ICML 2026 Papers — Page 11
International Conference on Machine Learning · 6554 papers
CombinationTS: A Modular Framework for Understanding Time-Series Forecasting Models
Xiaorui Wang (Institute of Computing Technology, Chinese Academy of Sciences), Jianfeng Zhan (Institute of Computing Technology, Chinese Academy of Sciences)
Explainability and InterpretabilityMixture of ExpertsTime SeriesBenchmark
🎯 What it does: Proposes the CombinationTS framework, which uses modular and probabilistic evaluation methods to decompose and attribute the performance of time series forecasting models.
Combinatorial Sparse PCA Beyond the Spiked Identity Model
Syamantak Kumar (University of Texas Austin), Peiyuan Zhang (University of Wisconsin Madison)
OptimizationComputational EfficiencyRepresentation LearningTabularReview/Survey PaperBenchmark
🎯 What it does: Proposed a composite algorithm for solving sparse PCA under the non-pulse identity covariance model (Model 2), and provided theoretical convergence guarantees;
CoME: Empowering Channel-of-Mobile-Experts with Informative Hybrid-Capabilities Reasoning
Yuxuan Liu (Renmin University of China), Rui Yan (Wuhan University)
Computational EfficiencyKnowledge DistillationRepresentation LearningRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsVision-Language-Action ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Designed and implemented a mobile agent architecture called CoME based on output-oriented expert activation, combining four specialized experts (screen summarization, subtask planning, action decision-making, and action function) to achieve multi-stage hybrid capability reasoning, and proposed an information gain-driven DPO training strategy to reduce error propagation in intermediate steps.
CoMem: Context Management with A Decoupled Long-Context Model
Yuwei Zhang (University of California, San Diego), Bing Yin (Amazon)
Computational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Propose the COMEM framework, which decouples long-context compression from the main inference model, using a k-step-off asynchronous pipeline to allow the memory compression model to complete in the background in parallel. The main inference model uses compressed summaries and a buffer for recent interactions to generate actions, significantly reducing inference latency.
Commit to the Bit: Reactive Reinforcement Learning Done Right
Onno Eberhard (Max Planck Institute for Intelligent Systems), Michael Muehlebach (Max Planck Institute for Intelligent Systems)
Reinforcement Learning
🎯 What it does: This paper proposes an algorithm called Committed Q-learning and proves that it can converge to the optimal reactive policy in finite, partially observable environments with deterministic observations.
Comp-Attn: Present-and-Align Attention for Compositional Video Generation
Hongyu Zhang (Peking University), Jie Chen (Peking University)
GenerationTransformerLarge Language ModelVision Language ModelDiffusion modelVideoText
🎯 What it does: Propose the Comp-Attn mechanism, achieving the 'Present-and-Align' paradigm of cross-attention, to address the issues of subject absence and relationship misalignment in multi-subject video generation.
Compact Conformal Subgraphs
Sreenivas Gollapudi (Google Research), Aravindan Vijayaraghavan (Northwestern University)
CompressionOptimizationGraph Neural NetworkGraph
🎯 What it does: Propose a graph-based compliance compression framework that compresses large-scale structured prediction ensembles into compact subgraphs while preserving statistical validity.
Compass-RoPE: Isotropic Rotary Position Embeddings for Vision Transformers
Chengxi Min (Beijing Jiaotong University), Yao Zhao (Beijing Jiaotong University)
ClassificationObject DetectionSegmentationTransformerImage
🎯 What it does: Propose Compass‑RoPE, a polar coordinate parameterized rotational position embedding, and enhance the robustness of Vision Transformer in resolution extrapolation by improving directional uniformity through uniform angular initialization combined with internal DFT mixing.
Compile to Compress: Boosting Formal Theorem Provers by Compiler Outputs
Guchan Li (Tsinghua University), Hongning Wang (Tsinghua University)
OptimizationComputational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Propose a learned distillation framework that leverages structured error information from the Lean compiler to compress failure modes, achieving efficient self-correction and search.
CompleteP for RL: Maintaining Feature Learning When Scaling Deep Reinforcement Learning
Adam Lee (Harvard University), Cengiz Pehlevan (Harvard University)
Convolutional Neural NetworkReinforcement LearningTabularTime SeriesSequential
🎯 What it does: This paper systematically evaluates the impact of scaling on feature learning, learning rate transfer, policy consistency, and computational and reward efficiency in residual networks by changing the parameterization of the RL agent from NTK ("lazy" mode) to CompleteP ("rich" mode).
Complexity Bounds for Dirichlet Process Slice Samplers
Beatrice Franzolini (King's College London), Francesco Gaffi (University of Bergamo)
OptimizationComputational EfficiencyReview/Survey Paper
🎯 What it does: Studies the computational complexity of slice sampling in Dirichlet process (DP) models, and provides a high-probability upper bound for its performance under arbitrary posterior clustering growth scenarios.
Complexity of Decentralized Optimization with Mixed Affine Constraints
Demyan Yarmoshik (MIRAI), Alexander Gasnikov (MIRAI)
OptimizationFederated Learning
🎯 What it does: This paper studies convex optimization problems with hybrid affine constraints in distributed networks, and unifies typical scenarios such as horizontal and vertical federated learning, and distributed control.
