NeurIPS 2025 Papers — Page 10
Conference on Neural Information Processing Systems · 5275 papers
Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers
Youmin Ko (Hanyang University), Hyunjoon Kim (Hanyang University)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes the CoopRAG framework, which first breaks down the problem into sub-problems and uncertain reasoning chains, utilizes a retriever to search for documents, performs hierarchical comparative re-ranking of the search results, and finally allows the LLM to fill in missing information and provide answers.
CoP: Agentic Red-teaming for Large Language Models using Composition of Principles
Chen Xiong (Chinese University of Hong Kong), Tsung-Yi Ho (Chinese University of Hong Kong)
Explainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: A framework for agentic red-teaming based on Composition-of-Principles (CoP) is proposed, which automatically generates and iteratively improves jailbreak prompts using human-provided 'red team principles' to achieve single-round cracking of large language models;
Copresheaf Topological Neural Networks: A Generalized Deep Learning Framework
Mustafa Hajij (Technical University of Munich), Tolga Birdal (Stanford University)
Graph Neural NetworkSpiking Neural NetworkTransformerGraphPhysics Related
🎯 What it does: The Copresheaf Topological Neural Networks (CTNN) framework is proposed, utilizing the directional information flow of Copresheaf on combinatorial complexes to achieve unified deep learning for multi-scale and heterogeneous data.
CORAL: Disentangling Latent Representations in Long-Tailed Diffusion
Esther Rodriguez (Arizona State University), Lalitha Sankar (Arizona State University)
GenerationData SynthesisDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes an improved diffusion model method called CORAL for long-tail data distribution, addressing the issue of low-quality generation in tail classes.
CORE: Collaborative Optimization with Reinforcement Learning and Evolutionary Algorithm for Floorplanning
Pengyi Li (Tianjin University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationReinforcement Learning
🎯 What it does: A hybrid evolutionary reinforcement learning framework called CORE based on B*-Tree representation is proposed to solve the floorplanning problem in electronic design automation.
CORE: Reducing UI Exposure in Mobile Agents via Collaboration Between Cloud and Local LLMs
Gucongcong Fan (Shanghai Jiao Tong University), Guihai Chen (Shanghai Jiao Tong University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: The CORE framework is proposed, which enables collaboration between a local lightweight LLM and a powerful cloud LLM to minimize the upload of mobile UI information while maintaining task success rates.
CoreGuard: Safeguarding Foundational Capabilities of LLMs Against Model Stealing in Edge Deployment
Qinfeng Li (Zhejiang University), Jianwei Yin (Zhejiang University)
Safty and PrivacyLarge Language ModelText
🎯 What it does: This study proposes CoreGuard, which provides protection against model theft for proprietary large language models (LLMs) deployed on edge devices;
Coreset for Robust Geometric Median: Eliminating Size Dependency on Outliers
Ziyi Fang (Nanjing University), Runkai Yang (Nanjing University)
OptimizationTabular
🎯 What it does: This study constructs a coreset for robust geometric medians, proposing to eliminate the dependence on the number of outliers m under the condition n≥4m, and provides the optimal coreset size in one dimension and upper bounds for multi-dimensional and k-clustering.
Coresets for Clustering Under Stochastic Noise
Lingxiao Huang (Nanjing University), Haoyu Zhao (Princeton University)
OptimizationTabular
🎯 What it does: When the data is contaminated by random noise from a known distribution, this paper studies and constructs a compact representative subset (coreset) that can approximate the original (noise-free) dataset, and provides corresponding quality guarantees.
Corporate Needs You to Find the Difference: Revisiting Submodular and Supermodular Ratio Optimization Problems
Elfarouk Harb (University of Illinois at Urbana Champaign), Chandra Chekuri (University of Illinois at Urbana Champaign)
Recommendation SystemOptimizationComputational EfficiencyGraphFinance Related
🎯 What it does: This paper proposes a unified 'universal solver'—SUPERGREEDY++, Frank-Wolfe, and Fujishige-Wolfe FW-MNP—by proving the exact and approximate equivalence of SFM, DSS, USSS, UDSS with the minimum norm point (MNP), which can directly solve multi-class sub/supermodular ratio problems and conduct large-scale experiments on tasks such as HNSN and minimum s-t-cut.
Correcting misinterpretations of additive models
Benedict Clark (Physikalisch-Technische Bundesanstalt), Stefan Haufe (Charité - Universitätsmedizin)
Explainability and InterpretabilityImageTabularElectronic Health Records
🎯 What it does: Two activation pattern-based explanation methods, PatternGAM and PatternQLR, are proposed to correct misinterpretations in additive models, particularly to suppress errors in variable importance caused by confounding variables.
Corrector Sampling in Language Models
Itai Gat (Meta), Yaron Lipman (Meta)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A Resample-Previous-Tokens (RPT) sampling method is proposed, allowing the model to backtrack and resample the most recent tokens during the generation process, thereby reducing error accumulation; fine-tuning on 100B corpus with an 8B parameter model achieves approximately a 10% relative improvement.
