ICLR 2026 Papers — Page 9
International Conference on Learning Representations · 5356 papers
Completing Missing Annotation: Multi-Agent Debate for Accurate and Scalable Relevant Assessment for IR Benchmarks
Minjeong Ban (Korea Advanced Institute of Science and Technology), Hwanjun Song (Korea Advanced Institute of Science and Technology)
RetrievalTransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Design and implement a multi-agent debate framework called DREAM for automated relevance assessment in information retrieval, and construct the BRIDGE benchmark based on this framework to fill missing text fragments.
Complexity Analysis of Normalizing Constant Estimation: from Jarzynski Equality to Annealed Importance Sampling and beyond
Wei Guo (Georgia Institute of Technology), Yongxin Chen (Georgia Institute of Technology)
Computational EfficiencyDiffusion modelScore-based ModelPhysics Related
🎯 What it does: Studied the complexity of estimating normalization constants using Jarzynski equality, annealed importance sampling (AIS), and reverse diffusion sampling (RDS), provided non-asymptotic upper bounds, and proposed a new RDS algorithm.
Complexity- and Statistics-Guided Anomaly Detection in Time Series Foundation Models
Jongwon Kim (POSTECH), JAEUNG TAE
Anomaly DetectionTransformerMixture of ExpertsAuto EncoderTime Series
🎯 What it does: Propose an anomaly detection method based on time series foundational models to address overgeneralization and overnormalization issues, constructing complexity-adaptive ensemble (CAE) and instance-statistical enhancement model (MOMENT-Stat)
CompMarkGS: Robust Watermarking for Compressed 3D Gaussian Splatting
Sumin In (Korea University), Sangpil Kim (Korea University)
CompressionSafty and PrivacyGaussian Splatting
🎯 What it does: Proposes a compression fault-tolerant watermarking method called CompMarkGS for 3D Gaussian Splatting models, which can maintain watermark integrity and rendering quality even after the model is quantized and compressed.
CompoDistill: Attention Distillation for Compositional Reasoning in Multimodal LLMs
Jiwan Kim (KAIST), Chanyoung Park (KAIST)
Knowledge DistillationTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: This paper addresses the problem of knowledge distillation failing to effectively enhance visual perception capabilities in multimodal large language models (MLLMs), proposing the CompoDistill framework. Through two modules, visual attention alignment (VAT) and teacher adapter grasping (TAF), it significantly improves the student model's performance in compositional reasoning tasks while maintaining visual question answering (VQA) performance.
Composable Sparse Subnetworks via Maximum-Entropy Principle
Francesco Caso (Sapienza University of Rome), Fabrizio Silvestri (Sapienza University of Rome)
ClassificationConvolutional Neural NetworkMixture of ExpertsImageTextTabular
🎯 What it does: Train sparse subnetworks to make accurate predictions only on specified class subsets while remaining uncertain on other classes; subsequently reconstruct the overall model by combining these subnetworks through weight summation or logit averaging.
Compose and Fuse: Revisiting the Foundational Bottlenecks in Multimodal Reasoning
Yucheng Wang (ETH Zurich), Mrinmaya Sachan (ETH Zurich)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Built a multi-modal evaluation framework based on logical reasoning, defined six modal interaction patterns, and evaluated the reasoning ability of multi-modal large language models on synthetic data.
Compose Your Policies! Improving Diffusion-based or Flow-based Robot Policies via Test-time Distribution-level Composition
Jiahang Cao (University of Hong Kong), Andrew Luo (University of Hong Kong)
Robotic IntelligenceDiffusion modelScore-based ModelFlow-based Model
🎯 What it does: Propose a training-free strategy combination framework called GPC, which enhances control performance by performing convex combination of distributional scores from multiple pre-trained diffusion/flow-based robot policies during testing.
Composer: A Search Framework for Hybrid Neural Architecture Design
Bilge Acun (Meta), Carole-Jean Wu (University of Texas at Austin)
Computational EfficiencyNeural Architecture SearchTransformerText
🎯 What it does: Proposed the Composer framework for automatically searching and constructing high-quality hybrid neural network architectures (Hybrid Neural Architecture Search, HNAS) to surpass traditional Transformers in large-scale language model training;
Composite Optimization with Error Feedback: the Dual Averaging Approach
Yuan Gao (CISPA Helmholtz Center for Information Security), Sebastian U Stich (CISPA Helmholtz Center for Information Security)
OptimizationFederated LearningImage
🎯 What it does: Proposes a distributed compressed optimization algorithm combining error feedback (EF) with dual averaging, achieving communication efficiency improvements in convex composite optimization (including smooth loss + non-smooth regularization or constraints), along with theoretical convergence analysis.
Composition of Memory Experts for Diffusion World Models
Sebastian Stapf (University of Bern), Paolo Favaro (University of Bern)
Autonomous DrivingSupervised Fine-TuningMixture of ExpertsDiffusion modelContrastive LearningVideo
🎯 What it does: Propose integrating short-term memory, long-term memory, and spatial long-term memory experts into a diffusion-based world model to achieve efficient fusion of information across different time scales
Composition of Pretrained Diffusion Models: A Logic-Based Calculus
Peter Blohm (Aalto University), Vikas K Garg (Aalto University)
GenerationDrug DiscoveryMixture of ExpertsDiffusion modelScore-based ModelImageBiomedical Data
🎯 What it does: This paper proposes a fuzzy logic-based Dombi operator that can online combine pre-trained diffusion models using logical operations (AND, OR, NOT), addressing the shortcomings of traditional PoE/MoE combinations in terms of mode coverage, sampling bias, and violation of logical laws.
