NeurIPS 2025 Papers — Page 8
Conference on Neural Information Processing Systems · 5275 papers
Channel Matters: Estimating Channel Influence for Multivariate Time Series
Muyao Wang (Xidian University), James Kwok (Hong Kong University of Science and Technology)
Anomaly DetectionGraph Neural NetworkTransformerTime Series
🎯 What it does: Proposed the Channel-wise Influence (ChInf) method, which quantifies the impact of each channel on model performance in multivariate time series, and based on this method, constructed two types of algorithms for anomaly detection and channel pruning.
Channel Simulation and Distributed Compression with Ensemble Rejection Sampling
Buu Phan (University of Toronto), Ashish J Khisti
CompressionAuto EncoderImage
🎯 What it does: This paper studies channel simulation and distributed compression problems, proposing a new coding scheme based on Ensemble Rejection Sampling (ERS) and providing theoretical bounds for distributed matching. Experiments are then conducted on synthetic Gaussian sources and the MNIST image dataset for validation.
Characterization and Learning of Causal Graphs from Hard Interventions
Zihan Zhou (Johns Hopkins University), Murat Kocaoglu (Johns Hopkins University)
Graph Neural NetworkGraph
🎯 What it does: This paper studies the learning of causal structures using multiple sets of hard intervention data (including optional observational distributions) in causal graphs with latent variables.
Characterizing control between interacting subsystems with deep Jacobian estimation
Adam Joseph Eisen, Ila R Fiete
Recurrent Neural NetworkTime SeriesOrdinary Differential Equation
🎯 What it does: A JacobianODE method based on deep learning is proposed to learn the Jacobian matrix of nonlinear systems from time series data and use it for control analysis between subsystems.
Characterizing the Expressivity of Fixed-Precision Transformer Language Models
Jiaoda Li (ETH Zurich), Ryan Cotterell (ETH Zurich)
TransformerText
🎯 What it does: This paper studies the expressive power of a Transformer language model with fixed precision, soft attention, strict future masking, and no positional encoding, and establishes a complete equivalence between this model and linear temporal logic fragments with only past operators, left deterministic polynomials, R-trivial monoids, and partially ordered DFAs.
ChartSketcher: Reasoning with Multimodal Feedback and Reflection for Chart Understanding
Muye Huang (Xi'an Jiaotong University), Jun Liu (Xi'an Jiaotong University)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: This paper presents ChartSketcher, a chart understanding method that achieves multimodal step-by-step reasoning through programmatic drawing and visual feedback.
ChatbotID: Identifying Chatbots with Granger Causality Test
Xiaoquan Yi (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextTime SeriesSequential
🎯 What it does: A dialogue-level LLM detection method named ChatbotID is proposed, which utilizes Granger causality tests to extract interactive dynamic features and jointly fine-tunes with LLM contextual embeddings to distinguish between human conversations and chatbot dialogues.
ChatVLA-2: Vision-Language-Action Model with Open-World Reasoning
Zhongyi Zhou (East China Normal University), Yi Xu (Midea Group)
Robotic IntelligenceTransformerMixture of ExpertsVision-Language-Action ModelImageTextMultimodality
🎯 What it does: This paper presents ChatVLA-2, a vision-language-action model based on dynamic Mixture-of-Experts, enabling robots to reason and execute tasks in an open world.
Checklists Are Better Than Reward Models For Aligning Language Models
Vijay Viswanathan (Carnegie Mellon University), Tongshuang Wu (Carnegie Mellon University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: By automatically generating checklists from instructions and using these checklists as reward signals, reinforcement learning is applied to large language models to enhance the model's ability to follow instructions.
ChemOrch: Empowering LLMs with Chemical Intelligence via Groundbreaking Synthetic Instructions
Yue Huang (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed and implemented the ChemOrch framework, which synthesizes instruction-response data in the field of chemistry through a two-stage process, and utilizes tools to ensure answer verifiability.
CHiQPM: Calibrated Hierarchical Interpretable Image Classification
Thomas Norrenbrock (Leibniz Universität Hannover), Bodo Rosenhahn (Leibniz Universität Hannover)
ClassificationExplainability and InterpretabilityImage
🎯 What it does: A calibratable hierarchical explanation model CHiQPM is proposed, which can generate interpretable global class representations and hierarchical local explanations while maintaining extremely high prediction accuracy, and incorporates calibratable ensemble predictions.
Chirality in Action: Time-Aware Video Representation Learning by Latent Straightening
Piyush Nitin Bagad (University of Oxford), Andrew Zisserman (University of Oxford)
RecognitionRepresentation LearningTransformerAuto EncoderContrastive LearningVideoBenchmark
🎯 What it does: A self-supervised video representation learning method called LiFT is designed, which utilizes linearized feature trajectories to extract time-sensitive video features, and introduces the chiral action recognition task to evaluate temporal sensitivity.