Correlated Low-Rank Adaptation for ConvNets
Wu Ran (Shanghai Jiao Tong University), Chao Ma (Shanghai Jiao Tong University)
Domain AdaptationOptimizationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A parameter-efficient fine-tuning method for convolutional networks called CoLoRA is proposed, which utilizes low-rank shared matrices of adjacent layers to achieve correlation learning.
Correlation Dimension of Autoregressive Large Language Models
Xin Du (Waseda University), Kumiko Tanaka-Ishii (Waseda University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes using the correlation dimension to evaluate the long-range text structure of autoregressive large language models, constructing an indicator that can instantly compute and capture the self-similarity of text during inference.
COS3D: Collaborative Open-Vocabulary 3D Segmentation
Runsong Zhu (Chinese University of Hong Kong), Chi-Wing Fu (Chinese University of Hong Kong)
SegmentationVision Language ModelGaussian SplattingPoint Cloud
🎯 What it does: The COS3D framework is proposed, utilizing collaborative fields (instance fields + language fields) to achieve open vocabulary 3D segmentation, supporting real-time 3D region extraction based on text or images.
Cost-Aware Contrastive Routing for LLMs
Reza Shirkavand (University of Maryland), Heng Huang (University of Maryland)
Recommendation SystemOptimizationComputational EfficiencyTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a cost-aware contrastive learning-based routing framework called CSCR, which selects the most suitable and cost-effective model for each input from a pool of multi-model LLMs.
Cost-aware LLM-based Online Dataset Annotation
Eray Can Elumar (Carnegie Mellon University), Osman Yagan (Carnegie Mellon University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: The Cost-aware Majority Voting (CaMVo) framework is proposed for online selection of the minimum cost subset of LLMs and weighted voting for automated dataset labeling without the premise of a training set or labels.
Cost-Efficient LLM Training with Lifetime-Aware Tensor Offloading via GPUDirect Storage
Ziqi Yuan (University of Illinois Urbana-Champaign), Jian Huang (University of Illinois Urbana-Champaign)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A lifespan-aware tensor offloading framework TERAIO based on PCIe SSD has been designed and implemented for training large language models, utilizing GPUDirect Storage for direct and efficient migration between GPU and SSD;
Cost-Sensitive Freeze-thaw Bayesian Optimization for Efficient Hyperparameter Tuning
Dong Bok Lee (Korea Advanced Institute of Science and Technology), Hae Beom Lee (Korea University)
OptimizationHyperparameter SearchTransformerTabular
🎯 What it does: A cost-sensitive freeze-thaw Bayesian optimization method is proposed to achieve early stopping in multi-precision hyperparameter optimization based on a user-defined utility function.
CoT Information: Improved Sample Complexity under Chain-of-Thought Supervision
Awni Altabaa (Yale University), John Lafferty (Yale University)
SequentialChain-of-Thought
🎯 What it does: This paper proposes and analyzes the statistical learning theory under Chain of Thought (CoT) supervision, focusing on how CoT supervision improves the learning efficiency of multi-step reasoning models.
CoT Red-Handed: Stress Testing Chain-of-Thought Monitoring
Benjamin Arnav (LASR Labs), Mary Phuong (LASR Labs)
Anomaly DetectionAI Code AssistantLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper studies the effectiveness of chain-of-thought (CoT) monitoring compared to action-only review in detecting potential malicious behavior in AI control frameworks, and proposes a hybrid monitoring scheme that integrates both approaches.
CoT-lized Diffusion: Let's Reinforce T2I Generation Step-by-step
Zheyuan Liu (Peking University), Li Yuan (Peking University)
GenerationData SynthesisLarge Language ModelDiffusion modelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the CoT-Diff framework, which tightly couples multimodal large language models (MLLM) with diffusion models in a single sampling trajectory, achieving stepwise generation and dynamic correction based on 3D scene layouts.
CoUn: Empowering Machine Unlearning via Contrastive Learning
Yasser H. Khalil (Huawei Noah's Ark Lab), Hongliang Li (Huawei Noah's Ark Lab)
Computational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Proposes the CoUn framework, which combines contrastive learning and supervised learning on retained data to indirectly adjust the representations of forgotten samples, achieving efficient machine unforgetting (MU).
Counteractive RL: Rethinking Core Principles for Efficient and Scalable Deep Reinforcement Learning
Ezgi Korkmaz
Reinforcement LearningSequentialBenchmark
🎯 What it does: This paper proposes a novel experience collection strategy called CoAct TD learning, which enhances temporal difference error by utilizing 'adversarial' actions that minimize the state-action value function, thereby significantly improving sample efficiency without increasing additional computational costs.
Counterfactual Evolution of Multimodal Datasets via Visual Programming
Minghe Gao (Zhejiang University), Juncheng Li (Zhejiang University)
GenerationData SynthesisAdversarial AttackLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A controllable dataset evolution framework SCOPE based on verifiable visual programming is proposed, and an evolvable SCOPE-Train/SCOPE-Test benchmark is constructed.
Counterfactual Identifiability via Dynamic Optimal Transport
Fabio De Sousa Ribeiro (Imperial College London), Ben Glocker (Imperial College London)
Flow-based ModelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper studies how to achieve counterfactual identification in high-dimensional multivariate settings through continuous-time flows and dynamic optimal transport.