ComplexMCP: Evaluation of LLM Agents in Dynamic, Interdependent, and Large-Scale Tool Sandbox
Yuanyang Li (Zhejiang University), Hongyang Chen (Zhejiang Lab)
TransformerLarge Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the ComplexMCP benchmark to evaluate the performance of LLM agents in large-scale, interdependent, randomized tool sandboxes.
Component-Wise Composite Likelihood Distillation for Censored Time-to-Event Data
Feiyang Deng (University of Michigan), Kevin He (University of Michigan)
Knowledge DistillationRepresentation LearningData-Centric LearningDiffusion modelScore-based ModelContrastive LearningTabularBiomedical DataElectronic Health Records
🎯 What it does: Propose a knowledge distillation framework for right-censored time-to-event data, aligning teacher and student models within each comparison set using a locally normalized composite likelihood.
Compositional Behavioral Semantics for State Abstraction in Reinforcement Learning
Yivan Zhang (University of Tokyo), Manuel Baltieri (Araya Inc.)
Reinforcement Learning
🎯 What it does: This paper proposes a unified and composable framework for defining, analyzing, and transferring behavioral structures (such as invariants, value functions, bisimulation relations/metrics, etc.) in reinforcement learning, and treats state abstraction as coalgebra homomorphism;
Compositional Generalization Requires Linear, Orthogonal Representations in Vision Embedding Models
Arnas Uselis (University of Tübingen), Seong Joon Oh (University of Tübingen)
ClassificationRetrievalRepresentation LearningTransformerContrastive LearningImageMultimodality
🎯 What it does: Investigated the representation structure of visual embedding models (such as CLIP, SigLIP, DINO) in compositional generalization tasks, proving that under three necessary conditions—separability, transferability, and stability—the representations must be linearly decomposable and the differential vectors of different concepts must be orthogonal, and provided a theoretical lower bound for the minimum embedding dimension d ≥ k;
Compositional Generative Modeling from Decentralized Data
Mashrur M. Morshed (Michigan State University), Vishnu Boddeti (Michigan State University)
GenerationData SynthesisFederated LearningKnowledge DistillationTransformerMixture of ExpertsScore-based ModelFlow-based ModelImageTabularTime Series
🎯 What it does: Build a generative model that can learn from distributed data silos and generate composite samples without sharing raw data.
Compositional Planning with Jumpy World Models
Jesse Farebrother (McGill University), Ahmed Touati (FAIR at Meta)
TransformerReinforcement LearningFlow-based ModelWorld ModelTabularTime SeriesSequentialBenchmark
🎯 What it does: This paper proposes a compositional planning framework based on a multi-step jumping world model (jumpy world model), which can combine pre-trained parameterized policies across time scales to solve long-term tasks without additional training.
Compositional Transduction with Latent Analogies for Offline Goal-Conditioned Reinforcement Learning
Junseok Kim (Seoul National University), Songhwai Oh (Seoul National University)
TransformerReinforcement LearningContrastive LearningTabularSequentialBenchmark
🎯 What it does: Propose a method for achieving compositional generalization in offline goal-conditioned reinforcement learning through 'analogy transduction,' constructing intrinsic task analogies and combining them with different environmental contexts to generate new plans.
Compress then Merge: From Multiple LoRAs into One Low-Rank Adapter
Zhengbao He (Shanghai Jiao Tong University), Xiaolin Huang (Shanghai Jiao Tong University)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningImageText
🎯 What it does: This paper proposes a method called Compress-then-Merge (CtM), which merges multiple low-rank adapters (LoRA) into a single low-rank update. It enforces the target rank constraint before merging, avoiding information loss caused by post-hoc truncation.
Compressed Sensing for Capability Localization in Large Language Models
Anna Bair (Carnegie Mellon University), J Zico Kolter (Carnegie Mellon University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Locate task-related attention heads in the Transformer model and evaluate their impact on various capabilities.
Compression as Adaptation: Implicit Visual Representation with Diffusion Foundation Models
Zongyu Guo (Microsoft Research Asia), Yan Lu (Microsoft Research Asia)
CompressionRepresentation LearningTransformerDiffusion modelAuto EncoderImageVideo
🎯 What it does: Treat visual signals (images or videos) as generative functions, perform low-rank adaptation (LoRA) learning on frozen large-scale diffusion generative models, and hash the adapted parameters into a single vector, achieving ultra-low bitrate compression and reconstruction.
Computational Arbitrage in AI Model Markets
Ricardo Olmedo (Max Planck Institute for Intelligent Systems), Moritz Hardt (Max Planck Institute for Intelligent Systems)
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextFinance RelatedRetrieval-Augmented Generation
🎯 What it does: This study explores the possibility of computational arbitrage in the AI model market, demonstrating how arbitrageurs can effectively allocate inference budgets across different model providers to create competitive products.
Computationally-efficient Graph Modeling with Refined Graph Random Features
Krzysztof Marcin Choromanski (Google DeepMind), Isaac Reid (Google DeepMind)
Computational EfficiencyRepresentation LearningGraph Neural NetworkContrastive LearningImageGraphTabularTime SeriesSequential
🎯 What it does: Proposes GRFs++, an improved graph random feature method for efficiently approximating graph kernels.
Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision
Dulhan Jayalath (University Of Oxford), Alan Schelten (Meta Superintelligence Labs)
Data SynthesisComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextTabularRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Compute as Teacher (CaT) framework, which generates pseudo reference answers by leveraging parallel rollout computations during inference, thereby enabling RL training in an unlabeled environment;
Compute Where it Counts: Self Optimizing Language Models
Yash Akhauri (Cornell University), Mohamed S. Abdelfattah (Cornell University)
OptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Proposes an Adaptive Computation Budget Language Model (SOL), which dynamically allocates computational resources for attention sparsification, MLP channel pruning, and activation quantization during inference based on the difficulty of each token.
Concept Concentration for Faithful Representation Intervention
Hongzheng Yang (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Safty and PrivacyExplainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the COCA method, which adds explicit concept reasoning labels (think, <concept>, <check>, <erase>, <response>) to the training data, enabling LLMs to aggregate harmful concepts into a linear subspace within the internal representation space, thereby achieving safe alignment under existing linear interventions (ReFT/LoFiT).
Concept Heterogeneity-aware Representation Steering
Laziz Abdullaev (National University Of Singapore), Tan Minh Nguyen (National University Of Singapore)
Safty and PrivacyExplainability and InterpretabilityRepresentation LearningLarge Language ModelImageTextMultimodality
🎯 What it does: Propose a concept heterogeneity-aware representation steering method called CHaRS (Concept Heterogeneity-aware Representation Steering), based on Gaussian Mixture Models and Optimal Transport. It clusters semantic concepts into multi-modal subspaces and solves discrete OT between these subspaces to obtain smooth activation vector fields.
Concept Removal for Frontier Image Generative Models
Aditya Kumar (CISPA Helmholtz Center for Information Security), Franziska Boenisch (CISPA Helmholtz Center for Information Security)
GenerationSafty and PrivacyExplainability and InterpretabilityTransformerPrompt EngineeringDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a concept elimination framework called BLOCK based on a transcoder, which directly replaces the bottleneck layer of the text-to-generator in state-of-the-art text-image generation models (such as SD3.5, Flux, Infinity-2B/8B), thereby achieving concept elimination without modifying the backend network.
Concept Removal Guidance: Evidence-Calibrated Negative Guidance for Safe Diffusion Sampling
Yoonseok Choi (KAIST), Kee-Eung Kim (KAIST)
Safty and PrivacyPrompt EngineeringDiffusion modelImageText
🎯 What it does: Propose a training-agnostic concept removal guidance (CRG) during inference, which suppresses unsafe concepts in text-to-image diffusion models by dynamically calibrating negative prompts.
Concept-Guided Tokenization: Closing the Gap Between Reconstruction and Generation
Yunqiao Yang (Nanyang Technological University), Ying Wei (Zhejiang University)
GenerationRepresentation LearningTransformerVision Language ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextMultimodality
🎯 What it does: Proposes a text-conditioned and concept-guided image tokenizer (ConceptTok), which improves generation quality while maintaining reconstruction quality by integrating text information into the encoder and using sparse autoencoders (SAE) to map pre-trained vision-language model features into a sparse, decoupled concept space.
ConceptMoE: Adaptive Token-to-Concept Compression for Implicit Compute Allocation
Zihao Huang, Ge Zhang (Bytedance Seed)
CompressionComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsAuto EncoderImageTextMultimodality
🎯 What it does: Propose ConceptMoE, which adaptively compresses tokens into concepts by learning dynamic chunking of similar tokens, achieving a token-to-concept compression.
CONCUR: High-Throughput Agentic Batch Inference of LLM via Congestion-Based Concurrency Control
qiaoling chen, Tianwei Zhang (Nanyang Technological University)
Computational EfficiencyTransformerLarge Language ModelAgentic AIText
🎯 What it does: Proposes CONCUR, a flow-control-based agent-level concurrent scheduling layer, aimed at preventing mid-term thrashing in LLM KV cache and improving the throughput of offline agent batch inference.
Condition Number Based Low-Bit Quantization for Image Super-Resolution
Kai Liu (Shanghai Jiao Tong University), Linghe Kong (Shanghai Jiao Tong University)
Super ResolutionTransformerImage
🎯 What it does: Designed a low-bit quantization method called CondiQuant based on the condition number, for post-training quantization of image super-resolution models.
Condition-Aware Graph Flow Matching for Modeling the Distributions of Complex Fluid Systems
Xiaochao Deng (Sichuan University), Xiaogang Deng (Sichuan University)
Graph Neural NetworkFlow-based ModelGraphPhysics Related
🎯 What it does: This paper proposes a condition-aware graph flow matching (CGFM) method to learn and sample equilibrium state distributions of complex fluid systems from short trajectories, avoiding long-term numerical simulations;
Conditional Clifford-Steerable CNNs for PDE Modeling
Bálint László Szarvas (University of Amsterdam), Maksim Zhdanov (University of Amsterdam)
Convolutional Neural NetworkTabularPhysics Related
🎯 What it does: Proposed a conditioned Clifford-Steerable CNN (C-CSCNN) to enhance the expressiveness of convolutional kernels in PDE modeling.
Conditional Coverage Diagnostics for Conformal Prediction
Sacha Braun (Inria), Francis Bach (Inria)
Anomaly DetectionExplainability and InterpretabilityData-Centric LearningSupervised Fine-TuningContrastive LearningTabularTime SeriesBenchmark
🎯 What it does: This paper proposes a new framework that transforms conditional coverage evaluation into a classification task, and quantifies the failure of conditional coverage in conformal prediction by constructing a metric called 'Excess Risk of Target Coverage' (ERT).