Composition-Grounded Data Synthesis for Visual Reasoning
Xinyi Gu (Massachusetts Institute Of Technology), Zexue He (Massachusetts Institute Of Technology)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Propose the COGS framework, which generates a large number of synthetic question-answer pairs by decomposing and recombining a small number of seed question-answer pairs through factorization, thereby enhancing the ability of multimodal large language models in chart and web reasoning tasks.
Compositional amortized inference for large-scale hierarchical Bayesian models
Jonas Arruda (University of Bonn), Stefan T. Radev (Rensselaer Polytechnic Institute)
Computational EfficiencyRepresentation LearningDiffusion modelScore-based ModelImageTime SeriesBiomedical DataStochastic Differential Equation
🎯 What it does: Proposes a compositional approximation inference framework for large-scale hierarchical Bayesian models, achieving efficient simulation-free inference through error-damping grouped score matching.
Compositional Diffusion with Guided search for Long-Horizon Planning
Utkarsh Aashu Mishra, Danfei Xu (Georgia Institute of Technology)
GenerationVision Language ModelDiffusion modelImageVideoTextMultimodality
🎯 What it does: Proposed Compositional Diffusion with Guided Search (CDGS), embedding search into the diffusion denoising process, addressing the mode averaging problem caused by multi-modal combinations through iterative resampling and likelihood pruning, generating globally consistent sequences for tasks such as long-term planning, panoramas, and long videos.
Compositional Generalization from Learned Skills via CoT Training: A Theoretical and Structural Analysis for Reasoning
Xinhao Yao (Renmin University of China), Yong Liu (Renmin University of China)
TransformerTextChain-of-Thought
🎯 What it does: This paper explains how Chain-of-Thought (CoT) training enhances the compositional generalization ability of large language models in both in-distribution (ID) and out-of-distribution (OOD) scenarios through theoretical derivation and structural analysis, and verifies its effectiveness on controllable synthetic data and real mathematical problems.
Compositional Generalization through Gradient Search in Nonparametric Latent Space
Haruki Shirakami (École Polytechnique Fédérale de Lausanne), James Henderson (Idiap Research Institute)
Representation LearningTransformer
🎯 What it does: Propose a novel transformer architecture called Abduction Transformer, which achieves combinatorial generalization capability for various discrete abstract reasoning tasks by leveraging non-parametric mixture distribution latent spaces, information theory regularization, and gradient search during testing.
Compositional Visual Planning via Inference-Time Diffusion Scaling
Yixin Zhang (Georgia Institute of Technology), Danfei Xu (Georgia Institute of Technology)
Robotic IntelligenceDiffusion modelImageVideo
🎯 What it does: By using diffusion models during inference, long-term visual planning is decomposed into overlapping short segments, with synchronous and asynchronous message passing performed on Tweedie estimates to achieve globally consistent visual trajectories. Subsequently, inverse dynamics models map image sequences to robot actions.
Compositional-ARC: Assessing Systematic Generalization in Abstract Spatial Reasoning
Philipp Mondorf (LMU Munich), Barbara Plank (LMU Munich)
Meta LearningTransformerImageBenchmark
🎯 What it does: Proposed a new benchmark dataset, Compositional-ARC, to evaluate models' systematic generalization ability in abstract spatial reasoning; trained a small Transformer encoder-decoder using meta-learning (MLC), achieving systematic generalization on unknown composite transformations;
Computational Bottlenecks for Denoising Diffusions
Viet Vu (Stanford University), Andrea Montanari (Stanford University)
GenerationComputational EfficiencyDiffusion modelScore-based ModelStochastic Differential Equation
🎯 What it does: Investigated the computational bottleneck in diffusion sampling, proving that under distributions with an information-computation gap, even if the drift function is approximately optimal under the score matching objective, sampling failure still occurs, and provided a strict reduction from estimation to diffusion sampling.
Compute-Optimal Quantization-Aware Training
Aleksandr Dremov (Apple Inc), Awni Hannun (Apple Inc)
Computational EfficiencyTransformerText
🎯 What it does: The study investigates how to allocate computational resources between full-precision pre-training and quantization-aware training (QAT), and explores the pattern of the optimal QAT ratio as the total computational resources vary through large-scale experiments.
Computer Agent Arena: Toward Human-Centric Evaluation and Analysis of Computer-Use Agents
Bowen Wang (The University of Hong Kong), Tao Yu (The University of Hong Kong)
Reinforcement Learning from Human FeedbackLarge Language ModelAgentic AITextMultimodalityBenchmark
🎯 What it does: Developed an open-source platform named COMPUTER AGENT ARENA for anonymous side-by-side, human-centric evaluation of Computer-Using Agents (CUA) in real-world cloud-diverse environments;
ComputerRL: Scaling End-to-End Online Reinforcement Learning for Computer Use Agents
Hanyu Lai (Tsinghua University), Jie Tang (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: Built a complete computer usage intelligent agent framework named COMPUTERRL, capable of performing complex tasks through API and GUI dual-mode interaction in a desktop environment, achieving end-to-end online reinforcement learning.