Counterfactual Image Editing with Disentangled Causal Latent Space
Yushu Pan (Columbia University), Elias Bareinboim (Columbia University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A causal structure-based image editing framework has been developed, utilizing a pre-trained diffusion model to achieve reverse causal editing of images without the need for retraining.
Counterfactual Implicit Feedback Modeling
Chuan Zhou (University of Melbourne), Mingming Gong (University of Melbourne)
Recommendation SystemContrastive LearningTabular
🎯 What it does: Modeling the relevance prediction problem of implicit feedback as a counterfactual estimation problem with missing data handling, and proposing the Counter-IF method.
Counterfactual Reasoning for Steerable Pluralistic Value Alignment of Large Language Models
Hanze Guo (Renmin University of China), Xing Xie (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes the COUPLE framework, which utilizes structural causal models and counterfactual reasoning to achieve multi-dimensional value alignment for LLMs, supporting adjustable fine-grained value priorities;
Counterfactual reasoning: an analysis of in-context emergence
Moritz Miller (Max Planck Institute for Intelligent Systems), Siyuan Guo (University of Cambridge)
TransformerLarge Language ModelTime SeriesSequentialStochastic Differential Equation
🎯 What it does: This paper studies the ability of large-scale language models (especially Transformers) to perform counterfactual reasoning in context, designing synthetic linear regression tasks and SDE-based dynamical system tasks to evaluate how models achieve noise elimination and counterfactual prediction through self-attention and depth, providing corresponding theoretical and mechanistic explanations.
Coupled Data and Measurement Space Dynamics for Enhanced Diffusion Posterior Sampling
Shayan Mohajer Hamidi (Stanford University), EN-HUI YANG
RestorationDiffusion modelImageBenchmark
🎯 What it does: A new framework called Coupled Data and Measurement Space Diffusion Posterior Sampling (C-DPS) is proposed to address signal recovery in inverse problems, particularly in the presence of noise or incomplete measurements.
Coupling Generative Modeling and an Autoencoder with the Causal Bridge
ruolin meng, Lawrence Carin (Duke University)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkTabularTime SeriesElectronic Health Records
🎯 What it does: In the presence of unobserved confounding variables, this study investigates the use of two sets of proxy variables to construct causal bridge functions and proposes a framework for joint learning through generative models and autoencoders to enhance the accuracy of causal effect estimation; this framework is also extended to time-to-event (survival) analysis.
Covariances for Free: Exploiting Mean Distributions for Training-free Federated Learning
Dipam Goswami (Universitat Autònoma de Barcelona), Joost van de Weijer (Universitat Autònoma de Barcelona)
Federated LearningImage
🎯 What it does: In federated learning, a pre-trained model is used to estimate class covariance solely through the class means sent by clients, initializing the global classifier to achieve unsupervised federated learning.
Covariate-moderated Empirical Bayes Matrix Factorization
William R.P. Denault, Matthew Stephens (University of Chicago)
Recommendation SystemTabular
🎯 What it does: A new covariate-modulated empirical Bayesian matrix factorization (cEBMF) framework is proposed, which can utilize any form of side information to improve matrix factorization.
Covering Multiple Objectives with a Small Set of Solutions Using Bayesian Optimization
Natalie Maus (University of Pennsylvania), Jacob R. Gardner (University of Pennsylvania)
OptimizationDrug DiscoveryTabular
🎯 What it does: A Bayesian optimization framework for black-box multi-objective coverage optimization, MOCOBO, is proposed, which can find K solutions that cover all objectives in high-dimensional structured search spaces.
CovMatch: Cross-Covariance Guided Multimodal Dataset Distillation with Trainable Text Encoder
Yongmin Lee (KAIST), Hye Won Chung (KAIST)
RetrievalKnowledge DistillationVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This study investigates how to perform multimodal dataset distillation in large-scale vision-language models to generate a small number of high-quality image-text pairs.
CoVoMix2: Advancing Zero-Shot Dialogue Generation with Fully Non-Autoregressive Flow Matching
leying zhang, sheng zhao
GenerationTransformerFlow-based ModelAudio
🎯 What it does: Developed CoVoMix2, a fully non-autoregressive zero-shot multi-speaker dialogue generation framework that directly predicts mixed Mel spectrograms from multi-stream text, supporting overlapping speech and temporal control.
CPathAgent: An Agent-based Foundation Model for Interpretable High-Resolution Pathology Image Analysis Mimicking Pathologists' Diagnostic Logic
Yuxuan Sun (Zhejiang University), Lin Yang (Westlake University)
ClassificationExplainability and InterpretabilityLarge Language ModelReinforcement LearningVision Language ModelImageBiomedical Data
🎯 What it does: This paper presents CPathAgent, an agent-based framework for pathological image analysis that mimics the diagnostic process of pathologists, including low-power scanning, locating areas of interest, gradually magnifying views, and performing multi-scale reasoning.