Chiron-o1: Igniting Multimodal Large Language Models towards Generalizable Medical Reasoning via Mentor-Intern Collaborative Search
Haoran Sun (Shanghai Artificial Intelligence Laboratory), Xiaosong Wang (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelSupervised Fine-TuningMultimodalityBiomedical DataChain-of-Thought
🎯 What it does: This paper proposes the Mentor-Intern Collaborative Search (MICS) multi-model collaborative search framework, constructs a multi-modal medical reasoning dataset MMRP, and trains a multi-modal medical model Chiron-o1 with strong reasoning capabilities based on this dataset.
CHPO: Constrained Hybrid-action Policy Optimization for Reinforcement Learning
Ao Zhou (Tongji University), Guang Chen (Tongji University)
OptimizationReinforcement Learning
🎯 What it does: A new constrained hybrid action policy optimization algorithm CHPO is proposed to simultaneously maximize rewards and meet cost constraints in parameterized action spaces.
ChromFound: Towards A Universal Foundation Model for Single-Cell Chromatin Accessibiltiy Data
Yifeng Jiao (Fudan University), Yuan Cheng (Fudan University)
ClassificationRepresentation LearningBiomedical Data
🎯 What it does: ChromFound is proposed, a foundational model specifically designed for single-cell chromatin accessibility data, aimed at addressing the challenges of high-dimensional sparsity and dynamic chromatin landscapes.
ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference
Xiang Liu (Hong Kong University of Science and Technology), Xiaowen Chu (Hong Kong University of Science and Technology)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper addresses the issue of excessive GPU memory usage by KV caching in large model inference, proposing the ChunkKV method, which compresses KV caches by treating continuous semantic chunks as compression units instead of individual tokens.
CIDD: Collaborative Intelligence for Structure-Based Drug Design Empowered by LLMs
Bowen Gao (Tsinghua University), Yanyan Lan (Tsinghua University)
Drug DiscoveryTransformerLarge Language ModelPrompt EngineeringBiomedical DataChain-of-Thought
🎯 What it does: A collaborative intelligent drug design framework (CIDD) has been designed and implemented, combining 3D molecular generation models with large language models (LLM). Through four modules of interactive analysis, design, reflection, and selection, it achieves a complete generation process from target binding pockets to drug candidate molecules.
Class conditional conformal prediction for multiple inputs by p-value aggregation
Jean-Baptiste Fermanian (University of Montpellier), Joseph Salmon (University of Montpellier)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: A category conditional conformal prediction method is proposed for instances with multiple images, utilizing the p-values generated from multiple images for aggregation to construct smaller and more reliable prediction sets.
Class-aware Domain Knowledge Fusion and Fission for Continual Test-Time Adaptation
Jiahuan Zhou (Peking University), Gang Hua (Amazon.com, Inc)
ClassificationDomain AdaptationTransformerPrompt EngineeringImage
🎯 What it does: For Continuous Testing Task Adaptation (CTTA), a Category-Aware Knowledge Fusion and Splitting (KFF) framework is proposed to dynamically learn and accumulate knowledge from different domains while avoiding catastrophic forgetting.
Class-wise Balancing Data Replay for Federated Class-Incremental Learning
Zhuang Qi (Shandong University), Xiangxu Meng (Shandong University)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: Proposes the FedCBDR method, which achieves class balance in federated incremental learning through global perspective data replay and task-aware temperature scaling.
Classical Planning with LLM-Generated Heuristics: Challenging the State of the Art with Python Code
Augusto B. Corrêa (University of Oxford), Jendrik Seipp (Linköping University)
OptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the use of large language models (LLMs) to automatically generate domain-specific heuristic functions, evaluates them on training tasks to select the best heuristics, and subsequently uses them to solve new discrete deterministic planning tasks.
CLAWS:Creativity detection for LLM-generated solutions using Attention Window of Sections
Keuntae Kim (Hanyang University), Yong Suk Choi (Hanyang University)
ClassificationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes a detection method called CLAWS, which uses the attention distribution of various parts of the prompt to determine whether the mathematical answers generated by LLM are Hallucinated, Creative, or Typical, achieving three-class classification without the need for manual evaluation.
CLEAR: Conv-Like Linearization Revs Pre-Trained Diffusion Transformers Up
Songhua Liu (Shanghai Jiao Tong University), Xinchao Wang (National University of Singapore)
GenerationComputational EfficiencyKnowledge DistillationTransformerDiffusion modelImage
🎯 What it does: Linearization of the pre-trained Diffusion Transformer (DiT) is performed, proposing the CLEAR (Convolutional Linearization) attention mechanism to significantly reduce the computational cost of high-resolution image generation.
CLiFT: Compressive Light-Field Tokens for Compute Efficient and Adaptive Neural Rendering
Zhengqing Wang (Simon Fraser University), Yasutaka Furukawa (Wayve)
GenerationComputational EfficiencyTransformerImage
🎯 What it does: A neural rendering framework based on Compressed Light Field Tokens (CLiFT) is proposed, achieving adjustable storage size and rendering computation for scene representation and view synthesis.