Computing Equilibrium beyond Unilateral Deviation
Mingyang Liu (Massachusetts Institute of Technology), Asuman E. Ozdaglar (Massachusetts Institute of Technology)
OptimizationReinforcement LearningGraph
🎯 What it does: This paper proposes and studies 'Minimum Average Strong Equilibrium' (MASE) — a novel concept of balance that measures the feasibility of multi-party deviations by minimizing the average improvement gain of any coalition, and provides its existence, complexity, and algorithms;
CoNavBench: Collaborative Long-Horizon Vision-Language Navigation Benchmark
Tianhang Wang (Tongji University), Guang Chen (Tongji University)
OptimizationRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelVision-Language-Action ModelTextMultimodalityGraphBenchmark
🎯 What it does: Designed and implemented a collaborative long-term visual-language navigation benchmark CoNavBench and the NavCraft task generation and scheduling platform, supporting automatic construction, planning, and evaluation of multi-robot collaborative tasks.
Concept-Aware Privacy Mechanisms for Defending Embedding Inversion Attacks
Yu-Che Tsai (National Taiwan University), Shou-De Lin (National Taiwan University)
Safty and PrivacyRepresentation LearningTransformerTextBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: This paper proposes a framework named SPARSE, which can identify and weight sensitive dimensions in text embeddings according to user-specified privacy concepts, then inject elliptical noise using the Mahalanobis distance mechanism to resist embedding inversion attacks while maintaining semantic quality.
Concept-based Adversarial Attack: a Probabilistic Perspective
Andi Zhang (University of Warwick), Samuel Kaski (University of Manchester)
Adversarial AttackLarge Language ModelSupervised Fine-TuningDiffusion modelImage
🎯 What it does: Proposed and implemented a concept-level adversarial attack framework that generates diverse adversarial samples from concept distributions rather than single images.
Concept-TRAK: Understanding how diffusion models learn concepts through concept attribution
Yong-Hyun Park (University of Pennsylvania), Yuki Mitsufuji
GenerationExplainability and InterpretabilityDiffusion modelScore-based ModelImageBenchmark
🎯 What it does: To address the concept-level attribution problem in diffusion models, the Concept-TRAK method is proposed, which evaluates the impact of training samples on specific concepts using a reward-optimized loss function and influence functions.
Concepts' Information Bottleneck Models
Karim Galliamov (University of Amsterdam), Adín Ramírez Rivera (University of Oslo)
Explainability and InterpretabilityRepresentation LearningImage
🎯 What it does: The study introduces information bottleneck regularization into conceptual bottleneck models, compressing input-concept mutual information while preserving concept-label mutual information to enhance interpretability and performance.
CONCUR: A Framework for Continual Constrained and Unconstrained Routing
Peter Baile Chen (Massachusetts Institute Of Technology), Jacob Andreas (Massachusetts Institute Of Technology)
OptimizationRepresentation LearningLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the CONCUR framework, which trains predictors separately for each computational strategy, estimates accuracy and inference cost using multiple representations of tasks and strategies, and performs task routing through optimization solving under both budget-constrained and unconstrained scenarios;
Condition Errors Refinement in Autoregressive Image Generation with Diffusion Loss
Yucheng Zhou (University of Macau), Jianbing Shen (University of Macau)
GenerationTransformerDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes using diffusion loss in autoregressive image generation and reduces conditional errors by block-wise denoising and conditional refinement, thereby improving generation quality;
Condition Matters in Full-head 3D GANs
Heyuan Li (Chinese University of Hong Kong), Xiaoguang Han (Chinese University of Hong Kong)
GenerationData SynthesisVision Language ModelDiffusion modelGenerative Adversarial NetworkImageMultimodality
🎯 What it does: Proposed a view-invariant semantic conditional full-head 3D GAN (BalanceHead), which uses CLIP features from the front view as conditions to achieve high-fidelity, rich, and diverse full-head 3D generation, and constructed a large-scale 360° synthetic dataset.
Conditional Advantage Estimation for Reinforcement Learning in Large Reasoning Models
Guanxu Chen (Shanghai Jiao Tong University), Jing Shao (Shanghai Artificial Intelligence Laboratory)
Large Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a novel reward estimation framework called CANON, which leverages conditional grouping to compare inter-group and intra-group advantages in reinforcement learning, thereby amplifying the impact of target metrics (e.g., entropy, length) on inference performance without presupposing the direction of the metric.
Conditional Independent Component Analysis for Estimating Causal Structure with Latent Variables
Yewei Xia (Fudan University), Shuigeng Zhou (Fudan University)
🎯 What it does: Proposed the principle of Conditional Independent Component Analysis (CICA) for identifying complete causal structures in linear non-Gaussian acyclic models with latent variables, and constructed the CICA-LiNGAM estimation algorithm based on this.
Conditionally Whitened Generative Models for Probabilistic Time Series Forecasting
Yanfeng Yang (Graduate University of Advanced Studies), Kenji Fukumizu (Institute of Statistical Mathematics)
GenerationDiffusion modelFlow-based ModelTime Series
🎯 What it does: This paper proposes a Conditional Whitening Generative Model (CW-Gen), which improves the quality of probabilistic forecasting for multivariate time series by jointly estimating conditional means and sliding window covariance and performing conditional whitening.