CPO: Condition Preference Optimization for Controllable Image Generation
Zonglin Lyu (University of Central Florida), Chen Chen (University of Central Florida)
SegmentationGenerationPose EstimationDiffusion modelContrastive LearningImage
🎯 What it does: A Conditional Preference Optimization (CPO) method is proposed, which enhances controllability in text-to-image generation by learning preferences at the control condition level.
CPPO: Accelerating the Training of Group Relative Policy Optimization-Based Reasoning Models
ZhiHang Lin, Rongrong Ji (Xiamen University)
OptimizationComputational EfficiencyReinforcement LearningAgentic AITabular
🎯 What it does: This study investigates the acceleration of training inference models based on GRPO through pruning high-advantage completions, proposing CPPO and a dynamic completion allocation strategy.
CQ-DINO: Mitigating Gradient Dilution via Category Queries for Vast Vocabulary Object Detection
Zhichao Sun (Wuhan University), Yongchao Xu (Wuhan University)
Object DetectionTransformerContrastive LearningImage
🎯 What it does: The CQ-DINO framework is proposed to address the issue of positive and negative gradient dilution in large vocabulary object detection through learnable category queries and image-guided query selection.
CREA: A Collaborative Multi-Agent Framework for Creative Image Editing and Generation
Kavana Venkatesh (Virginia Tech), Pinar Yanardag (Virginia Tech)
GenerationLarge Language ModelAgentic AIDiffusion modelImageVideoChain-of-Thought
🎯 What it does: The CREA multi-agent framework is proposed, specifically designed for creative image editing and generation, capable of achieving diverse and artistic image transformations and creations with minimal user intervention.
Creativity or Brute Force? Using Brainteasers as a Window into the Problem-Solving Abilities of Large Language Models
Simeng Han (Yale University), R. Thomas McCoy (Yale University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Designed and evaluated the BRAINGLE BRAINTEASER benchmark, which systematically measures the reasoning process, creativity, tendency for brute-force search, step decomposition, and prompt utilization of LLMs in long narrative brainteasers.
Credal Prediction based on Relative Likelihood
Timo Löhr (LMU Munich), Eyke Hüllermeier (LMU Munich)
Anomaly DetectionNeural Architecture SearchConvolutional Neural NetworkImageMagnetic Resonance Imaging
🎯 What it does: A trustworthy prediction method based on relative likelihood thresholds is proposed, and an approximately adjustable trustworthy set is integrated through neural networks.
CReFT-CAD: Boosting Orthographic Projection Reasoning for CAD via Reinforcement Fine-Tuning
Ke Niu (Fudan University), Xiangyang Xue (Fudan University)
TransformerReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: A two-stage fine-tuning framework called CReFT-CAD is proposed, which combines curriculum-driven reinforcement learning with supervised post-tuning to enhance CAD projection reasoning capabilities, and a large-scale projection reasoning benchmark called TriView2CAD is released.
Critical Batch Size Revisited: A Simple Empirical Approach to Large-Batch Language Model Training
William Merrill (Allen Institute for Artificial Intelligence), Hannaneh Hajishirzi
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A branch training method for directly measuring the critical batch size (CBS) threshold is proposed, and this method is used to study the evolution of CBS during training. Based on the changes in CBS, a batch size warm-up strategy is designed to accelerate large batch training without compromising performance.
CroPe: Cross-Modal Semantic Compensation Adaptation for All Adverse Scene Understanding
Qin Xu (Anhui University), Bo Jiang (Anhui University)
SegmentationDomain AdaptationAutonomous DrivingTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: A cross-modal semantic compensation adaptation method called CroPe is proposed, which enhances the performance of semantic segmentation tasks in unsafe domains under extreme weather conditions using visual + text.
Cross City Traffic Flow Generation via Retrieval Augmented Diffusion Model
Yudong Li (Beihang University), Qian Huang (Huawei Technologies Co., Ltd)
GenerationDomain AdaptationTransformerDiffusion modelTime SeriesRetrieval-Augmented Generation
🎯 What it does: Proposed a cross-city zero-shot traffic flow generation model CRAFT
Cross-Domain Graph Data Scaling: A Showcase with Diffusion Models
Wenzhuo Tang (Michigan State University), Jiliang Tang (Michigan State University)
ClassificationData SynthesisGraph Neural NetworkDiffusion modelGraphBiomedical Data
🎯 What it does: A universal graph structure enhancer called UniAug is proposed, aimed at improving the performance of downstream tasks by expanding the scale of graph data.
Cross-fluctuation phase transitions reveal sampling dynamics in diffusion models
Sai Niranjan Ramachandran (Technical University of Munich), Suvrit Sra (Technical University of Munich)
GenerationData SynthesisOptimizationDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper analyzes the sampling dynamics of the score diffusion model using cross-fluctuation indicators, discovering and locating discrete phase transitions that occur at different time points, thereby achieving fine diagnosis and intervention of the sampling paths.
Cross-modal Associations in Vision and Language Models: Revisiting the Bouba-Kiki Effect
Tom Kouwenhoven (Leiden University), Tessa Verhoef (Leiden University)
TransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Evaluate whether CLIP (ResNet and ViT) exhibits human-like bouba-kiki cross-modal associations in visual-language tasks.