Clip-and-Verify: Linear Constraint-Driven Domain Clipping for Accelerating Neural Network Verification
Duo Zhou (University of Illinois Urbana-Champaign), Huan Zhang (University of Illinois Urbana-Champaign)
OptimizationComputational EfficiencyTransformerTabularBenchmark
🎯 What it does: This paper proposes the Clip-and-Verify framework, which utilizes linear constraints to trim the input domain and improve the intermediate layer boundaries during the Branch-and-Bound process of neural network verification, significantly reducing the number of subproblems and increasing the verification success rate.
CLIPGaussian: Universal and Multimodal Style Transfer Based on Gaussian Splatting
Kornel Howil (Jagiellonian University), Przemysław Spurek (Jagiellonian University)
Image TranslationGenerationOptimizationContrastive LearningGaussian SplattingImageVideoMultimodality
🎯 What it does: A general style transfer plugin based on Gaussian Splatting (CLIPGaussian) has been developed, which can achieve style transfer based on image or text prompts on 2D images, videos, 3D objects, and 4D dynamic scenes without increasing the model size.
CLIPTTA: Robust Contrastive Vision-Language Test-Time Adaptation
Marc Lafon (Conservatoire National des Arts et Métiers), Nicolas THOME
ClassificationDomain AdaptationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes CLIPTTA, a test-time adaptive method for visual language models like CLIP, which utilizes soft contrastive loss to achieve gradient updates consistent with the CLIP pre-training objective.
Closed-Form Training Dynamics Reveal Learned Features and Linear Structure in Word2Vec-like Models
Dhruva Karkada (University of California Berkeley), Michael R DeWeese
Contrastive LearningText
🎯 What it does: This paper derives a quadratic word vector model called QWEM through a fourth-order Taylor approximation of the word2vec loss, and deduces its gradient flow training dynamics and the closed-form solution of the final embedding, revealing that the model learns orthogonal linear subspaces in a rank-increasing manner.
Cloud4D: Estimating Cloud Properties at a High Spatial and Temporal Resolution
Jacob Lin (University of Oxford), Ronald Clark (University of Oxford)
TransformerImageVideo
🎯 What it does: Using a ground-based synchronized camera system, we estimate the liquid water content of clouds, cloud base/cloud top heights, and horizontal wind speed, constructing a four-dimensional cloud model with a resolution of 25 m and an update frequency of 5 s.
ClusterFusion: Expanding Operator Fusion Scope for LLM Inference via Cluster-Level Collective Primitive
Xinhao Luo (Shanghai Jiao Tong University), Minyi Guo (Shanghai Jiao Tong University)
TransformerLarge Language ModelText
🎯 What it does: Developed the ClusterFusion framework, utilizing cluster-level communication primitives to achieve a broader range of operator fusion in LLM inference.
Clustering via Hedonic Games: New Concepts and Algorithms
Gergely Csáji (ELTE Centre of Economic and Regional Studies), Ildikó Schlotter (ELTE Centre of Economic and Regional Studies)
GraphTabular
🎯 What it does: This study explores the connection between local stability (local popularity and local stability) based on individual agent preferences in Friend-Enmity Network (FEN) games and clustering problems, providing corresponding polynomial algorithms and hardness proofs.
CMoB: Modality Valuation via Causal Effect for Balanced Multimodal Learning
Jun Wang (Shanxi University), Jiye Liang (Shanxi University)
ClassificationVideoMultimodalityAudio
🎯 What it does: A sample-level multimodal contribution assessment method based on causal effects, CMoB, is proposed. It calculates the contribution of each modality in each sample through an information uncertainty benefit function and causal intervention, and then dynamically enhances and masks weak modalities to achieve balance in multimodal learning.
Co-PatcheR: Collaborative Software Patching with Component-specific Small Reasoning Models
Yuheng Tang (University of California), Wenbo Guo (University of California)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes Co-PatcheR, a collaborative software patch generation system that employs multiple models to complete the tasks of localization, generation, and verification.
Co-Regularization Enhances Knowledge Transfer in High Dimensions
Shuo Shuo Liu (Pennsylvania State University), Runze Li (Pennsylvania State University)
Domain AdaptationOptimizationSupervised Fine-TuningTabular
🎯 What it does: A common regularization transfer (CoRT) framework for multi-source high-dimensional generalized linear models is proposed, which directly imposes regularization constraints between the target and source models to achieve knowledge transfer.
Co-Reinforcement Learning for Unified Multimodal Understanding and Generation
Jingjing Jiang (Nanyang Technological University), Chao Ma (Shanghai Jiao Tong University)
GenerationReinforcement LearningImageTextMultimodalityBenchmark
🎯 What it does: By proposing the CoRL framework, the understanding and generation capabilities of the unified multimodal model are enhanced through a joint reinforcement learning approach, ultimately resulting in ULM-R1.
COALA: Numerically Stable and Efficient Framework for Context-Aware Low-Rank Approximation
Uliana Parkina (Higher School of Economics University), Maxim Rakhuba (Higher School of Economics University)
CompressionOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A context-aware low-rank approximation framework COALA based on regularization and without matrix inversion is proposed for compressing and fine-tuning large language models, maintaining numerical stability in scenarios with large calibration matrices and near-singular activation matrices.
