ICML 2025 Papers — Page 6
International Conference on Machine Learning · 3257 papers
CogReact: A Reinforced Framework to Model Human Cognitive Reaction Modulated by Dynamic Intervention
Songlin Xu (University of California San Diego), Xinyu Zhang (University of California San Diego)
Explainability and InterpretabilityComputational EfficiencyRecurrent Neural NetworkReinforcement LearningDiffusion modelTabularTime SeriesSequential
🎯 What it does: The CogReact framework is proposed, which combines the drift-diffusion model with deep reinforcement learning to simulate the fine effects of dynamic environmental stimuli (such as time pressure) on human cognitive responses.
COKE: Core Kernel for More Efficient Approximation of Kernel Weights in Multiple Kernel Clustering
Weixuan Liang (National University of Defense Technology), En Zhu (National University of Defense Technology)
Computational EfficiencyImage
🎯 What it does: This paper proposes the concept of Core Kernel and designs a Core Kernel construction method based on Singular Value Decomposition (SVD) (SVD-CK), thereby achieving a theoretical guarantee for the large-scale extension of the Multi-Kernel Clustering (MKC) algorithm.
CollabLLM: From Passive Responders to Active Collaborators
Shirley Wu (Stanford University), Jianfeng Gao (Microsoft)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes the COLLABLLM framework, which utilizes collaborative simulation and multi-turn perceptual rewards to train LLMs to actively inquire about user intentions and improve task completion effectiveness during multi-turn interactions.
Collaborative Mean Estimation Among Heterogeneous Strategic Agents: Individual Rationality, Fairness, and Truthful Contribution
Alex Clinton (University of Wisconsin Madison), Kirthevasan Kandasamy (University of Wisconsin Madison)
🎯 What it does: A mechanism is designed that allows heterogeneous cost agents to jointly estimate an unknown vector mean by sharing data in an environment without monetary transactions.
Collapse or Thrive: Perils and Promises of Synthetic Data in a Self-Generating World
Joshua Kazdan (Stanford University), Sanmi Koyejo (Stanford University)
GenerationData SynthesisLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically evaluates the impact of three training workflows—full replacement (replace), accumulation (accumulate), and subsampling with a fixed computational budget based on accumulation (accumulate-subsample)—on model collapse across various generative model tasks (multivariate Gaussian estimation, kernel density estimation, language model fine-tuning, linear regression, and pre-training).
Collapse-Proof Non-Contrastive Self-Supervised Learning
Emanuele Sansone (KU Leuven), Tinne Tuytelaars (KU Leuven)
Representation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: This paper proposes a contrast-free self-supervised learning framework called CPLearn based on hyperdimensional computing. It designs a projector and a loss function that can automatically avoid four common collapse modes (representation collapse, dimension collapse, clustering collapse, and sub-clustering collapse) during the training process, resulting in representations that are both decorrelated and clustered.
CombiMOTS: Combinatorial Multi-Objective Tree Search for Dual-Target Molecule Generation
Thibaud Southiratn (Seoul National University), Sun Kim (Seoul National University)
OptimizationDrug DiscoveryGraph Neural NetworkBiomedical Data
🎯 What it does: A fragment combination method called CombiMOTS based on Pareto Monte Carlo Tree Search (PMCTS) is proposed for generating candidate drugs targeting two objectives simultaneously within the synthetic space.
Combinatorial Reinforcement Learning with Preference Feedback
Joongkyu Lee (Seoul National University), Min-hwan Oh (Seoul National University)
Recommendation SystemReinforcement LearningTabular
🎯 What it does: This paper studies combinatorial reinforcement learning based on Multinomial Logit (MNL) preference feedback and proposes the MNL-VQL algorithm, which maximizes long-term user engagement in scenarios with unknown item values, non-zero outside options, and multi-step state transitions.
Come Together, But Not Right Now: A Progressive Strategy to Boost Low-Rank Adaptation
Zhan Zhuang (Southern University of Science and Technology), Ying Wei (Zhejiang University)
OptimizationTransformerSupervised Fine-TuningDiffusion modelImageTextMultimodality
🎯 What it does: This paper proposes a progressive training strategy named CoTo, aimed at improving the issues of hierarchical gradient imbalance and susceptibility to local optima during the fine-tuning process of LoRA (Low-Rank Adaptation) models.