Cross-Modal Representational Knowledge Distillation for Enhanced Spike-informed LFP Modeling
Eray Erturk (University of Southern California), Maryam M. Shanechi (University of Southern California)
Knowledge DistillationRepresentation LearningTransformerAuto EncoderMultimodalityBiomedical Data
🎯 What it does: This paper proposes a cross-modal knowledge distillation framework that transfers the representational knowledge from a multi-session, cross-animal multi-pulse Transformer pre-trained model (teacher) to a Transformer that only uses local field potentials (LFP) (student), and trains it under both completely unsupervised and partially supervised settings.
CrossAD: Time Series Anomaly Detection with Cross-scale Associations and Cross-window Modeling
Beibu Li (East China Normal University), Chenjuan Guo (East China Normal University)
Anomaly DetectionTransformerTime Series
🎯 What it does: Proposes the CrossAD framework, which utilizes cross-scale reconstruction and cross-window modeling for time series anomaly detection.
CrossSpectra: Exploiting Cross-Layer Smoothness for Parameter-Efficient Fine-Tuning
Yifei Zhang (Northwestern Polytechnical University), Han Yu (Nanyang Technological University)
TransformerSupervised Fine-TuningImageText
🎯 What it does: A method is proposed that utilizes the cross-layer smoothness of Transformer and achieves parameter-efficient fine-tuning through 3D frequency domain sparse representation.
CRRL: Learning Channel-invariant Neural Representations for High-performance Cross-day Decoding
Xianhan Tan (Zhejiang University), Yueming Wang (Zhejiang University)
Representation LearningGenerative Adversarial NetworkBiomedical Data
🎯 What it does: Proposes the CRRL framework to achieve channel-level invariant representation for cross-day BCI decoding.
Crucible: Quantifying the Potential of Control Algorithms through LLM Agents
Lianchen Jia (Tsinghua University), Lifeng Sun (Tsinghua University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTabularTime Series
🎯 What it does: The Crucible framework is proposed, which systematically evaluates and quantifies the tuning potential of control algorithms through LLM-driven multi-layer expert simulation and formal tunability metrics.
CryptoMoE: Privacy-Preserving and Scalable Mixture of Experts Inference via Balanced Expert Routing
Yifan Zhou (Peking University), Meng Li (Peking University)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes the CryptoMoE framework, which implements private inference for sparse gated Mixture of Experts (MoE) large language models;
CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding
Yuchen Zhou (Shanghai Artificial Intelligence Laboratory), Chao Gou (Shanghai Artificial Intelligence Laboratory)
ClassificationRecognitionAnomaly DetectionConvolutional Neural NetworkSupervised Fine-TuningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: A cross-scale spatiotemporal structure-aware EEG foundational model, CSBrain, is proposed to achieve multi-task EEG decoding.
CSGO: Content-Style Composition in Text-to-Image Generation
Peng Xing (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: This paper proposes a scalable automated content-style-stylization image triplet construction pipeline and builds a large-scale dataset IMAGStyle with 210K samples based on this pipeline. An end-to-end CSGO framework is then trained on this dataset, achieving image-driven, text-driven, and text-editing-driven style transfer.
CSPCL: Category Semantic Prior Contrastive Learning for Deformable DETR-Based Prohibited Item Detectors
Mingyuan Li (Northeastern University), Dongyue Chen (Northeastern University)
Object DetectionTransformerContrastive LearningImage
🎯 What it does: This study investigates the detection of contraband based on X-ray images and proposes a Category Semantic Prior Contrastive Learning mechanism (CSPCL), which enhances the model's robustness against interference from overlapping features by aligning classifier weights (category prototypes) with content queries in the decoder.
CTRL-ALT-DECEIT Sabotage Evaluations for Automated AI R&D
Francis Rhys Ward (LawZero), Lauren Robson (Imperial College London)
Anomaly DetectionOptimizationAI Code AssistantTransformerLarge Language ModelAgentic AIImageTextTabularAudio
🎯 What it does: This paper expands on MLE-Bench by adding 20 code sabotage tasks to evaluate whether cutting-edge large language models (Claude-3.7) intentionally sabotage the model or obscure performance when executing ML engineering tasks, and detects such behaviors through automatic monitoring.
Ctrl-DNA: Controllable Cell-Type-Specific Regulatory DNA Design via Constrained RL
Xingyu Chen (University of Toronto), BO WANG
OptimizationReinforcement LearningBiomedical Data
🎯 What it does: In response to the design of cell type-specific transcriptional regulatory elements (CREs), a method is proposed to automatically generate highly active and specific promoter and enhancer sequences through constrained reinforcement learning.
CTSketch: Compositional Tensor Sketching for Scalable Neurosymbolic Learning
Seewon Choi (University of Pennsylvania), Eric Wong (University of Pennsylvania)
ClassificationOptimizationComputational EfficiencyTabularBenchmark
🎯 What it does: A neural symbolic learning framework named CTSketch is proposed, which achieves differentiable inference for large-scale symbolic reasoning by decomposing symbolic programs into subprograms and performing low-rank tensor compression on the tensor summaries of each subprogram.