Coarse-to-Fine 3D Part Assembly via Semantic Super-Parts and Symmetry-Aware Pose Estimation
Xinyi Zhang (Shandong University), Jian Sun (Xi'an Jiaotong University)
Pose EstimationOptimizationTransformerPoint Cloud
🎯 What it does: A two-stage coarse-fine granularity 3D component assembly framework CFPA is proposed, utilizing semantic superparts and symmetry-aware pose prediction.
Coarse-to-fine Q-Network with Action Sequence for Data-Efficient Reinforcement Learning
Younggyo Seo (University of California Berkeley), Pieter Abbeel (University of California Berkeley)
Robotic IntelligenceReinforcement Learning
🎯 What it does: Proposed the Coarse-to-fine Q-Network with Action Sequence (CQNAS), which learns the value function of action sequences in critic-only RL to improve data efficiency;
CoC-VLA: Delving into Adversarial Domain Transfer for Explainable Autonomous Driving via Chain-of-Causality Visual-Language-Action Model
Dapeng Zhang (Lanzhou University), Qingguo Zhou (Lanzhou University)
Domain AdaptationAutonomous DrivingExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelGenerative Adversarial NetworkImageMultimodality
🎯 What it does: This paper proposes an end-to-end adaptive driving framework CoC-VLA based on a visual language model, achieving long-tail case transfer from simulation to the real world through a teacher-student model and a discriminator.
CoCoA: A Minimum Bayes Risk Framework Bridging Confidence and Consistency for Uncertainty Quantification in LLMs
Roman Vashurin (Mohamed bin Zayed University of Artificial Intelligence), Maxim Panov (Mohamed bin Zayed University of Artificial Intelligence)
TransformerLarge Language ModelText
🎯 What it does: A unified CoCoA framework is proposed, combining the confidence of LLMs with the semantic consistency of diverse outputs for uncertainty quantification;
CoDA: Coordinated Diffusion Noise Optimization for Whole-Body Manipulation of Articulated Objects
Huaijin Pi (University of Hong Kong), Taku Komura (University of Hong Kong)
GenerationOptimizationRobotic IntelligenceDiffusion modelText
🎯 What it does: This paper proposes a coordinated diffusion noise optimization framework that can generate synchronized movements of the entire human body, hands, and joint objects from text input.
Code Graph Model (CGM): A Graph-Integrated Large Language Model for Repository-Level Software Engineering Tasks
Hongyuan Tao (Ant Group), Peng Di (Ant Group)
AI Code AssistantGraph Neural NetworkTransformerLarge Language ModelTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes the Code Graph Model (CGM) and the Graph RAG framework to achieve agent-free warehouse-level software engineering tasks;
CodeCrash: Exposing LLM Fragility to Misleading Natural Language in Code Reasoning
Man Ho LAM, Michael Lyu
Explainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Constructed the CODECRASH framework to stress test the robustness of LLM code reasoning through structural and natural language embedding perturbations.
CodeGEMM: A Codebook-Centric Approach to Efficient GEMM in Quantized LLMs
Gunho Park (NAVER Cloud), Dongsoo Lee (NAVER Cloud)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: We propose CodeGEMM, an efficient GEMM kernel for codebook quantization of LLMs, which pre-computes and caches the inner product of center vectors and activations, avoiding real-time dequantization.
CodeMerge: Codebook-Guided Model Merging for Robust Test-Time Adaptation in Autonomous Driving
Huitong Yang (University of Queensland), Yadan Luo (University of Queensland)
Domain AdaptationAutonomous DrivingOptimizationPoint Cloud
🎯 What it does: CodeMerge is proposed, a model fusion framework based on low-dimensional fingerprint codebooks for adaptive 3D perception during online testing, addressing the unstable optimization problem in high-variance tasks.
Codifying Character Logic in Role-Playing
Letian Peng (University of California San Diego), Jingbo Shang (University of California San Diego)
Large Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Convert character description text into executable structured code functions to achieve more persistent, updatable, and controllable role-playing logic;
CoFFT: Chain of Foresight-Focus Thought for Visual Language Models
Xinyu Zhang (Xi'an Jiaotong University), Mike Zheng Shou (National University of Singapore)
RecognitionGenerationComputational EfficiencyTransformerVision Language ModelImageMultimodalityChain-of-Thought
🎯 What it does: We propose CoFFT, a training-independent visual language model inference framework that significantly enhances visual reasoning capabilities by simulating human visual foresight and focused thinking through multi-stage iterations.
Cognitive Mirrors: Exploring the Diverse Functional Roles of Attention Heads in LLM Reasoning
Xueqi Ma (University of Melbourne), James Bailey (University of Melbourne)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes an explanatory framework that identifies attention heads responsible for different cognitive functions (such as retrieval, logical reasoning, mathematical computation, etc.) in large language models through multi-class detection methods, and validates the key role of these 'cognitive heads' in reasoning tasks.
Cognitive Predictive Processing: A Human-inspired Framework for Adaptive Exploration in Open-World Reinforcement Learning
boheng liu, Xia Wu (Beijing Institute of Technology)
Reinforcement LearningWorld ModelSequential
🎯 What it does: A Cognitive Predictive Processing (CPP) framework based on human cognitive processes is proposed for adaptive exploration in open-world reinforcement learning.