CoMemo: LVLMs Need Image Context with Image Memory
Shi Liu (Shanghai Artificial Intelligence Laboratory), Jifeng Dai (Tsinghua University)
TransformerVision Language ModelImageText
🎯 What it does: Proposed a dual-path visual processing architecture CoMemo, and introduced RoPE-DHR positional encoding to address the issues of neglected visual information and 'intermediate loss'.
Communicating Activations Between Language Model Agents
Vignav Ramesh (Harvard University), Kenneth Li (Harvard University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A multi-model communication method is proposed that facilitates information exchange within the activation layers of the Transformer, eliminating the high computational cost of natural language transmission and enhancing inference performance.
Commute Graph Neural Networks
Wei Zhuo (Sun Yat-sen University), Xiaoxiao Li (University of British Columbia)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes Commute Graph Neural Networks (CGNN), which utilize commute time between nodes to weight neighbor information in message passing, thereby better capturing asymmetric path dependencies in directed graphs.
CommVQ: Commutative Vector Quantization for KV Cache Compression
Junyan Li (University of Massachusetts), Chuang Gan
CompressionTransformerLarge Language ModelText
🎯 What it does: A KV cache compression method based on Commutable Vector Quantization (CommVQ) is proposed, which can significantly reduce memory requirements during long-context LLM inference.
Compact Matrix Quantum Group Equivariant Neural Networks
Edward Pearce-Crump (Imperial College London)
🎯 What it does: A new compact matrix quantum group equivariant neural network is proposed, addressing the issue that traditional group isomorphic networks cannot model in non-commutative geometric spaces.
Comparing Comparisons: Informative and Easy Human Feedback with Distinguishability Queries
Xuening Feng (Shanghai Jiao Tong University), Yifei Zhu (Shanghai Jiao Tong University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: This paper proposes a new 'discriminative query' that allows humans to first select a more distinguishable pair when comparing two pairs of trajectories, and then provide preference feedback, thereby directly capturing preference intensity.
Comparing Few to Rank Many: Active Human Preference Learning Using Randomized Frank-Wolfe Method
Kiran Koshy Thekumparampil, Branislav Kveton (Adobe Research)
Recommendation SystemOptimizationTabular
🎯 What it does: This paper studies the problem of learning human preferences under limited comparative feedback and proposes an active learning framework that collects the most informative subset through D-optimal design in limited K-way (K≪N) comparisons, thereby efficiently learning the Plackett-Luce (PL) model and ranking N items.
Compelling ReLU Networks to Exhibit Exponentially Many Linear Regions at Initialization and During Training
Max Milkert (Vanderbilt University), Forrest John Laine
Convolutional Neural NetworkImage
🎯 What it does: A reparameterization and pre-training method for ReLU networks is proposed, which maintains exponentially more linear regions during initialization and training, significantly improving approximation accuracy.
Competing Bandits in Matching Markets via Super Stability
Soumya Basu (Google)
Reinforcement Learning
🎯 What it does: A centralized and decentralized bandit learning framework utilizing the concept of super stability in a market with uncertainty on both sides is proposed to achieve the goal of low regret (stable regret).
Competitively Consistent Clustering
Niv Buchbinder (Tel Aviv University), Yue Yang (Rutgers University)
OptimizationTabular
🎯 What it does: This paper proposes a series of online algorithms for maintaining approximately consistent clustering (k-center, facility location, k-median) in fully dynamic environments, aiming to keep the total amount of center additions and deletions (recourse) within a multiplicative factor of the optimal offline solution while maintaining an approximately optimal solution at each time step.
Complete-Tree Space Favors Data-Efficient Link Prediction
Chi Gao (Tsinghua University), Shangqi Guo (Tsinghua University)
Recommendation SystemOptimizationGraph Neural NetworkGraph
🎯 What it does: Proposes Complete Tree Space (CT Space) and Leaf Matching Algorithm to achieve efficient link prediction for sparse graphs.