Conditioned Initialization for Attention
Hemanth Saratchandran (Adelaide University), Simon Lucey (Adelaide University)
OptimizationComputational EfficiencyTransformerImageVideoText
🎯 What it does: Proposes a conditional initialization method for the Transformer attention layer, aiming to improve the spectral condition number of attention weights, thereby enhancing optimization stability and performance.
ConfHit: Conformal Generative Design with Oracle-Free Guarantees
Siddhartha Laghuvarapu (University Of Illinois Urbana Champaign), Jimeng Sun (University Of Illinois Urbana Champaign)
GenerationDrug DiscoveryTransformerFlow-based ModelAuto EncoderBiomedical Data
🎯 What it does: Proposes the CONFHIT framework, which provides finite-sample reliability guarantees for sample sets generated by conditional generative models without relying on experimental oracles, and achieves compact candidate set design.
Confident and Adaptive Generative Speech Recognition via Risk Control
Amit Damri (Tel Aviv University), Bracha Laufer-Goldshtein (Tel Aviv University)
RecognitionLarge Language ModelAudio
🎯 What it does: Propose an adaptive generative speech recognition error correction framework based on risk control, dynamically determining the required number of N-best options for each input to achieve efficient and reliable post-processing.
Confident Block Diagonal Structure-Aware Invariable Graph Completion for Incomplete Multi-view Clustering
Shuping Zhao (Guangdong University of Technology), Tingting Chai (Harbin Institute of Technology)
OptimizationRepresentation LearningMultimodalityBenchmark
🎯 What it does: Propose a method named CBDS-IMVC, which utilizes confidence block diagonal structure and invariant graph completion techniques to reconstruct missing multi-view data and achieve clustering.
Conformal Prediction for Long-Tailed Classification
Tiffany Ding (University of California), Joseph Salmon (University of Montpellier)
ClassificationImage
🎯 What it does: This study proposes a conformal prediction method for long-tail classification, aiming to balance the prediction set size and class-conditional coverage while maintaining marginal coverage.
Conformal Prediction with Corrupted Labels: Uncertain Imputation and Robust Re-weighting
Shai Feldman (Technion), Yaniv Romano (Technion)
Data-Centric LearningTabularElectronic Health Records
🎯 What it does: Proposed a robust calibration framework that achieves effective coverage even when labels are noisy or missing, incorporating Uncertain Interpolation (UI) and Privileged Conformal Prediction (PCP)/Weighted Conformal Prediction (WCP) that are robust to weight errors;
Conformal Robustness Control: A New Strategy for Robust Decision
Yang Hu (Shanghai Jiao Tong University), Haojie Ren (Shanghai Jiao Tong University)
OptimizationTabularTime SeriesFinance Related
🎯 What it does: Proposes a new Conformal Robustness Control (CRC) framework that directly minimizes the expected risk certificate during the construction of prediction sets, guided by robustness constraints rather than coverage constraints for robust decision-making.
Conformalized Decision Risk Assessment
Wenbin Zhou (Carnegie Mellon University), Shixiang Zhu (Carnegie Mellon University)
OptimizationExplainability and InterpretabilityTabular
🎯 What it does: Propose the CREDO framework, which can provide distribution-agnostic upper bounds on sub-optimality probabilities for any candidate decision;
Conformalized Hierarchical Calibration for Uncertainty-Aware Adaptive Hashing
Junyu Luo (Peking University), Ming Zhang (University of International Business and Economics)
RetrievalDomain AdaptationImage
🎯 What it does: Proposed an unsupervised domain adaptive hashing method utilizing hierarchical synthetic calibration to enhance cross-domain retrieval performance.
Conformalized Survival Counterfactuals Prediction for General Right-Censored Data
Sijie Ren (Fudan University), Xinwei Sun (Fudan University)
TabularBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a method for predicting lower prediction bounds (LPB) of survival time under different treatment regimens in the presence of general right-censored data.
Conjuring Semantic Similarity
Tian Yu Liu (University of California, Los Angeles), Stefano Soatto (University of California, Los Angeles)
GenerationData SynthesisRepresentation LearningDiffusion modelScore-based ModelTextBenchmarkStochastic Differential Equation
🎯 What it does: This paper proposes a method to measure the semantic similarity between text expressions by generating image distributions based on text-conditioned diffusion models;
ConRep4CO: Contrastive Representation Learning of Combinatorial Optimization Instances across Types
Ziao Guo (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
OptimizationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Built a unified contrastive learning framework ConRep4CO, which converts multi-type graph decision problems into SAT, using the corresponding SAT graph as positive samples and other SAT graphs as negative samples to pre-train cross-domain generalizable graph representations.
CONSIGN: Conformal Segmentation Informed by Spatial Groupings via Decomposition
Bruno Viti (University of Graz), Martin Holler (University of Graz)
SegmentationBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Propose the CONSIGN method, which leverages SVD principal components to capture spatial correlations and constructs a segmentation prediction set with statistical guarantees.
Consis-GCPO: Consistency-Preserving Group Causal Preference Optimization for Vision Customization
Qiaoqiao Jin, Jidong Jiang (The Hong Kong University Of Science And Technology)
GenerationData SynthesisReinforcement LearningDiffusion modelFlow-based ModelImageVideoTextMultimodality
🎯 What it does: Propose Consis-GCPO, which quantifies the temporal impact of text and visual conditions using discrete-time causal models and stepwise causal interventions, thereby achieving subjective identity preservation and text consistency during personalized image and video generation.