CogVLA: Cognition-Aligned Vision-Language-Action Models via Instruction-Driven Routing & Sparsification
Wei Li (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
Computational EfficiencyRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelMultimodality
🎯 What it does: This paper proposes CogVLA, a cognitive alignment visual-language-action framework based on instruction-driven routing and sparsification;
CoIDO: Efficient Data Selection for Visual Instruction Tuning via Coupled Importance-Diversity Optimization
Yichen Yan (Zhejiang University), Huan Li (Zhejiang University)
OptimizationData-Centric LearningLarge Language ModelSupervised Fine-TuningImageMultimodality
🎯 What it does: Proposes the COIDO framework, which efficiently selects visual instruction fine-tuning data using the importance and diversity of joint optimization;
COLA: Towards Efficient Multi-Objective Reinforcement Learning with Conflict Objective Regularization in Latent Space
Pengyi Li (Tianjin University), YAN ZHENG
OptimizationReinforcement Learning
🎯 What it does: This paper proposes the COLA framework for multi-objective reinforcement learning, utilizing shared latent space and conflict objective regularization to enhance sample efficiency and performance.
Collaborative and Confidential Junction Trees for Hybrid Bayesian Networks
Roberto Gheda (Delft University of Technology), Lydia Y. Chen (Delft University of Technology)
Federated LearningSafty and PrivacyTabular
🎯 What it does: This paper proposes the Hybrid CCJT framework, which achieves multi-party collaborative and confidential hybrid Bayesian network inference.
Collaborative Reasoner: Self-Improving Social Agents with Synthetic Conversations
Ansong Ni (Meta), Asli Celikyilmaz (Meta)
TransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes a collaborative reasoning framework called Coral, which utilizes multi-turn dialogue to evaluate and enhance the collaborative reasoning capabilities of LLMs;
Collapsing Taylor Mode Automatic Differentiation
Felix Dangel (Vector Institute), Andrea Walther (Humboldt Universitaet zu Berlin)
OptimizationComputational EfficiencyAuto Encoder
🎯 What it does: This paper proposes a 'contraction Taylor mode' optimization technique for the computation of high-order derivatives (such as PDE operators), which combines linear structures to achieve graph rewriting, significantly reducing the required vector passing. It is applicable to linear PDEs such as Laplace, weighted Laplace, biharmonic, and their stochastic approximations.
Collective Counterfactual Explanations: Balancing Individual Goals and Collective Dynamics
Ahmad Reza Ehyaei, Samira Samadi (Max Planck Institute for Intelligent Systems)
Recommendation SystemOptimizationExplainability and InterpretabilityTabular
🎯 What it does: This paper proposes a collective counterfactual explanation framework (CCE) that considers population dynamics and resource competition, capable of generating recommendations that minimize individual costs while also balancing group interests.
Color Conditional Generation with Sliced Wasserstein Guidance
Alexander Lobashev (Glam AI), Dmitry Guskov (Glam AI)
Image TranslationGenerationData SynthesisDiffusion modelImageText
🎯 What it does: A training-free color conditional generation method named SW-Guidance is proposed, which uses the Sliced Wasserstein distance to guide the diffusion model to match the color distribution of reference images.
Coloring Learning for Heterophilic Graph Representation
Miaomiao Huang (Northeastern University), Xingwei Wang (Northeastern University)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Proposes the CoRep framework, which utilizes graph coloring learning to achieve self-supervised representation of heterogeneous graphs;
CoLT: The conditional localization test for assessing the accuracy of neural posterior estimates
Tianyu Chen (University of Texas at Austin), James G. Scott (University of Texas at Austin)
🎯 What it does: A conditional location test (CoLT) is proposed to verify whether the posterior distribution output by the neural posterior estimator (NPE) is consistent with the true posterior.
Combinatorial Ski Rental Problem: Robust and Learning-Augmented Algorithms
Ziwei Li (Nanjing University), Yang Gao (Nanjing University)
OptimizationTabular
🎯 What it does: This paper proposes and analyzes the Combinatorial Ski Rental (CSR) problem, designs the optimal stochastic online algorithm SOAC, and builds a learning-enhanced algorithm LA-SOAC based on it.
Combining Cost Constrained Runtime Monitors for AI Safety
Tim Tian Hua (MARS), Tyler Tracy (Redwood Research)
OptimizationSafty and PrivacyAI Code AssistantLarge Language ModelText
🎯 What it does: A cost-constrained runtime monitoring protocol is proposed, which can combine multiple monitors into an optimal security intervention strategy.
COME: Adding Scene-Centric Forecasting Control to Occupancy World Model
Yining Shi (Tsinghua University), Diange Yang (Tsinghua University)
GenerationAutonomous DrivingOptimizationTransformerDiffusion modelWorld ModelPoint Cloud
🎯 What it does: A framework named COME is proposed, which enhances the generation effect of occupancy world models through scene-centered predictive control.
