chunk-based Sequence Modeling

Xiuying Wei (Γ‰cole Polytechnique FΓ©dΓ©rale de Lausanne), Caglar Gulcehre

CodeOptimizationComputational EfficiencyRecurrent Neural NetworkSupervised Fine-TuningTextSequential

🎯 What it does: This paper proposes a block cyclic attention architecture named RAT, which divides long sequences into several blocks. Within each block, a lightweight RNN is used to handle short-range dependencies, while global information interaction between blocks is achieved through softmax attention, thus balancing efficiency and accuracy.

Rationalized All-Atom Protein Design with Unified Multi-Modal Bayesian Flow

Hanlin Wu (Tsinghua University), Jingjing Liu (Tsinghua University)

CodeDrug DiscoveryProtein Structure PredictionFlow-based ModelMultimodalityBiomedical Data

🎯 What it does: A unified all-atom protein design model called ProBayes is proposed, capable of end-to-end generation of sequences, backbone, and side chains.

RayFusion: Ray Fusion Enhanced Collaborative Visual Perception

Shaohong Wang (Zhejiang University), Eryun Liu (Zhejiang University)

CodeObject DetectionAutonomous DrivingImage

🎯 What it does: Proposes RayFusion, a framework for collaborative visual perception using multi-vehicle perspective ray occupancy information;

Re-coding for Uncertainties: Edge-awareness Semantic Concordance for Resilient Event-RGB Segmentation

Nan Bao (Beihang University), Jia Li (Beihang University)

CodeSegmentationOptimizationAuto EncoderImageMultimodality

🎯 What it does: This paper proposes an Edge-awareness Semantic Concordance (ESC) framework that utilizes edge information to unify event and RGB features, achieving semantic segmentation under extreme conditions.

Reaction Prediction via Interaction Modeling of Symmetric Difference Shingle Sets

Runhan Shi (Shanghai Jiao Tong University), Yang Yang (Shanghai Jiao Tong University)

CodeDrug DiscoveryTransformerGraph

🎯 What it does: This paper proposes a chemical reaction prediction model called ReaDISH, based on molecular shingle symmetric difference and interactive attention, aimed at addressing the issues of traditional models being sensitive to input order and insufficient in substructure interaction modeling.

ReAgent-V: A Reward-Driven Multi-Agent Framework for Video Understanding

Yiyang Zhou (University of North Carolina Chapel Hill), Huaxiu Yao (University of North Carolina Chapel Hill)

CodeLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodality

🎯 What it does: Proposes the ReAgent-V framework, which enhances video understanding performance through multi-agent real-time rewards and multi-perspective reflection.

Real-DRL: Teach and Learn in Reality

Yanbing Mao (Wayne State University), Lui Sha (University of Illinois)

CodeRobotic IntelligenceReinforcement Learning

🎯 What it does: Proposes the Real-DRL framework, which implements safety-critical deep reinforcement learning (DRL) on real physical systems, achieving real-time learning and safety assurance through three interactive components: DRL-Student, PHY-Teacher, and Trigger.

Real-Time Execution of Action Chunking Flow Policies

Kevin Black (Physical Intelligence), Sergey Levine (UC Berkeley)

CodeOptimizationRobotic IntelligenceDiffusion modelFlow-based ModelVideo

🎯 What it does: A real-time block execution method RTC based on flow matching has been designed, capable of achieving asynchronous execution and continuous action streams without retraining.

Real-World Adverse Weather Image Restoration via Dual-Level Reinforcement Learning with High-Quality Cold Start

Fuyang Liu (Nanjing University of Science and Technology), Xiaowei Hu (South China University of Technology)

CodeRestorationReinforcement LearningImage

🎯 What it does: A high-precision large-scale weather dataset HFLS-Weather has been constructed, and a dual-layer reinforcement learning framework has been proposed to achieve real-time recovery of adverse weather images through local model refinement and global meta-controller dynamic scheduling.

Reasoning Beyond Points: A Visual Introspective Approach for Few-Shot 3D Segmentation

Changshuo Wang (University College London), Prayag Tiwari (Halmstad University)

CodeSegmentationDomain AdaptationContrastive LearningPoint Cloud

🎯 What it does: A pre-training-free point cloud few-shot semantic segmentation framework VIP-Seg is proposed.

Reasoning by Superposition: A Theoretical Perspective on Chain of Continuous Thought

Hanlin Zhu (University of California Berkeley), Yuandong Tian (Meta AI)

CodeTransformerLarge Language ModelGraphChain-of-Thought

🎯 What it does: This paper studies and demonstrates that using a continuous chain of thought (COC) allows a two-layer Transformer to efficiently solve the directed graph reachability problem, and the experimental results validate the superior performance of this method on real datasets.

REASONING COMPILER: LLM-Guided Optimizations for Efficient Model Serving

Sujun Tang (University of California San Diego), Hadi Esmaeilzadeh (University of California San Diego)

CodeOptimizationComputational EfficiencyLarge Language ModelMixture of ExpertsText

🎯 What it does: A compiler optimization framework called REASONING COMPILER is proposed, which combines large language models with Monte Carlo tree search to generate efficient transformation sequences for neural network code.

Reasoning is Periodicity? Improving Large Language Models Through Effective Periodicity Modeling

Yihong Dong (Peking University), Hong Mei (Advanced Institute of Big Data)

CodeTransformerLarge Language ModelText

🎯 What it does: The FANformer structure is proposed, embedding a Fourier Analysis Network into the attention mechanism of the Transformer to enhance periodic modeling, thereby improving the learning efficiency and performance of large language models.

Reasoning Models Better Express Their Confidence

Dongkeun Yoon (Korea Advanced Institute of Science and Technology), Minjoon Seo (Korea Advanced Institute of Science and Technology)

CodeLarge Language ModelTextChain-of-Thought

🎯 What it does: This study explores how large language models can express confidence more accurately during chain-of-thought (CoT) reasoning and demonstrates that slow thinking behavior can significantly enhance confidence calibration.