Cue3D: Quantifying the Role of Image Cues in Single-Image 3D Generation
Xiang Li (University of Illinois at Urbana-Champaign), James Matthew Rehg
GenerationImage
🎯 What it does: This paper proposes the Cue3D framework, which systematically quantifies the influence of various visual cues (such as shadows, textures, contours, etc.) in the 3D generation from a single image.
CURE: Co-Evolving Coders and Unit Testers via Reinforcement Learning
Yinjie Wang (University of Chicago), Mengdi Wang (Princeton University)
AI Code AssistantReinforcement LearningBenchmarkChain-of-Thought
🎯 What it does: A CURE framework is proposed, utilizing reinforcement learning to achieve the co-evolution of the encoder and unit test generator without requiring real code as supervision;
CURE: Concept Unlearning via Orthogonal Representation Editing in Diffusion Models
Shristi Das Biswas (Purdue University), Kaushik Roy (Purdue University)
GenerationData SynthesisExplainability and InterpretabilityTransformerDiffusion modelImage
🎯 What it does: A training-free concept forgetting framework called CURE is proposed, which directly edits the cross-attention weights in the weight space of pre-trained diffusion models in a closed-form manner, achieving rapid, interpretable, and highly specific erasure of target concepts.
Curious Causality-Seeking Agents Learn Meta Causal World
Zhiyu Zhao, Mengyue Yang (University of Bristol)
Robotic IntelligenceMeta LearningReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerReinforcement LearningAgentic AIAuto EncoderGraph
🎯 What it does: Construct a Meta-Causal Graph and propose a Curious Causality-Seeking Agent that actively intervenes to explore and learn variable causal structures in open environments.
Curl Descent : Non-Gradient Learning Dynamics with Sign-Diverse Plasticity
Hugo Ninou (Ecole Normale Superieure PSL), N Alex Cayco Gajic
OptimizationTabular
🎯 What it does: This paper systematically studies the naturally occurring non-gradient 'curl' terms in biologically plausible synaptic plasticity rules (such as excitatory-inhibitory networks or mixed Hebbian/anti-Hebbian rules) and proposes a 'curl descent' learning rule—introducing the curl term by reversing the signs of some synaptic updates. The authors perform analytical stability analysis in a two-layer linear student-teacher framework, utilize random matrix theory to plot phase diagrams under different network architectures and curl strengths, and validate through numerical simulations the phenomena of chaotic learning, instability, and accelerated convergence caused by the curl term.
Curly Flow Matching for Learning Non-gradient Field Dynamics
Katarina Petrović (University of Oxford), Alexander Tong (Université de Montréal)
Flow-based ModelTime SeriesBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A new method called Curly Flow Matching (CURLY-FM) is proposed for learning non-gradient field dynamics, capturing periodic behavior by solving the Schrödinger bridge problem with a non-zero drift reference process.
Curriculum Abductive Learning
Wen-Chao Hu (Nanjing University), Zhi-Hua Zhou (Nanjing University)
Convolutional Neural NetworkTransformerImageTabular
🎯 What it does: A course-based inductive reasoning (C-ABL) framework is proposed, which divides the knowledge base into sub-libraries according to logical structure and gradually introduces them, forming a training process that progresses from simple to complex.
Curriculum Design for Trajectory-Constrained Agent: Compressing Chain-of-Thought Tokens in LLMs
Georgios Tzannetos (Max Planck Institute for Software Systems), Adish Singla (Max Planck Institute for Software Systems)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: An adaptive constraint adjustment strategy based on curriculum learning is proposed, allowing reinforcement learning and large language models to gradually tighten constraints under strict trajectory constraints and ultimately meet hard constraints during deployment; validation is conducted on binary tree MDP, PuddleGrid, and mathematical reasoning tasks.
Curriculum Model Merging: Harmonizing Chemical LLMs for Enhanced Cross-Task Generalization
Baoyi He (Zhejiang University), Fei Wu (Zhejiang University)
Drug DiscoveryTransformerLarge Language ModelTextBenchmark
🎯 What it does: A curriculum learning-based chemical LLM fusion method is proposed—Curriculum Model Merging (CMM), which integrates multi-task chemical expert models by merging them progressively based on expert performance ranking;
CURV: Coherent Uncertainty-Aware Reasoning in Vision-Language Models for X-Ray Report Generation
Ziao Wang (Hong Kong Baptist University), William K. Cheung (Hong Kong Baptist University)
GenerationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelTextMultimodalityBiomedical DataElectronic Health Records
🎯 What it does: The CURV framework is proposed, utilizing a vision-language model to generate chest X-ray reports, incorporating structured uncertainty expression and an explicit reasoning process (Findings-Thinking-Impression) in the reports.
Curvature Tuning: Provable Training-free Model Steering From a Single Parameter
Leyang Hu (Brown University), Randall Balestriero (Brown University)
ClassificationAdversarial AttackConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: A Curvature Tuning (CT) method is proposed to achieve model Steering by injecting a single hyperparameter into the activation function, which can serve as untrained Steering (S-CT) or as a trainable parameter-efficient Fine-Tuning (T-CT).