ConsisDrive: Identity-Preserving Driving World Models for Video Generation by Instance Mask
Zhuoran Yang (University of Science and Technology of China), Yanyong Zhang (University of Science and Technology of China)
GenerationData SynthesisAutonomous DrivingTransformerVision Language ModelDiffusion modelAuto EncoderVideoTextPoint Cloud
🎯 What it does: Proposes a driving scene video generation model called ConsisDrive, focusing on solving the problem of identity drift for the same instance across multiple frames.
Consistency-Driven Calibration and Matching for Few-Shot Class Incremental Learning
Qinzhe Wang (Central South University), Chang Xu (University of Sydney)
ClassificationMeta LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes the Consistency-driven Calibration and Matching (ConCM) framework to address knowledge conflicts in Few-Shot Class Incremental Learning (FSCIL). It eliminates prototype bias through memory-aware prototype calibration and achieves dual consistency between features and structure via dynamic structural matching.
Consistent Low-Rank Approximation
David Woodruff, Samson Zhou (Texas A&M University)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes and studies the problem of consistent low-rank approximation under row-stream input, presenting two algorithms that maintain (1+ε) approximation accuracy while achieving lower recurrence cost than traditional online algorithms.
Consistent Noisy Latent Rewards for Trajectory Preference Optimization in Diffusion Models
Xiaole Xian (MMLab), Xiangyu Yue (MMLab)
GenerationOptimizationDiffusion modelScore-based ModelContrastive LearningImageVideoTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: In text-to-image/video generation with diffusion models, this paper proposes a reward model based on noise compatibility (SLRM) and a cross-time trajectory preference optimization framework (TAPO) to achieve more stable preference alignment.
Consistent Text-to-Image Generation via Scene De-Contextualization
Song Tang (University of Shanghai for Science and Technology), Xiatian Zhu (University of Surrey)
GenerationTransformerImageTextMultimodality
🎯 What it does: This paper proposes a training-agnostic scene decontextualization method called SDeC, which maintains identity consistency in text-to-image generation without prior knowledge of all scenes.
Consolidating Reinforcement Learning for Multimodal Discrete Diffusion Models
Tianren Ma (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
OptimizationSupervised Fine-TuningReinforcement LearningDiffusion modelMultimodality
🎯 What it does: Introduce MaskGRPO to achieve reinforcement learning optimization for multimodal discrete diffusion models, combining modality-specific importance estimation and sampling methods.
Constant Degree Matrix-Driven Incomplete Multi-View Clustering via Connectivity-Structure and Embedding Tensor Learning
Zhibin Gu (Hebei Normal University), Bing Li (China University of Labor Relations)
OptimizationRepresentation LearningMultimodality
🎯 What it does: Propose an end-to-end multi-view incomplete clustering framework called CAMEL, which directly learns view-specific latent embedding matrices and organizes them into low-rank tensors, while simplifying the Laplacian operator through a constant degree matrix to achieve k-means clustering without post-processing.
Constantly Improving Image Models Need Constantly Improving Benchmarks
Jiaxin Ge (University of California Berkeley), David M. Chan (University of California Berkeley)
GenerationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed and implemented the ECHO framework, automatically constructing an image generation model benchmark that aligns with users' real-world usage scenarios using social media posts.
Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks
Hoagy Cunningham (Anthropic), Mrinank Sharma (Anthropic)
Computational EfficiencyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed and implemented Constitutional Classifiers++, significantly enhancing LLM's anti-jailbreaking capability through methods such as full-context exchange classifier, two-stage classification cascade, and linear probe;
Constrained Decoding of Diffusion LLMs with Context-Free Grammars
Niels Mündler (ETH Zurich), Martin Vechev (ETH Zurich)
GenerationAI Code AssistantLarge Language ModelDiffusion modelText
🎯 What it does: Proposes the first restricted decoding method for diffusion language models (DLMs), enabling the generation of words in arbitrary order while maintaining syntactic constraints and supporting context-free grammar (CFG) constraints.
Constrained Diffusion for Protein Design with Hard Structural Constraints
Jacob K Christopher, Ferdinando Fioretto (University of Virginia)
Protein Structure PredictionDiffusion modelBiomedical Data
🎯 What it does: Proposes a design framework that performs random proximal projection only at the terminal state in restricted diffusion, combined with ADMM decomposition to achieve strict hard constraint satisfaction and maintain diversity in protein structures.
Constraint Matters: Multi-Modal Representation for Reducing Mixed-Integer Linear programming
jiajun Li, Wanyuan Wang (Southeast University)
OptimizationGraph Neural NetworkGraphBenchmark
🎯 What it does: This paper proposes a constraint-based model reduction method that utilizes multi-modal representations to simultaneously consider instance-level and abstract-level MILP information. It identifies and corrects critical tight constraints through an information-theoretic inspired heuristic module, thereby narrowing the search space and accelerating solving.