ComfyMind: Toward General-Purpose Generation via Tree-Based Planning and Reactive Feedback
Litao Guo (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)
GenerationLarge Language ModelVision Language ModelImageVideoBenchmarkRetrieval-Augmented Generation
🎯 What it does: A collaborative AI system called ComfyMind has been built on the open-source ComfyUI platform to achieve general visual generation tasks (image and video generation and editing).
Communication-Efficient Diffusion Denoising Parallelization via Reuse-then-Predict Mechanism
Kunyun Wang (Shanghai Jiao Tong University), Jieru Zhao (Shanghai Jiao Tong University)
RestorationComputational EfficiencyDiffusion modelImageAudio
🎯 What it does: Two methods, ParaStep and BatchStep, are proposed to accelerate diffusion model inference through a reuse-repredict mechanism, significantly reducing communication overhead.
Communication-Efficient Language Model Training Scales Reliably and Robustly: Scaling Laws for DiLoCo
Zachary Charles (Google Research), Arthur Douillard (Google DeepMind)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A systematic study of the communication-efficient algorithm DiLoCo used in large-scale language model training is conducted, establishing scaling laws for evaluation loss and hyperparameters as they vary with model size, number of replicas, and synchronization periods, and comparing it with traditional data parallel training.
Compact Memory for Continual Logistic Regression
Yohan Jung (RIKEN Center for AI Project), Mohammad Emtiyaz Khan (Imperial College London)
ClassificationOptimizationImage
🎯 What it does: A PPCA method based on Hessian matching is proposed to construct compact memory (memory vectors and weights) for logistic regression in continual learning, and to achieve approximate reconstruction of past task gradients through this memory, thereby alleviating catastrophic forgetting.
Comparator-Adaptive $\Phi$-Regret: Improved Bounds, Simpler Algorithms, and Applications to Games
Soumita Hait (University of Southern California), Mengxiao Zhang (University of Iowa)
Optimization
🎯 What it does: This paper focuses on the study of Φ-regret in expert problems, proposing two simpler and stronger comparative adaptive algorithms, and generalizing one of the algorithms to general multi-player games, achieving accelerated and adaptive Φ-equilibrium convergence speed.
Comparing Uniform Price and Discriminatory Multi-Unit Auctions through Regret Minimization
Marius Potfer (ENSAE), Vianney Perchet (ENSAE)
OptimizationReinforcement Learning
🎯 What it does: The study compares the learning difficulty of uniform price auctions and discriminatory price auctions in repeated multi-unit auctions, providing optimal regret upper and lower bounds under full information and bandit settings, and further analyzes the impact of instance differences and special opponent distributions.
Competitive Advantage Attacks to Decentralized Federated Learning
Yuqi Jia (Duke University), Neil Zhenqiang Gong
OptimizationFederated LearningAdversarial AttackImageText
🎯 What it does: This paper proposes SelfishAttack, targeting selfish clients in decentralized federated learning (DFL). It customizes the shared model to enhance the accuracy of its own model while suppressing the competitive advantage of non-selfish clients.
Compiler-R1: Towards Agentic Compiler Auto-tuning with Reinforcement Learning
Haolin Pan (Hangzhou Institute for Advanced Study), Yanjun Wu (Institute of Software Chinese Academy of Sciences)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningChain-of-Thought
🎯 What it does: Developed the Compiler-R1 framework, which utilizes reinforcement learning to drive large language models (LLM) for automatic compiler tuning, automatically generating optimization pass sequences for LLVM IR.
Complete Structure Guided Point Cloud Completion via Cluster- and Instance-Level Contrastive Learning
Yang Chen (Fudan University), Cheng Jin (Fudan University)
GenerationData SynthesisContrastive LearningPoint Cloud
🎯 What it does: A self-supervised point cloud completion framework CSG-PCC is proposed, which guides point cloud completion through self-learning of complete structures.
Complexity Scaling Laws for Neural Models using Combinatorial Optimization
Lowell Weissman (Virginia Tech), A. Lynn Abbott (Virginia Tech)
OptimizationTransformerSupervised Fine-TuningReinforcement LearningTabularBenchmark
🎯 What it does: This paper conducts large-scale experiments on the Traveling Salesman Problem (TSP) to study the scalability of neural networks under fixed model capacity, proposing a 'problem complexity scaling law' based on the size of the solution space and the representation space.
Compliant Residual DAgger: Improving Real-World Contact-Rich Manipulation with Human Corrections
Xiaomeng Xu (Stanford University), Shuran Song (Stanford University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelMultimodality
🎯 What it does: The CR-DAgger method is proposed, combining compliant intervention interfaces and residual strategies to improve contact manipulation in real scenarios.
ComPO: Preference Alignment via Comparison Oracles
Peter Chen (Columbia University), Tianyi Lin (Columbia University)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes a zero-order preference alignment method based on comparison or acle, utilizing noisy preference alignment data to improve LLM.