Conditional Diffusion Sampling
Francisco M Castro-Macías, José Miguel Hernández-Lobato (University of Cambridge)
OptimizationDiffusion modelScore-based ModelMultimodalityPoint CloudGraphTabularTime SeriesPhysics RelatedStochastic Differential Equation
🎯 What it does: A new training-free, two-stage sampling framework called Conditional Diffusion Sampling (CDS) is proposed, which combines Parallel Tempering with closed-form SDE transport to address the sampling problem of unnormalized multi-modal distributions.
Conditional Distributional Treatment Effects: Doubly Robust Estimation and Testing
Saksham Jain (University of Washington), Alex Luedtke (Harvard Medical School)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyData-Centric LearningImageTabularFinance Related
🎯 What it does: Proposed a new conditional distribution effect estimator (SCoDiTE), and provided a doubly robust, locally asymptotically optimal estimator and a permutation-free kernel-based test to test global conditional distribution homogeneity.
Conditional Equivalence of DPO and RLHF: Assumptions, Failure Modes, and Provable Alignment
Zhiqin Yang (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Investigate the conditional equivalence between DPO and RLHF, reveal the implicit assumptions of DPO, and propose Constrained Preference Optimization (CPO) and its reward-model-free version E-CPOC to address the mismatch problem of DPO when the reference policy is poor.
Conditional KRR: Injecting Unpenalized Features into Kernel Methods with Applications to Kernel Thresholding
Rustem Takhanov (Nazarbayev University), Zhenisbek Assylbekov (Purdue University Fort Wayne)
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningImageTabular
🎯 What it does: Propose and study Conditional Kernel Ridge Regression (Conditional KRR), by injecting unregularized features into kernel methods, and interpreting it as equivalent to a two-step regression using residual kernels.
Conditional Quantile Adjusted Conformal Prediction for Time Series
Cheng Yu (University of Chicago), Ke Zhu (University of Hong Kong)
Anomaly DetectionOptimizationFederated LearningComputational EfficiencyData-Centric LearningDiffusion modelScore-based ModelAuto EncoderContrastive LearningTabularTime SeriesFinance Related
🎯 What it does: Proposed a new method for constructing adaptive confidence intervals for time series — Conditional Quantile Adjusted Conformal Prediction (CQACP), which achieves smooth, nested, and coverage-accurate prediction intervals by fitting the Cornish-Fisher approximation to the conditional quantile curve of residual non-consistency scores;
Conditionally Site-Independent Neural Evolution of Antibody Sequences
Stephen Zhewen Lu (University of California Berkeley), Yun S. Song (University of California Berkeley)
Drug DiscoveryProtein Structure PredictionTransformerLarge Language ModelSequentialBiomedical DataStochastic Differential Equation
🎯 What it does: Propose the COSINE model, which combines deep neural networks with continuous-time Markov chains to simulate the antibody affinity maturation process.
ConEx: Human-Interpretable Saliency Maps via Concept-Aware Attribution
Yehonatan Elisha (Tel Aviv University), Noam Koenigstein (Tel Aviv University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: By automating concept discovery and spatial localization, ConEx generates interpretable concept heatmaps and combines gradients to achieve both local and global explanations for classification.
Conf-Gen: Conformal Uncertainty Quantification for Generative Models
Gabriel Loaiza-Ganem (Layer 6 AI), Kin Kwan Leung (Layer 6 AI)
GenerationData SynthesisExplainability and InterpretabilityComputational EfficiencyTransformerDiffusion modelScore-based ModelGenerative Adversarial NetworkImageTextSequential
🎯 What it does: Propose the Conf-Gen framework, extending conformal risk control (CRC) to generative models, achieving distribution-free uncertainty quantification for generated results.
Confidence is Not Universal: Task-Dependent Calibration and Emergent Behavior in LLMs
Chaeyun Jang (Korea Advanced Institute of Science and Technology), Hyungi Lee (Kookmin University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper investigates the transferability of large language models (LLMs) in verbalizing confidence calibration across different tasks, finding an essential conflict between a unified confidence scale for reasoning tasks and retrieval/copypaste tasks, and proposes a discretization method that splits confidence into reasoning confidence and evidence confidence, achieving reliable confidence alignment across task families without using reinforcement learning.
Configurable Reward Model for Balanced Safety Alignment
Zhengping Jiang (Johns Hopkins University), Li Chen (Meta Superintelligence Labs)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningText
🎯 What it does: Proposed a configurable safety reward model called CSRM, which can output dense and calibrated reward signals during inference based on natural language safety configurations;
Conflict-Aware Adaptive Alignment for LLM Hallucination Mitigation
Ruohan Zong (University of Illinois UrbanaChampaign), Dong Wang (Miami University)
OptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes the Conflict-Aware Adaptive Margin Preference Alignment (CAMP) method to reduce hallucinations in large language models while maintaining helpfulness.
Conflict-Aware Additive Guidance for Flow Models under Compositional Rewards
Xuehui Yu (National University of Singapore), Harold Soh (National University of Singapore)
OptimizationRobotic IntelligenceReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelImageTextPoint CloudTabular
🎯 What it does: This paper proposes a conflict-aware incremental guidance (CAR Guidance) method for multi-constraint (multi-reward) guidance during inference on flow models (Flow Matching/Conditional Flow Matching), addressing the problem of trajectory deviation from the true data manifold (off-manifold drift) caused by multiple constraints.