Constraint-guided Hardware-aware NAS through Gradient Modification
Gregory De Ruyter (KU Leuven), Hans Hallez (KU Leuven)
Computational EfficiencyNeural Architecture SearchConvolutional Neural NetworkBenchmark
🎯 What it does: Proposes a NAS framework called CONNAS that directly enforces hardware constraints through gradient modifications during gradient updates.
Constructive Distortion: Improving MLLMs with Attention-Guided Image Warping
Dwip Dalal (University of Illinois Urbana-Champaign), Unnat Jain (University of California, Irvine)
Image TranslationKnowledge DistillationTransformerVision Language ModelMultimodality
🎯 What it does: During the inference phase, adaptive linear stretching/compression of the input image is performed using cross-attention from a multimodal large language model to magnify regions relevant to the query and compress irrelevant regions while maintaining the global context unchanged.
Contact Wasserstein Geodesics for Non-Conservative Schrödinger Bridges
Andrea Testa (Bosch Center for Artificial Intelligence), Leonel Rozo (Technical University of Denmark)
GenerationOptimizationComputational EfficiencyConvolutional Neural NetworkImageVideoMultimodalityPoint CloudTime SeriesStochastic Differential Equation
🎯 What it does: Propose a non-conservative generalized Schrödinger bridge (NCGSB) and implement a non-iterative near-linear solver (CWG) based on contact Wasserstein geometry, capable of generating energy-variable optimal stochastic paths in probability space.
Contact-guided Real2Sim from Monocular Video with Planar Scene Primitives
Zihan Wang (Carnegie Mellon University), Deva Ramanan (Carnegie Mellon University)
GenerationData SynthesisDepth EstimationReinforcement LearningVision Language ModelVideo
🎯 What it does: Built a complete pipeline CRISP for simulatable 3D human and scene reconstruction from monocular video.
Contamination Detection for VLMs Using Multi‑Modal Semantic Perturbations
Jaden Park (University of Wisconsin-Madison), Yong Jae Lee (University of Wisconsin-Madison)
Anomaly DetectionLarge Language ModelVision Language ModelDiffusion modelMultimodality
🎯 What it does: Addressing the test set leakage issue in Vision-Language Models, we propose a detection method based on multimodal semantic perturbations, and evaluate model generalization differences under varying pollution levels.
Content-Aware Mamba for Learned Image Compression
Yunuo Chen (Shanghai Jiao Tong University), Guo Lu (Shanghai Jiao Tong University)
CompressionAuto EncoderImage
🎯 What it does: Proposed Content-Aware Mamba for Image Compression (CMIC), which enhances the rate-distortion performance of image compression by incorporating content-adaptive token ordering and global prior hints into the Mamba state space model.
Context and Diversity Matter: The Emergence of In-Context Learning in World Models
Fan Wang (Shenzhen Institute of Artificial Intelligence and Robotics for Society), Yu Kang (University of Science and Technology of China)
TransformerAuto EncoderWorld ModelImage
🎯 What it does: Investigated in-context learning (ICL) in world models, theoretically derived and empirically validated two mechanisms: environment recognition (ER) and environment learning (EL), and proposed an efficient long-context linear attention world model, L2World, demonstrating its cross-environment adaptability.
Context Learning for Multi-Agent Discussion
Xingyuan Hua (Tsinghua University), Ju Ren (Tsinghua University)
OptimizationTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Propose M2CL, a context learning method designed for multi-agent discussion (MAD), which dynamically generates and evolves context instructions for each LLM during the discussion process.
Context parroting: A simple but tough-to-beat baseline for foundation models in scientific machine learning
Yuanzhao Zhang (Santa Fe Institute), William Gilpin (University of Texas at Austin)
Recurrent Neural NetworkTransformerTime SeriesElectrocardiogramPhysics Related
🎯 What it does: Propose a zero-shot prediction strategy called Context Parroting, and compare it with various time series baseline models to verify its superiority in low-dimensional chaotic systems and broader SciML tasks.
Context Tokens are Anchors: Understanding the Repeat Curse in dMLLMs from an Information Flow Perspective
Qiyan Zhao (Shanghai Jiao Tong University), Da-Han Wang (Xiamen University of Technology)
Computational EfficiencyTransformerLarge Language ModelVision Language ModelDiffusion modelMultimodality
🎯 What it does: This paper studies the repeat curse problem that occurs during cache acceleration in diffusion-based multimodal LLM (dMLLM) from the information flow perspective, and proposes the CoTA method to alleviate it.
ContextBench: Modifying Contexts for Targeted Latent Activation and Behaviour Elicitation
Robert Graham, Joseph Isaac Bloom
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringDiffusion modelAuto EncoderTextBenchmark
🎯 What it does: Propose the ContextBench benchmark to evaluate the ability of language models to activate specific latent features or behaviors through context modification while maintaining fluency, and improve Evolutionary Prompt Optimisation (EPO) by introducing EPO-Assist and EPO-Inpainting methods.
ContextGen: Contextual Layout Anchoring for Identity-Consistent Multi-Instance Generation
Ruihang Xu (Zhejiang University), Yi Yang (Zhejiang University)
GenerationTransformerSupervised Fine-TuningReinforcement LearningDiffusion modelImageMultimodality
🎯 What it does: Propose ContextGen, a multi-instance generation framework based on Diffusion Transformer, capable of achieving precise layout control and identity consistency simultaneously.