Composing Global Solutions to Reasoning Tasks via Algebraic Objects in Neural Nets
Yuandong Tian (Meta Superintelligence Lab)
OptimizationTabular
🎯 What it does: This paper studies a two-dimensional hidden layer neural network trained on Abelian group reasoning tasks (such as modular addition), revealing the semiring algebraic structure of its weight space and loss function, and constructs a global optimal solution based on this;
Composing Linear Layers from Irreducibles
Travis Pence (University of Wisconsin-Madison), Vikas Singh (University of Wisconsin-Madison)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Using Clifford algebra to decompose linear layers into geometric primitives (bivectors) and achieve differentiable linear mappings through rotor sandwiching, forming a parameter-compressed linear module.
Composite Flow Matching for Reinforcement Learning with Shifted-Dynamics Data
Lingkai Kong (Harvard University), Milind Tambe (Harvard University)
Reinforcement LearningFlow-based ModelTabular
🎯 What it does: To address the dynamic shift between offline and online data in reinforcement learning, the COMPFLOW method is proposed, which uses combined flow matching to estimate the dynamic gap and designs an active exploration strategy based on the gap results.
Composition and Alignment of Diffusion Models using Constrained Learning
Shervin Khalafi (University of Pennsylvania), Alejandro Ribeiro (University of Pennsylvania)
GenerationOptimizationDiffusion modelImageText
🎯 What it does: This paper proposes a unified constrained learning framework for aligning and composing diffusion models, generating samples that meet multiple constraints by imposing constraints on the KL divergence between pre-trained models and reward functions or multiple pre-trained models.
Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models
Samuel Lavoie (Mila Université de Montréal), Aaron Courville (Mila Université de Montréal)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: This paper proposes and utilizes Discrete Latent Code (DLC) to condition diffusion models, achieving high-fidelity unconditional image generation on ImageNet, and enabling image composability and text-to-image generation through DLC.
Compositional Monte Carlo Tree Diffusion for Extendable Planning
Jaesik Yoon (KAIST & SAP), Sungjin Ahn (KAIST & NYU)
Diffusion modelGraph
🎯 What it does: The C-MCTD framework is proposed, which utilizes diffusion models and Monte Carlo tree search to concatenate short trajectories into long, globally consistent planning trajectories during inference.
Compositional Neural Network Verification via Assume-Guarantee Reasoning
Hai Duong (George Mason University), Matthew B. Dwyer (University of Virginia)
Computational EfficiencyConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A parameterizable hypothesis-guarantee compositional framework called CoVeNN is proposed to decompose large-scale neural networks into sub-networks and verify them incrementally, significantly reducing memory consumption and enhancing verifiability.
Compositional Reasoning with Transformers, RNNs, and Chain of Thought
Gilad Yehudai (New York University), Joan Bruna (New York University)
Recurrent Neural NetworkTransformerChain-of-Thought
🎯 What it does: This paper proposes and studies a class of tree-structured reasoning tasks known as Compositional Reasoning Questions (CRQ); it systematically compares the expressive power and resource requirements of different models (deep Transformers, RNNs, and shallow Transformers with Chain of Thought (CoT)) in solving CRQs; it proves that under standard hardware assumptions, all three architectures require an increase in some hyperparameter (number of Transformer layers, RNN hidden dimensions, or number of CoT tokens) as the input size grows.
Compress & Cache: Vision token compression for efficient generation and retrieval
Adrian Bulat (Samsung AI Cambridge), Georgios Tzimiropoulos (Technical University of Iasi)
GenerationRetrievalCompressionTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: The Compress & Cache (C&C) method is proposed, utilizing the LVLM itself to compress visual tokens during the offline phase and cache them, and then directly using the compressed summary tokens for generation and retrieval during online inference;
Compress Large Language Models via Collaboration Between Learning and Matrix Approximation
Yuesen Liao (Fudan University), WEIZHONG ZHANG
CompressionOptimizationLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes a dual-layer optimization-based LLM compression framework that jointly learns hierarchical sparse allocation and matrix low-rank approximation to achieve a composite compression of sparsity and low-rank.
Compress to Impress: Efficient LLM Adaptation Using a Single Gradient Step on 100 Samples
Shiva Sreeram (Massachusetts Institute of Technology), Daniela Rus (Massachusetts Institute of Technology)
CompressionDomain AdaptationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A method for adapting LLMs without fine-tuning is proposed, which only requires a single gradient computation on 100 samples and quickly compresses part of the weight matrix;
Compress, Gather, and Recompute: REFORMing Long-Context Processing in Transformers
Woomin Song (Korea Advanced Institute of Science and Technology), Sravan Babu Bodapati (Amazon Web Services)
CompressionComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: A long-context reasoning framework called REFORM is proposed, which can handle up to millions of contexts without changing the model parameters.
Compressed and Smooth Latent Space for Text Diffusion Modeling
Viacheslav Meshchaninov (Higher School of Economics University), Dmitry Vetrov (Constructor University)
GenerationCompressionTransformerDiffusion modelAuto EncoderText
🎯 What it does: Designed and implemented COSMOS, which utilizes a compressed and smooth latent space to achieve text generation in diffusion models.
Computable universal online learning
Dariusz Kalociński (Institute of Computer Science Polish Academy of Sciences), Tomasz Steifer (Institute of Fundamental Technological Research Polish Academy of Sciences)
🎯 What it does: This paper studies the introduction of computability constraints in online binary classification learning, exploring which hypothesis classes can be computationally completed by learning programs.