CVGL: Causal Learning and Geometric Topology
Songsong Ouyang (Shenzhen University), Yingying Zhu (Shenzhen University)
RetrievalDomain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The CLGT framework is proposed, combining causal learning and geometric topology fusion to address the cross-view geographic localization problem between street scenes and aerial maps.
Cycle-Sync: Robust Global Camera Pose Estimation through Enhanced Cycle-Consistent Synchronization
Shaohan Li (University of Minnesota), Gilad Lerman (University of Minnesota)
Pose EstimationSimultaneous Localization and MappingImage
🎯 What it does: Proposes the Cycle-Sync method, constructing a global and robust camera pose estimation framework that can simultaneously estimate camera rotation and translation without the need for traditional bundle adjustment steps.
Cyclic Counterfactuals under Shift–Scale Interventions
Saptarshi Saha (Indian Statistical Institute), Utpal Garain (Indian Statistical Institute)
🎯 What it does: This paper proposes a method for counterfactual reasoning in structural causal models (SCM) with cyclic dependencies, specifically targeting shift-scale soft interventions.
CyIN: Cyclic Informative Latent Space for Bridging Complete and Incomplete Multimodal Learning
Ronghao Lin (Sun Yat-Sen University), Haifeng Hu (Sun Yat-Sen University)
ClassificationData SynthesisRepresentation LearningTransformerAuto EncoderMultimodality
🎯 What it does: A cyclic information bottleneck space (CyIN) has been constructed, capable of handling both complete and incomplete multimodal learning, and completing missing modalities through cross-modal cyclic translation.
CymbaDiff: Structured Spatial Diffusion for Sketch-based 3D Semantic Urban Scene Generation
Li Liang (University of Western Australia), Ajmal Saeed Mian
SegmentationGenerationData SynthesisDiffusion modelImagePoint CloudBenchmark
🎯 What it does: This study investigates the use of sketches and pseudo-labeled satellite images to generate 3D outdoor semantic scenes, proposing the CymbaDiff diffusion model and the SketchSem3D dataset.
Cypher-RI: Reinforcement Learning for Integrating Schema Selection into Cypher Generation
Hanchen Su (Zhejiang University), Yang Yang (Zhejiang University)
TransformerLarge Language ModelReinforcement LearningTextGraph
🎯 What it does: Developed a Cypher-RI framework based on reinforcement learning, integrating schema selection into the process of generating text to Cypher, and training the LLM through unsupervised reasoning trajectories.
D-VST: Diffusion Transformer for Pathology-Correct Tone-Controllable Cross-Dye Virtual Staining of Whole Slide Images
shurong yang, Liansheng Wang (Xiamen University)
Image TranslationGenerationTransformerDiffusion modelImage
🎯 What it does: A cross-staining virtual staining method D-VST based on diffusion Transformer is proposed, which achieves controllable staining by adjusting the color tone while maintaining the accuracy of pathological structures.
D$^2$GS: Dense Depth Regularization for LiDAR-free Urban Scene Reconstruction
Kejing Xia (Wuhan University), Youjian Zhang (Bosch)
RestorationDepth EstimationAutonomous DrivingDiffusion modelGaussian SplattingPoint Cloud
🎯 What it does: This paper proposes a LiDAR-free dynamic urban scene reconstruction framework called D2GS, which utilizes camera data to achieve high-quality 3D Gaussian Splatting reconstruction through multi-view depth estimation, progressive cropping, and depth enhancement driven by a diffusion model.
d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning
Siyan Zhao (University of California Los Angeles), Aditya Grover (University of California Los Angeles)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelText
🎯 What it does: The d1 framework is proposed, combining supervised fine-tuning and reinforcement learning to enhance the reasoning ability of masked diffusion large language models.
D2SA: Dual-Stage Distribution and Slice Adaptation for Efficient Test-Time Adaptation in MRI Reconstruction
Lipei Zhang (University of Cambridge), Angelica I Aviles-Rivero (Tsinghua University)
RestorationDomain AdaptationConvolutional Neural NetworkNeural Radiance FieldImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a dual-stage adaptive framework for testing (D2SA), which first learns patient-level distribution shifts using implicit neural representations (INR), and then refines single slices through a learnable anisotropic diffusion (AD) module to enhance the quality and efficiency of MRI reconstruction under different distributions.
DAA: Amplifying Unknown Discrepancy for Test-Time Discovery
Tianle Liu (Suzhou University of Science and Technology), Liang Wang (Chinese Academy of Sciences)
ClassificationRecognitionDomain AdaptationTransformerContrastive LearningImage
🎯 What it does: A trainable Discrepancy-Amplifying Adapter (DAA) and Short-Term Memory Renewal (STMR) mechanism are proposed to achieve real-time adaptation to unknown categories during Test-Time Discovery while maintaining performance on known categories.
DAAC: Discrepancy-Aware Adaptive Contrastive Learning for Medical Time series
Yifan WANG, Chenzhong Li (Chinese University of Hong Kong)
Anomaly DetectionRepresentation LearningRecurrent Neural NetworkAuto EncoderGenerative Adversarial NetworkContrastive LearningTime SeriesBiomedical DataAlzheimer's Disease
🎯 What it does: The DAAC framework is proposed, which enhances the representation learning and generalization ability of medical time series diagnostic models through a difference estimator generated from external normal samples and multi-view adaptive contrastive learning.