ContextIF: Enhancing Instruction-Following through Context Reward
Yule Zhong (East China Normal University), Guoxiu He (East China Normal University)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Propose ContextIF, which automatically generates high-quality context (constraint summary + examples) for any instruction using reinforcement learning, thereby enhancing the instruction-following capability of LLMs.
ContextNav: Towards Agentic Multimodal In-Context Learning
Honghao Fu (University of Queensland), Yujun Cai (Nanjing University)
Graph Neural NetworkTransformerLarge Language ModelAgentic AIVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposes ContextNav, an agent-based multimodal contextualization framework that integrates automatic retrieval with human-like planning;
ContextPRM: Leveraging Contextual Coherence for multi-domain Test-Time Scaling
Haotian Zhang (Beihang University), Xianglong Liu (Beihang University)
Domain AdaptationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes ContextPRM, a Process Reward Model based on context coherence evaluation, to enhance test-time generalization performance in multi-domain reasoning.
Contextual and Seasonal LSTMs for Time Series Anomaly Detection
Lingpei Zhang (Zhejiang University), Shouling Ji (Zhejiang University)
Anomaly DetectionRecurrent Neural NetworkTime SeriesBenchmark
🎯 What it does: Proposed a prediction-based dual-branch model called CS-LSTMs for detecting subtle anomalies and chronic trend anomalies in univariate time series;
Contextual Causal Bayesian Optimisation
Vahan Arsenyan (CREST), Arnak S. Dalalyan
OptimizationBiomedical Data
🎯 What it does: This paper proposes a unified context and causal Bayesian optimization framework, CoCa-BO, which can simultaneously search for the optimal intervention variable set (policy scope) and corresponding intervention strategies under known causal graphs;
Contextual Multi-Armed Bandits with Minimum Aggregated Revenue Constraints
Ahmed Ben Yahmed (Criteo AI Lab), Vianney Perchet (Criteo AI Lab)
OptimizationReinforcement Learning
🎯 What it does: Proposed a contextual multi-armed bandit model with a minimum aggregate reward constraint, and provided two optimization strategies based on linear programming.
Contextual Similarity Distillation: Ensemble Uncertainties with a Single Model
Moritz Akiya Zanger (Delft University of Technology), Matthijs T. J. Spaan (Delft University of Technology)
Explainability and InterpretabilityKnowledge DistillationReinforcement LearningImage
🎯 What it does: Propose Contextual Similarity Distillation (CSD), a single-model framework for estimating the variance of deep ensemble predictions;
Continual Low-Rank Adapters for LLM-based Generative Recommender Systems
Hyunsik Yoo (University of Illinois Urbana-Champaign), Hanghang Tong (University of Illinois Urbana-Champaign)
Recommendation SystemTransformerLarge Language ModelText
🎯 What it does: This study investigates continuous learning methods for generative recommendation systems based on large language models, proposing a PESO algorithm that incorporates proximal regularization into a single LoRA adapter to address issues of catastrophic forgetting in traditional single adapters and the excessive rigidity of cumulative adapters.
Continual Unlearning for Text-to-Image Diffusion Models: A Regularization Perspective
Justin Lee (Ohio State University), Wei-Lun Chao (Ohio State University)
GenerationTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: Propose and systematically evaluate the continual unlearning problem in text-to-image diffusion models, revealing that existing methods are prone to performance degradation and loss of retained knowledge under sequential requests.
Continuous Audio Language Models
Simon Rouard (Kyutai), Alexandre Défossez (Kyutai)
GenerationTransformerAuto EncoderAudio
🎯 What it does: Propose a Continuous Audio Language Model (CALM) that directly performs autoregressive modeling of audio in the VAE latent space, replacing traditional discrete quantization encoding.
Continuous Chain of Thought Enables Parallel Exploration and Reasoning
Halil Alperen Gozeten (University of Michigan - Ann Arbor), Samet Oymak (University of Michigan - Ann Arbor)
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningReinforcement LearningGraphTabularChain-of-Thought
🎯 What it does: Propose a Continuous Token Chain-of-Thought (CoT2) that enables models to parallelly track multiple reasoning trajectories within a single inference path, significantly enhancing reasoning efficiency and accuracy.
Continuous multinomial logistic regression for neural decoding
Anuththara Rupasinghe (Princeton University), Jonathan W. Pillow (Princeton University)
Explainability and InterpretabilityBiomedical Data
🎯 What it does: Proposed a Continuous Polynomial Logistic Regression (CMLR) model to map neural activity to the full conditional probability density of continuous output variables, enabling nonparametric conditional density estimation for behavioral or perceptual variables.
Continuous Space-Time Video Super-Resolution with 3D Fourier Fields
Alexander Becker (ETH Zurich), Konrad Schindler (ETH Zurich)
Super ResolutionVideo
🎯 What it does: Proposed a continuous spatiotemporal video super-resolution framework V3 based on 3D Fourier field (VFF), capable of super-resolving both space and time at arbitrary scales.
Continuous-Time Value Iteration for Multi-Agent Reinforcement Learning
Xuefeng Wang (Purdue University), Husheng Li (Purdue University)
OptimizationReinforcement LearningBenchmarkPhysics RelatedOrdinary Differential Equation
🎯 What it does: Developed a continuous-time multi-agent reinforcement learning framework CT-MARL based on physics-informed neural networks (PINN), utilizing the Hamilton-Jacobi-Bellman (HJB) equation and value gradient iteration to achieve high-dimensional value function approximation and train distributed policies.