Computation and Memory-Efficient Model Compression with Gradient Reweighting
Zhiwei Li (Fudan University), WEIZHONG ZHANG
CompressionComputational EfficiencyLarge Language ModelImageText
🎯 What it does: A sparse pruning framework is proposed that utilizes gradient reweighting and on-demand instantiation of sub-models, significantly reducing memory usage and computational costs during the pruning process.
Computational Algebra with Attention: Transformer Oracles for Border Basis Algorithms
Hiroshi Kera (Chiba University), Sebastian Pokutta (Zuse Institute Berlin)
Computational EfficiencyTransformer
🎯 What it does: A new boundary basis algorithm is proposed, utilizing deep learning to accelerate the solving of polynomial equation systems while ensuring the correctness of the output.
Computational Budget Should Be Considered in Data Selection
Weilin Wan (Fudan University), Cheng Jin (Fudan University)
OptimizationComputational EfficiencyData-Centric LearningSupervised Fine-TuningImageText
🎯 What it does: This paper proposes a budget-aware data selection framework called CADS, which selects the optimal subset to enhance model training performance under a fixed computational budget.
Computational Efficiency under Covariate Shift in Kernel Ridge Regression
Andrea Della Vecchia (Swiss Finance Institute), Lorenzo Rosasco (University of Genova)
Computational EfficiencyTime Series
🎯 What it does: The study improves the computational efficiency of kernel ridge regression under covariate shift conditions through Nyström random projection and provides the corresponding statistical error upper bound.
Computational Hardness of Reinforcement Learning with Partial $q^{\pi}$-Realizability
Shayan Karimi (University of Alberta), Xiaoqi Tan (University of Alberta)
Reinforcement Learning
🎯 What it does: This paper studies reinforcement learning under the framework of partial qπ-realizability, proving that finding ε-optimal policies is computationally difficult in two common policy sets (greedy policy set and softmax policy set); specifically, it is NP-hard under the greedy policy set, and there exists an exponential lower bound under the softmax policy set assuming rETH.
Compute-Optimal Scaling for Value-Based Deep RL
Preston Fu (University of California Berkeley), Aviral Kumar (Carnegie Mellon University)
Computational EfficiencyRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: The research focuses on the computationally optimal scaling of deep reinforcement learning, exploring how model size, batch size, and update-to-data ratio (UTD) can maximize data efficiency under a fixed computational budget.
ComRank: Ranking Loss for Multi-Label Complementary Label Learning
Jing-Yi Zhu (Southeast University), Min-Ling Zhang (University of Queensland)
ClassificationOptimizationImageTextBiomedical Data
🎯 What it does: A ranking loss framework called ComRank is proposed for multi-label complementary label learning, addressing the failure of traditional unbiased risk estimation methods under non-uniform complementary label distributions.
Concentration and excess risk bounds for imbalanced classification with synthetic oversampling
Touqeer Ahmad (University of Angers), Gilles Stupfler (University of Rennes)
ClassificationData SynthesisTabular
🎯 What it does: This paper constructs a theoretical framework for synthetic oversampling methods such as SMOTE and KDEO in imbalanced classification, deriving their convergence properties for empirical risk and overall risk, and providing an upper bound on the excess risk of kernel methods and empirical risk minimization based on these synthetic samples.
Concept Incongruence: An Exploration of Time and Death in Role Playing
Xiaoyan Bai (University of Chicago), Chenhao Tan (University of Chicago)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the conceptual inconsistency caused by the conflict between the time of character death and query time in role-playing scenarios using large language models. It proposes three behavioral metrics (abandonment rate, response rate, conditional accuracy) and reveals the model's insufficient internal encoding of death states and times, leading to a decrease in accuracy through dialogue experiments and linear probe analysis.
Concept-Guided Interpretability via Neural Chunking
Shuchen Wu (Allen Institute), Zeynep Akata (Institute for Human-Centered AI)
Explainability and InterpretabilityRecurrent Neural NetworkTransformerLarge Language ModelTextSequential
🎯 What it does: Proposes the 'Reflection Hypothesis', suggesting that the internal activities of neural networks can reflect the structure of training data, and utilizes the chunking principle in cognition to extract interpretable entities from neural population activities;
ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts
Jinho Choi (KAIST), Jaegul Choo (KAIST)
ClassificationData-Centric LearningLarge Language ModelVision Language ModelAuto EncoderImageMultimodality
🎯 What it does: This paper proposes a framework called ConceptScope for the automatic discovery and quantification of visual concepts in image datasets, aimed at identifying dataset biases and assessing model robustness.
Concerto: Joint 2D-3D Self-Supervised Learning Emerges Spatial Representations
Yujia Zhang (Chinese University of Hong Kong), Hengshuang Zhao (Chinese University of Hong Kong)
SegmentationRepresentation LearningTransformerContrastive LearningImageVideoPoint Cloud
🎯 What it does: Proposes Concerto - a joint 2D-3D self-supervised learning framework that simulates human multi-sensory concept learning.