DAIL: Beyond Task Ambiguity for Language-Conditioned Reinforcement Learning
Runpeng Xie (Chinese Academy of Sciences), Bo XU
Robotic IntelligenceReinforcement LearningContrastive LearningText
🎯 What it does: The DAIL framework is proposed, which reduces task ambiguity caused by language instructions through distributed value estimation and trajectory semantic alignment, achieving more accurate task identification and execution.
DAMamba: Vision State Space Model with Dynamic Adaptive Scan
Tanzhe Li (Xiamen University), Rongrong Ji (Xiamen University)
ClassificationObject DetectionSegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: A dynamic adaptive scanning-based visual state space model called DAMamba is proposed to enhance the performance of image classification, object detection, instance segmentation, and semantic segmentation.
DAPO : Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage-Based Policy Optimization
Jiacai Liu (Fudan University), Yang Liu (Skywork AI)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a new offline step-level reinforcement learning framework—Direct Advantage-Based Policy Optimization (DAPO)—to enhance the multi-step reasoning capabilities of large language models (LLMs).
DAPO: An Open-Source LLM Reinforcement Learning System at Scale
Qiying Yu (ByteDance Seed), Mingxuan Wang (ByteDance Seed)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A large-scale LLM reinforcement learning system DAPO is proposed and open-sourced, using the Qwen2.5-32B base model, achieving a score of 50 at AIME 2024.
DartQuant: Efficient Rotational Distribution Calibration for LLM Quantization
Yuantian Shao (Nanjing University of Science and Technology), Jian Cheng (Chinese Academy of Sciences)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: DartsQuant is proposed, a distribution calibration-based rotation matrix quantization method that efficiently completes activation quantization of large-scale language models without the need for end-to-end fine-tuning.
Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning
Xiao Han (Zhejiang University of Technology), Xiangyu Zhao (City University of Hong Kong)
Domain AdaptationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A continuous low-rank fine-tuning framework named DEAL is proposed for efficient and continuous adaptation on large language models.
Data Fusion for Partial Identification of Causal Effects
Quinn Lanners (Duke University), Harsh Parikh (Yale University)
TabularTime Series
🎯 What it does: A partial identification framework is proposed for situations where both the experimental group and the observational group simultaneously violate the assumptions of no unobserved confounding and research exchangeability, providing interpretable sensitivity parameters γ and ρ to quantify the violations of assumptions, and subsequently deriving upper and lower bounds for treatment effects.
Data Mixing Can Induce Phase Transitions in Knowledge Acquisition
Xinran Gu (Tsinghua University), Jingzhao Zhang (Tsinghua University)
Large Language ModelText
🎯 What it does: The study investigates the impact of knowledge-dense data in mixed training on the knowledge acquisition of large language models, revealing a phase transition phenomenon related to model size and mixing ratio.
Data Mixture Optimization: A Multi-fidelity Multi-scale Bayesian Framework
Thomson Yen (Columbia Business School), Hongseok Namkoong (Columbia Business School)
OptimizationLarge Language ModelText
🎯 What it does: This paper proposes a multi-fidelity multi-scale Bayesian optimization framework for automatically optimizing the training data mixing ratio during the pre-training process of large language models.
Data Selection Matters: Towards Robust Instruction Tuning of Large Multimodal Models
Xu Yang (City University of Hong Kong), Ying Wei (Zhejiang University)
Data-Centric LearningTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: A gradient-free robustness-oriented data selection framework (ARDS) has been constructed for visual instruction tuning, selecting a training subset that can enhance the robustness of multimodal models against positional bias and spurious correlations.
Data-Adaptive Exposure Thresholds under Network Interference
Vydhourie Thiyageswaran (University of Washington), Jennifer Rogers Brennan
Graph
🎯 What it does: This paper proposes an adaptive threshold Horvitz-Thompson estimator to estimate the average treatment effect under network interference.
Data-Dependent Regret Bounds for Constrained MABs
Gianmarco Genalti (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
OptimizationReinforcement Learning
🎯 What it does: In the constrained adversarial multi-armed bandit problem, a safety OMD algorithm based on logarithmic barriers (COLB and SOLB) is proposed, achieving a data-dependent upper bound on small loss regret;
Data-Free Model Extraction for Black-box Recommender Systems via Graph Convolutions
Zeyu Wang (Zhejiang University of Technology), Xin Zheng (Griffith University)
Recommendation SystemAdversarial AttackGraph Neural NetworkGenerative Adversarial NetworkTabular
🎯 What it does: This study investigates model extraction attacks on recommendation systems in a black-box, no-data environment and proposes a data-free model extraction method based on graph convolution (DBGRME).
DataRater: Meta-Learned Dataset Curation
Dan A. Calian (Google DeepMind), David Silver (Google DeepMind)
Computational EfficiencyData-Centric LearningMeta LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes a meta-learning framework called DataRater, which is used to automatically estimate the value of training data points and filter or weight them, thereby improving the training efficiency of large-scale foundational models.