Conflicting Biases at the Edge of Stability: Norm versus Sharpness Regularization
Maria Matveev (LMU Munich), Johannes Maly (LMU Munich)
OptimizationRepresentation LearningHyperparameter SearchConvolutional Neural NetworkRecurrent Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Systematic experiments on implicit regularization of gradient descent (GD) under different learning rates are conducted to explore the interaction between norm regularization and sharpness regularization, and theoretical analysis is provided using a simplified diagonal linear network, demonstrating that a single implicit bias is insufficient to explain good generalization.
ConFlux: Multivariate Time Series in Flux, One Unified Forecast in Confluence
Shiyu Wang (ByteDance), Yang Xiang (ByteDance)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningTabularTime Series
🎯 What it does: Proposes ConFlux, a general-purpose multivariate time series foundation model that can uniformly predict multivariate sequences in zero-shot, fine-tuning, and from-scratch training scenarios.
Conformal C2ST: Turning weak classifiers into strong two-sample tests
Vansh Bansal (University of Texas at Austin), James G. Scott (University of Texas at Austin)
Anomaly DetectionComputational EfficiencyRepresentation LearningContrastive LearningImagePoint CloudGraphTabularBenchmarkPhysics Related
🎯 What it does: Propose a conformal two-sample test (Conformal C2ST) that utilizes deformation correction, converting any weak, biased, or overfitted classifier into a strong test with strict finite-sample control of type I error.
Conformal Calibration Transfer
Achref Doula (Technical University of Darmstadt)
Domain AdaptationContrastive LearningImageMultimodality
🎯 What it does: In scenarios where there are labeled calibration data in the source space and unlabeled data in the target space, we propose Transported Conformal Calibration (TCC), which learns input mappings through unlabeled paired data, migrates source calibration samples to the target space, and then uses the unlabeled information in the target space to correct residuals, thereby achieving confidence set generation for the unlabeled target domain.
Conformal Path Reasoning: Trustworthy Knowledge Graph Question Answering via Path-Level Calibration
Shuhang Lin (Rutgers University), Dimitris N. Metaxas (Rutgers University)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: Proposed the Conformal Path Reasoning (CPR) framework, which achieves trustworthy reasoning for knowledge graph question answering through path-level non-conformity measures;
Conformal Policy Control
Drew Prinster (Prescient Design, Genentech), Samuel Don Stanton
OptimizationSafty and PrivacyReinforcement LearningTextTabularBiomedical DataElectronic Health Records
🎯 What it does: Proposes a safety exploration framework called Conformal Policy Control (CPC), which utilizes calibration data from existing safe policies. By interpolating between safe policies and optimized policies through thresholding the likelihood ratio, it ensures risk tolerance during deployment.
Conformal Prediction for Early Stopping in Mixed Integer Optimization
Stefan Clarke (Princeton University), Bartolomeo Stellato (Princeton University)
OptimizationRecurrent Neural NetworkTransformerTime Series
🎯 What it does: Propose a technique that uses a neural network to predict the current true optimality gap of a mixed integer programming solver, and calibrates the threshold using conformal prediction, thereby allowing the solver to stop early while satisfying a given error tolerance and confidence level.
Conformal Reliability: A New Evaluation Metric for Conditional Generation
Yachen Gao (Fudan University), Yanwei Fu (Shanghai Innovation Institute)
GenerationData SynthesisExplainability and InterpretabilityComputational EfficiencyTransformerDiffusion modelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Propose a reliability score based on conformal prediction to evaluate the worst-case performance of conditional generative models under a given confidence level, and design the CReL framework to achieve efficient computation;
Conformal Risk-Averse Decision Making with Action Conditional Guarantee
Zihan Zhu (University of Pennsylvania), Hamed Hassani (University of Pennsylvania)
Recommendation SystemOptimizationImageTabular
🎯 What it does: Propose action-conditional synthetic prediction (AC-RAC) for risk-averse decision making, providing distribution-free action-conditional safety guarantees.
Conformal Thinking: Risk Control for Reasoning on a Compute Budget
Xi Wang (Johns Hopkins University), Eric Nalisnick (Johns Hopkins University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a dual-threshold early stopping framework based on risk control (Conformal Thinking), which automatically determines when to stop inference in large language models while meeting a specified error rate.
ConFu: Contemplate the Future for Better Speculative Sampling
Zongyue Qin (University of California Los Angeles), Yizhou Sun (University of California Los Angeles)
GenerationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose ConFu, a reasoning framework that improves the quality of draft models by 'thinking about the future,' combining pause (think) tokens, soft prompts, dynamic MoE mechanisms, and robust training;
CONGA:Confidence-and-Gradient-Aware Learning Rate Schedule for Test Time Adaptation
Lv Shaoran (University of Electronic Science and Technology of China), Jingjing Li (University of Electronic Science and Technology of China)
Domain AdaptationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Proposed CONGA, a learning rate scheduler based on confidence and gradient, to improve the optimization process of test-time adaptation (TTA);
Connecting Independently Trained Modes via Layer-Wise Connectivity
Yongding Tian (Delft University of Technology), H Peter Hofstee
OptimizationComputational EfficiencyKnowledge DistillationRepresentation LearningConvolutional Neural NetworkAuto EncoderContrastive LearningImage
🎯 What it does: Proposed a new empirical algorithm called LLPF (Low-Loss Path Finding), which can construct continuous low-loss paths between independently trained models under different random seeds or different training hyperparameters.