Complex Wavelet Mutual Information Loss: A Multi-Scale Loss Function for Semantic Segmentation
Renhao Lu (Cornell University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: A multi-scale loss function based on Complex Waveform Mutual Information (CWMI) is proposed to enhance the pixel accuracy and structural consistency of semantic segmentation models.
Componential Prompt-Knowledge Alignment for Domain Incremental Learning
Kunlun Xu (Peking University), Jiahuan Zhou (Peking University)
Domain AdaptationTransformerPrompt EngineeringImage
🎯 What it does: A prompt-based domain incremental learning method called KA-Prompt is proposed to address the knowledge conflict issue caused by inconsistent alignment of prompt components across different domains.
Compositional Causal Reasoning Evaluation in Language Models
Jacqueline R. M. A. Maasch (Cornell Tech), Javier Gonzalez (Microsoft Research)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A combined causal reasoning (CCR) evaluation framework is proposed and implemented, assessing the average treatment effect (ATE) and the reasoning quality of necessary and sufficient probabilities (PNS) across seven language models.
Compositional Condition Question Answering in Tabular Understanding
Jun-Peng Jiang (Nanjing University), Han-Jia Ye (Nanjing University)
RecognitionTransformerLarge Language ModelVision Language ModelMultimodalityTabularBenchmark
🎯 What it does: A new multimodal large language model, COCOTAB, is proposed to specifically address the combinatorial conditional question-answering problem in table images.
Compositional Flows for 3D Molecule and Synthesis Pathway Co-design
Tony Shen (Simon Fraser University), Martin Ester (Simon Fraser University)
OptimizationDrug DiscoveryFlow-based ModelGraphTabular
🎯 What it does: A general generative framework called CGFlow is proposed, which combines GFlowNet and flow matching to jointly generate molecular synthesis pathways and 3D conformations, achieving joint optimization in drug design tasks.
Compositional Generalization via Forced Rendering of Disentangled Latents
Qiyao Liang (Massachusetts Institute of Technology), Ila R Fiete
GenerationData SynthesisConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper explores the failure of combinatorial generalization in a two-dimensional Gaussian 'convex' generation task, even when the input latent representations are completely decoupled, due to remixing in subsequent layers of standard generative models. It proposes achieving data-efficient, zero-shot combinatorial generalization by rendering decoupled latents in pixel space and applying low-rank regularization or single-factor training data (stripes).
Compositional Risk Minimization
Divyat Mahajan (Mila, Université de Montréal), Pascal Vincent (Meta)
Domain AdaptationOptimizationImageText
🎯 What it does: This paper proposes a method called Compositional Risk Minimization (CRM), specifically designed to address the issue of distribution shift when certain attribute combinations are missing in the training set but appear in the test set;
Compositional Scene Understanding through Inverse Generative Modeling
Yanbo Wang (Delft University of Technology), Yilun Du (Harvard University)
RecognitionObject DetectionGenerationDiffusion modelImage
🎯 What it does: A reverse generative modeling framework is proposed to achieve scene understanding through compositional generative models, including object localization, facial attribute inference, and zero-shot multi-object perception.
Compress then Serve: Serving Thousands of LoRA Adapters with Little Overhead
Rickard Brüel Gabrielsson (Massachusetts Institute of Technology), Justin Solomon (Massachusetts Institute of Technology)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a method for compressing LoRA adapters, enabling efficient storage and rapid switching of thousands of LoRAs on GPUs to support multi-task LLM services.
Compressed Image Generation with Denoising Diffusion Codebook Models
Guy Ohayon (Technion Israel Institute of Technology), Michael Elad (Technion Israel Institute of Technology)
RestorationGenerationCompressionDiffusion modelImage
🎯 What it does: This paper proposes a denoising diffusion generative model (DDCM) that changes noise sampling to randomly selecting from a predefined discrete codebook, and utilizes this model to achieve image compression, compression conditional generation (such as image restoration), and generate high-quality images under untrained conditions.
Compressing tree ensembles through Level-wise Optimization and Pruning
Laurens Devos (KU Leuven), Jesse Davis (KU Leuven)
CompressionOptimizationTabular
🎯 What it does: A hierarchical optimization and pruning algorithm named LOP has been developed, which can significantly compress existing decision tree ensembles (reducing the number of leaves) while maintaining prediction accuracy.