Continuously Augmented Discrete Diffusion model for Categorical Generative Modeling
Huangjie Zheng (Apple), Yizhe Zhang (Apple)
GenerationDiffusion modelImageTextSequential
🎯 What it does: Propose a Continuous-Augmented Discrete Diffusion Model (CADD) that combines traditional masked diffusion with Gaussian diffusion in a continuous latent space;
Continuum Transformers Perform In-Context Learning by Operator Gradient Descent
Yash Patel (University of Michigan), Ambuj Tewari (University of Michigan)
OptimizationRepresentation LearningMeta LearningTransformerPhysics Related
🎯 What it does: Investigated the theoretical mechanisms of continuous Transformers in context-free learning, proving their equivalence to gradient descent in operator RKHS.
Contraction and Hourglass Persistence for Learning on Graphs, Simplices, and Cells
Mattie Ji (University of Pennsylvania), Vikas K Garg
Representation LearningDrug DiscoveryGraph Neural NetworkGraph
🎯 What it does: Proposes Contraction Homology (CH), novel Backward PH, Forward-Backward (FB) persistence, and 'Hourglass persistence,' extending them to simplicial and cell complexes; designs an efficient algorithm embeddable into end-to-end differentiable graph neural networks.
Contractive Diffusion Policies
Amin Abyaneh (McGill University), Hsiu-Chin Lin (McGill University)
Robotic IntelligenceReinforcement LearningDiffusion modelScore-based ModelStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed Contractive Diffusion Policies (CDPs), which enhance the robustness and performance of offline learning by introducing contraction during the diffusion sampling process.
Contrastive Diffusion Guidance for Spatial Inverse Problems
Sattwik Basu (University of Illinois Urbana-Champaign), Romit Roy Choudhury (University of Illinois Urbana-Champaign)
RestorationGenerationDiffusion modelContrastive LearningImageSequential
🎯 What it does: Propose constructing a smooth embedding space via contrastive learning, guiding diffusion models to solve spatial inverse problems (e.g., reconstructing floor plans from human walking trajectories) and broader blind inverse problems.
Contrastive Predictive Coding Done Right for Mutual Information Estimation
Jongha Jon Ryu (Massachusetts Institute of Technology), Gregory W. Wornell (Massachusetts Institute of Technology)
Representation LearningContrastive LearningImageTextBiomedical DataBenchmark
🎯 What it does: Investigate the defects of InfoNCE in mutual information estimation, propose InfoNCE-anchor, and implement consistent density ratio estimation through the proper scoring rule framework, thereby improving MI estimation accuracy and downstream prediction task performance.
Control Tax: The Price of Keeping AI in Check
Mikhail Terekhov (EPFL), Samuel Albanie
Safty and PrivacyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed and quantified the concept of 'control tax' to measure the operational costs and safety benefits generated by integrating AI supervision mechanisms (such as monitoring models) into practical systems.
Controllable diffusion-based generation for multi-channel biological data
Haoran Zhang (University of Texas at Austin), Wesley Tansey (Memorial Sloan Kettering Cancer Center)
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposed a controllable multi-channel diffusion generation framework (MCD) that can generate high-dimensional biological multi-channel data with spatial alignment under any combination of observed and missing channels, supporting single-cell, spatial protein imaging, and missing modality MRI completion.
Controllable Exploration in Hybrid-Policy RLVR for Multi-Modal Reasoning
Zhuoxu Huang, Jungong Han (Tsinghua University)
Reinforcement LearningMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a hybrid strategy RLVR framework named CalibRL for controlled exploration in multimodal reasoning models.
Controllable First-Frame-Guided Video Editing via Mask-Aware LoRA Fine-Tuning
Chenjian Gao (Chinese University of Hong Kong), Tianfan Xue (SenseTime Research)
GenerationSupervised Fine-TuningDiffusion modelVideo
🎯 What it does: Proposes a controllable first-frame guided video editing framework utilizing mask-aware LoRA fine-tuning, enabling flexible and fine-grained control over the entire video editing process without modifying the model architecture.
Controllable Logical Hypothesis Generation for Abductive Reasoning in Knowledge Graphs
Yisen Gao (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningGraph
🎯 What it does: Studied controlled inductive reasoning in knowledge graphs, proposing the CtrlHGen framework to generate logical hypotheses that comply with semantic and structural constraints.
Controllable Sequence Editing for Biological and Clinical Trajectories
Michelle M Li, Marinka Zitnik (Harvard University)
Recurrent Neural NetworkTransformerTime SeriesSequentialBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes CLEF, a controllable sequence editing framework that utilizes learned temporal concepts to achieve conditional long-sequence generation and causal inference;
Controllable Video Generation with Provable Disentanglement
Yifan Shen (Mohamed bin Zayed University of Artificial Intelligence), Kun Zhang (Carnegie Mellon University)
GenerationData SynthesisRecurrent Neural NetworkFlow-based ModelGenerative Adversarial NetworkVideo
🎯 What it does: In this work, we propose a controllable video generation framework called CoVoGAN based on GAN, and achieve separable and fine-grained control of static factors (content) and dynamic factors (style) in videos by introducing a Temporal Transition Module (TTM);