ConPress: Learning Efficient Reasoning from Multi-Question Contextual Pressure
Jie Deng (University of Chinese Academy of Sciences), Yutao Xie (Microsoft)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Studied the naturally occurring self-compression phenomenon in large reasoning models under multi-question prompting, and proposed the ConPress self-supervised fine-tuning framework to achieve shorter chain-of-thought reasoning trajectories in single-question reasoning.
Conservation Laws for Modern Neural Architectures
Viet-Hoang Tran (National University of Singapore), Tan Minh Nguyen
OptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerMixture of ExpertsContrastive LearningImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a unified theoretical framework that provides a complete characterization of conserved quantities during the gradient descent process in modern deep networks (including GeLU, SiLU, SwiGLU activation, MHA, RoPE position encoding, and multi-expert architectures).
ConServe: Fine-Grained GPU Harvesting for LLM Online and Offline Co-Serving
Yifan Qiao (University of California Berkeley), Harry Xu (University of California Los Angeles)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the ConServe system, which allows online low-latency requests and offline batch processing tasks to coexist on the same GPU, leveraging fine-grained scheduling, hierarchical preemption, and incremental KV cache management to achieve efficient GPU resource utilization.
Consistency Deep Equilibrium Models
Junchao Lin (Huazhong University of Science and Technology), Robert C Qiu
OptimizationComputational EfficiencyKnowledge DistillationTransformerImageTextGraphTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose Consistency Deep Equilibrium Model (C-DEQ), which accelerates inference significantly by mapping intermediate states directly to the equilibrium point of DEQ through consistency distillation
Consistency Training Can Entrench Misalignment
David Demitri Africa (UKAI Security Institute), Arathi Mani (UKAI Security Institute)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper studies the impact of consistency training on language model alignment, exploring whether it amplifies or suppresses existing misleading behaviors.
Consistent Diffusion Language Models
Hasan Amin (Purdue University), Xia Song (Microsoft)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelScore-based ModelText
🎯 What it does: Train a consistent diffusion language model using the multi-path discrete consistency (MPDC) principle to learn path-agnostic denoisers, achieving high-quality text generation with few steps.
Consistent Zero-Shot Imitation with Contrastive Goal Inference
Kathryn Wantlin (Princeton University), Benjamin Eysenbach (Princeton University)
Robotic IntelligenceReinforcement LearningContrastive LearningBenchmark
🎯 What it does: Proposes the Contrastive Inverse Reinforcement Learning (CIRL) framework, which achieves zero-shot imitation by utilizing self-supervised contrastive RL pre-training, automatic goal sampling, and variational goal inference;
ConsMSA: Semantic Distribution Consistency Learning for Multimodal Sentiment Analysis
Pan Wang (University of Pittsburgh), Jingtong Hu (University of Pittsburgh)
ClassificationComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringAuto EncoderContrastive LearningTextMultimodalityAudio
🎯 What it does: Propose the ConsMSA framework, which calculates consistency scores (I2CS) for intra-modal and inter-modal tokens in the semantic distribution space, combines consistency with predictive relevance to form importance signals, and applies them for consistency regularization, soft/hard token weighting, and compression, thereby enhancing the accuracy and efficiency of multi-modal sentiment analysis.
Constitutional Black-Box Monitoring for Scheming in LLM Agents
Simon Storf (MATS Research), Marius Hobbhahn (Apollo Research)
Anomaly DetectionOptimizationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Develop a constitutional-style black-box monitor that trains models using synthetic trajectories to detect conspiratorial behavior in LLM agents.
Constrained Adaptive Rejection Sampling
Paweł Parys (University of Warsaw), Loris D'Antoni (University of California-San Diego)
GenerationData SynthesisOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextSequential
🎯 What it does: Propose an algorithm for precise sampling under constrained generation in language models—Constrained Adaptive Rejection Sampling (CARS)—which significantly improves sampling efficiency while maintaining distribution integrity by dynamically recording invalid prefixes and reweighting the distribution.
Constrained Bayesian Experimental Design via Online Planning
Yujia Guo (ELLIS Institute Finland), Ayush Bharti (Aalto University)
OptimizationComputational EfficiencyTransformerReinforcement LearningContrastive LearningTabularBenchmarkPhysics RelatedStochastic Differential Equation
🎯 What it does: Proposed a semi-consumptive method called COPEx for handling dynamic constraints in Bayesian Experimental Design (BED), combining offline pre-trained design policies and posterior networks, and performing online planning through multi-step scenario trees during runtime;
Constrained Flow Optimization via Sequential Fine-Tuning for Molecular Design
Sven Gutjahr (ETH Zurich), Andreas Krause (ETH Zurich)
OptimizationDrug DiscoverySupervised Fine-TuningReinforcement LearningDiffusion modelFlow-based ModelGraphBiomedical Data
🎯 What it does: This paper proposes a sequential fine-tuning algorithm (CFO) under an augmented Lagrangian framework, achieving constrained generation optimization on pre-trained flow/diffusion models.