Compression via Pre-trained Transformers: A Study on Byte-Level Multimodal Data
David Heurtel-Depeiges (Chandar Research Lab MILA Quebec AI Institute Polytechnique Montreal), Tim Genewein (Google DeepMind)
CompressionTransformerImageTextMultimodalityAudio
🎯 What it does: This study investigates the performance of a small-scale pre-trained decoder-only Transformer in byte-level lossless compression tasks across multimodal data such as text, images, and audio.
Compute Optimal Inference and Provable Amortisation Gap in Sparse Autoencoders
Charles O'Neill (Australian National University), David Klindt (Cold Spring Harbor Laboratory)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelAuto EncoderText
🎯 What it does: This paper proves and empirically demonstrates that sparse autoencoders (SAE) have an unavoidable amortization gap in sparse inference. It compares various encoding strategies such as SAE, MLP, sparse coding, and SAE+ITO by decoupling encoding and decoding, and evaluates them on synthetic data and the activations of GPT-2 Small.
Compute or Load KV Cache? Why Not Both?
Shuowei Jin (University of Michigan), Zhuoqing Mao
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper presents Cake, a system that simultaneously utilizes computation and I/O parallel loading of KV caches, employing a bidirectional scheduling strategy to minimize the latency of the first token.
Computing Optimal Transport Maps and Wasserstein Barycenters Using Conditional Normalizing Flows
Gabriele Visentin (ETH Zurich), Patrick Cheridito (ETH Zurich)
OptimizationFlow-based ModelTabular
🎯 What it does: This paper proposes a method that uses Conditional Normalizing Flow to directly solve the optimal transport (OT) mapping in high-dimensional space and the Wasserstein-2 barycenter, avoiding the complexity of traditional dual or adversarial training.
Computing Voting Rules with Improvement Feedback
Evi Micha (University of Southern California), Vasilis Varsamis (University of Southern California)
🎯 What it does: This paper studies whether it is possible to accurately compute the winners of various voting rules under improved feedback (local improvements made by users on initial candidates). It provides the learnability boundaries for positional scoring rules and Condorcet consistent rules, and validates their performance through experiments.
COMRECGC: Global Graph Counterfactual Explainer through Common Recourse
Gregoire Fournier, Sourav Medya (University of Illinois Chicago)
Explainability and InterpretabilityDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
🎯 What it does: A global counterfactual explanation framework for graph neural networks, COMRECGC, is proposed and implemented, which can generate a set of shared recourses from a large number of rejected samples to achieve accepted outcomes.
Concentration Distribution Learning from Label Distributions
Jiawei Tang (Southeast University), Yuheng Jia (Southeast University)
Data-Centric LearningTabular
🎯 What it does: A concentration distribution learning (CDL) paradigm is proposed, which incorporates background concentration into the label distribution and designs a CDLLD model that can directly learn the concentration distribution from the LDL dataset.
Concept Reachability in Diffusion Models: Beyond Dataset Constraints
Marta Aparicio Rodriguez (Imperial College London), Anastasia Borovykh
GenerationData SynthesisPrompt EngineeringDiffusion modelImage
🎯 What it does: This study investigates the concept accessibility in diffusion models, designing synthetic data experiments to evaluate the impact of scarcity, insufficient descriptions, and bias on accessibility, and demonstrates that steering can maintain accessibility when prompting fails.
Concept-Based Unsupervised Domain Adaptation
Xinyue Xu (Hong Kong University of Science and Technology), Xiaomeng Li (Hong Kong University of Science and Technology)
Domain AdaptationGenerative Adversarial NetworkImage
🎯 What it does: Proposed the Concept-based Unsupervised Domain Adaptation (CUDA) framework to address the interpretability and performance degradation issues of CBM during domain transfer.