Constrained hybrid modelling to predict microbial dynamics and organic matter turnover in soil systems
Paul Collart (Forschungszentrum Jülich GmbH), Lars Doorenbos (University of Bonn)
OptimizationRecurrent Neural NetworkTabularTime SeriesAgriculture RelatedOrdinary Differential Equation
🎯 What it does: Proposed the HySoMi hybrid model, which uses genomic features to predict soil microbial dynamics and the biodynamic parameters of organic matter decomposition;
Constrained Meta Reinforcement Learning with Provable Test-Time Safety
Tingting Ni (EPFL), Maryam Kamgarpour (EPFL)
OptimizationSafty and PrivacyMeta LearningReinforcement LearningMixture of ExpertsTabularTime Series
🎯 What it does: Proposed a restricted meta reinforcement learning algorithm that learns a set of approximately optimal policies and a safe policy feasible for all tasks through a CMDP collection covering the task distribution during the training phase; during the testing phase, a hybrid policy is used to gradually improve, ensuring safe exploration and providing theoretical proofs of sample complexity and convergence.
Constrained Multi-Objective Reinforcement Learning with Max-Min Criterion
Giseung Park (University of Toronto), Youngchul Sung (KAIST)
OptimizationReinforcement LearningTabularTime SeriesSequential
🎯 What it does: Proposes a multi-objective reinforcement learning framework that integrates max-min fairness with constraint satisfaction, along with theoretical analysis and convergence proof;
Constructing Industrial-Scale Optimization Modeling Benchmark
Zhong Li (Great Bay University), Zaiwen Wen (Peking University)
OptimizationLarge Language ModelPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a new benchmark, MIPLIB-NL, with large scale and complete structure, by reverse constructing natural language descriptions and executable code from real industrial MILP instances in MIPLIB 2017.
Content-Style Identification via Differential Independence
Subash Timilsina (Oregon State University), Xiao Fu (Oregon State University)
Image TranslationGenerationDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Propose the Content-Style Differential Independence (CSDI) theory, proving that content and style can be identifiable under unpaired multi-domain data through orthogonal Jacobian subspace constraints, and implement the CSDI-GAN generative model.
Context Forcing: Consistent Autoregressive Video Generation with Long Context
Shuo Chen (University of California Merced), Wenhu Chen (University of Waterloo)
GenerationKnowledge DistillationTransformerDiffusion modelScore-based ModelVideo
🎯 What it does: Propose the Context Forcing framework, which achieves long-term consistent autoregressive video generation by guiding the student model with a long-context teacher.
Context Tuning for In-Context Optimization
Jack Lu (Agentic Learning AI Lab, New York University), Mengye Ren (Agentic Learning AI Lab, New York University)
OptimizationRepresentation LearningMeta LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose the Context Tuning (CT) method, which significantly improves the few-shot adaptation performance of LLMs by directly using a few examples as learnable prompts or key-value prefixes and performing gradient optimization during inference.
Context-Aware Reasoner: Enhancing Contextual Reasoning in Multimodal Large Language Models
Zhe Zheng (Zhejiang University), Weiming Lu (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose Context-Aware Reasoner (CAR), an improved contextual reasoning framework for multimodal large language models, addressing the issue of models ignoring context and relying on surface-level matching.
Context-Driven Incremental Compression for Multi-Turn Dialogue Generation
Yeongseo Jung (Hong Kong University of Science and Technology), Lei Chen (Hong Kong University of Science and Technology)
GenerationRetrievalCompressionTransformerPrompt EngineeringAuto EncoderTextRetrieval-Augmented Generation
🎯 What it does: Propose a context-driven incremental compression framework (C-DIC) that maintains retrievable and updatable compressed memories in a thread-based manner during multi-turn conversations;
Context-free Recognition with Transformers
Selim Jerad (ETH Zürich), William Merrill (Allen Institute for AI)
RecognitionTransformerPrompt EngineeringContrastive LearningText
🎯 What it does: Prove that Transformers with O(log N) cyclic layers and sufficient padding can recognize any context-free language (CFL), and provide the corresponding construction algorithm; give better padding and layer upper bounds for irreducible CFL and linear unambiguous CFL respectively; meanwhile, propose a Transformer architecture that can evaluate variable-free Boolean formulas under log-depth.
Context-level Language Modeling by Learning Predictive Context Embeddings
beiya dai, Zhouhan Lin (Shanghai Jiao Tong University)
GenerationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningText
🎯 What it does: Build the ContextLM framework on top of the standard Transformer, adding an autoregressive context predictor (Context Predictor). It achieves implicit multi-token prediction and guides token generation by generating multi-token level context embeddings between the Token Encoder and Token Decoder.
CONTEXTOR: Contextualized High-order Contrastive Learning
Ze Cai (Hangzhou City University), Jun Wen (Mohamed bin Zayed University of Artificial Intelligence)
Drug DiscoveryGraph Neural NetworkTransformerContrastive LearningBiomedical Data
🎯 What it does: Propose a context-based high-order contrastive learning framework called CONTEXTOR for high-order relationship reasoning
Contextual Rollout Bandits for Reinforcement Learning with Verifiable Rewards
Xiaodong Lu (Beihang University), deqing wang
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmark
🎯 What it does: To address the noise and myopia issues in trajectory sampling within Reinforcement Learning with Verifiable Rewards (RLVR), this paper proposes an adaptive scheduling framework based on contextual bandits (CBS), which dynamically selects high-value trajectories at each training step and efficiently reuses historical trajectories.
Contextual Slate GLM Bandits with Limited Adaptivity
Tanmay Goyal (Microsoft Research India), Gaurav Sinha (Microsoft Research India)
OptimizationTransformerReinforcement LearningPrompt EngineeringContrastive LearningTextTabularBenchmark
🎯 What it does: Proposes an algorithm for solving the contextual slide GLM bandit problem under limited adaptation conditions (batch processing and few switches), achieving a sublinear cumulative regret upper bound that is κ‑free.
Contextualized Privacy Defense for LLM Agents
Yule Wen (Tsinghua University), Diyi Yang (Stanford University)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Propose Contextualized Defense Instructing (CDI) and optimize it using experience-driven reinforcement learning, constructing a unified privacy risk simulation framework;
Contextualized Visual Personalization in Vision-Language Models
Yeongtak Oh (Seoul National University), Sungroh Yoon (Seoul National University)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose the CoViP framework, which uses personalized image descriptions as an intermediate task, achieving contextual personalization of vision-language models through RL fine-tuning and Caption-Augmented Generation;
Continual GUI Agents
Ziwei Liu (Sichuan University), Tao Feng (Tsinghua University)
Domain AdaptationTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision-Language-Action ModelImageMultimodality
🎯 What it does: Propose the "continual GUI agent" task and design the GUI‑AiF framework, enabling GUI agents to perform continual learning under domain migration and resolution changes.
Continual Learning of Domain-Invariant Representations
Pascal Janetzky (LMU Munich), Stefan Feuerriegel (LMU Munich)
Domain AdaptationKnowledge DistillationRepresentation LearningContrastive LearningImageTabular
🎯 What it does: Studies how to improve the model's generalization ability on unseen target domains during continuous learning (CL) by learning domain-invariant representations, and proposes a class of CL algorithms that combine replay, domain-invariant computation, and alignment.
Continual Learning through Control Minimization
Sander de Haan (University of Zurich and ETH Zurich), Benjamin F Grewe
ClassificationComputational EfficiencyRepresentation LearningMeta LearningContrastive LearningImage
🎯 What it does: Reformulate continual learning as a control minimization problem, introducing the Equilibrium Fisher Control (EFC) framework, which utilizes competing retention signals and learning signals from neural activity dynamics to achieve continual learning without replay.
Continual Learning With Participation Privacy: An Auditable Buffering-Aggregation Recipe
T-H. Hubert Chan (University of Hong Kong), Mingxun Zhou (Hong Kong University of Science and Technology)
Federated LearningSafty and PrivacyAgentic AIDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTabular
🎯 What it does: This paper studies an auditable buffer-aggregation process for continuous differential privacy release in federated and streaming learning systems, addressing participant changes with single insertions/deletions (single-edit adjacency).
Continual Model Routing in Evolving Model Hubs
Jack Bell (University of Pisa), Vincenzo Lomonaco (LUISS University)
ClassificationRecommendation SystemFederated LearningComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a model routing framework for continual learning in an expanding AI model hub, defines the Continual Model Routing (CMR) problem, and provides a dedicated benchmark called CMRBench.
Continual Segmentation under Joint Nonstationarity
Prashant Pandey (Indian Institute of Technology), Brejesh Lall (Indian Institute of Technology)
SegmentationDomain AdaptationConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Proposes a continuous semantic segmentation framework called JASCL that simultaneously adapts to changes in classes, domains, and supervision, achieving continuous learning under conditions of limited labeled and abundant unlabeled data.
Continuity-Regularized Flow Matching for Offline Reinforcement Learning
Xiaocong Chen (CSIRO's Data 61), Lina Yao (University of New South Wales)
Reinforcement LearningScore-based ModelFlow-based ModelTabularTime SeriesSequentialBenchmarkStochastic Differential Equation
🎯 What it does: This paper proposes a PQL algorithm, which regularizes the vector field of the flow matching policy using PDE in offline reinforcement learning, controlling its global geometric properties to avoid path distortion caused by value guidance.
Continuous Diffusion Models Can Obey Formal Syntax
Jinwoo Kim (University of California-San Diego), Loris D'Antoni (University of California-San Diego)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelScore-based ModelTextBenchmark
🎯 What it does: Propose a training-free guidance method that enables continuous diffusion language models to satisfy formal syntax constraints defined by regular expressions during generation.
Continuous Variable Hamiltonian Learning at Heisenberg Limit via Displacement-Random Unitary Transformation
Xi Huang (Peking University), Di Luo (Tsinghua University)
Diffusion modelScore-based ModelFlow-based ModelRectified FlowContrastive LearningPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a protocol named Displacement‑Random Unitary Transformation (D‑RUT) for active data acquisition and structured coefficient recovery, which can precisely learn finite-order Hamiltonians of continuous-variable systems under the Heisenberg limit.
Continuous Viewpoint Adaptation for Single View 3D Object Reconstruction
Seunghyun Hwang (Purdue University), Qiang Qiu (Purdue University)
GenerationData SynthesisPose EstimationDepth EstimationDomain AdaptationConvolutional Neural NetworkTransformerDiffusion modelNeural Radiance FieldGaussian SplattingImagePoint CloudMeshOrdinary Differential Equation
🎯 What it does: This paper proposes a continuous view adaptation method for single-view 3D reconstruction, using 3D Gaussian Splatting technology to predict complete 3D objects from a single image;