Concept-Centric Token Interpretation for Vector-Quantized Generative Models
Tianze Yang (University of Georgia), Ninghao Liu (University of Georgia)
GenerationExplainability and InterpretabilityConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: The CORTEX framework is proposed, utilizing an Information Extractor (IEM) and two explanation methods (sample-level and codebook-level) to provide concept-level explanations for discrete tokens in VQGMs, revealing key tokens in the model generation process, and used for bias detection and directed image editing.
ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features
Alec Helbling (Georgia Tech), Duen Horng Chau (Georgia Tech)
SegmentationGenerationExplainability and InterpretabilityTransformerDiffusion modelImageVideoMultimodality
🎯 What it does: This paper proposes the ConceptAttention method, which utilizes the attention layer of the multimodal diffusion transformer (DiT) to generate high-quality concept saliency maps without additional training.
Concurrent Reinforcement Learning with Aggregated States via Randomized Least Squares Value Iteration
Yan Chen (Duke University), Zhengyuan Zhou (Arena Technologies)
Reinforcement Learning
🎯 What it does: This paper proposes a parallel reinforcement learning framework using Randomized Least Squares Value Iteration (RLSVI) under an aggregated state representation to study how multiple agents can collaboratively explore and achieve low loss.
Conditional Diffusion Model with Nonlinear Data Transformation for Time Series Forecasting
J Rishi (Indian Institute of Science), Deepak Subramani (Indian Institute of Science)
GenerationOptimizationDiffusion modelTime Series
🎯 What it does: A time series forecasting framework named CN-Diff is proposed, which incorporates learnable nonlinear time transformations and conditions during the forward diffusion process, along with a corresponding variational training objective.
Conditioning Diffusions Using Malliavin Calculus
Jakiw Pidstrigach (University of Oxford), Nikolas Nüsken
GenerationData SynthesisOptimizationDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper proposes a general framework based on Malliavin calculus for conditioning diffusion processes in the presence of singular rewards (such as Dirac δ), and provides the corresponding loss function and learning algorithm (BEL).
Confidence Difference Reflects Various Supervised Signals in Confidence-Difference Classification
Yuanchao Dai (Jilin University), Changchun Li (Jilin University)
ClassificationContrastive LearningImageTabular
🎯 What it does: This study proposes a ConfDiff classification method based on consistency risk and consistency regularization (CRCR), which reduces the impact of noisy supervision by grouping confidence differences, thereby enhancing the performance of weakly supervised binary classifiers.
Confidential Guardian: Cryptographically Prohibiting the Abuse of Model Abstention
Stephan Rabanser (University of Toronto), Nicolas Papernot (University of Toronto)
Adversarial AttackImageTabular
🎯 What it does: Identifies and demonstrates the Mirage attack, which deliberately lowers confidence within a specified input range to induce a withdrawal, and proposes the Confidential Guardian audit framework based on zero-knowledge proofs to detect and prevent such attacks.
Conformal Anomaly Detection in Event Sequences
Shuai Zhang (Chinese Academy of Sciences), Shirui Pan (Griffith University)
Anomaly DetectionTime SeriesSequential
🎯 What it does: A continuous-time event sequence anomaly detection method based on quantile inference (conformal inference) called CADES is proposed, which constructs two new nonconformity scores through time recalibration and kernel density estimation, thereby achieving explicit statistical determination of anomalous sequences.
Conformal Prediction as Bayesian Quadrature
Jake C. Snell (Princeton University), Thomas L. Griffiths (Princeton University)
Tabular
🎯 What it does: This paper proposes a distribution-independent uncertainty quantification method based on Bayesian quadratic integration and relates it to traditional conformal prediction methods.
Conformal Prediction with Cellwise Outliers: A Detect-then-Impute Approach
Qian Peng (Nankai University), Changliang Zou (Nankai University)
Anomaly DetectionTabularBiomedical Data
🎯 What it does: A detect-then-impute framework is proposed to construct distribution-free prediction intervals when cell anomalies occur in test samples.
Conformal Tail Risk Control for Large Language Model Alignment
Catherine Chen (Stanford University), Lihua Lei (Stanford University)
GenerationOptimizationTransformerLarge Language ModelText
🎯 What it does: A lightweight calibration framework is constructed, which controls the tail risk of LLM-generated text through an adaptive threshold λ, providing finite sample guarantees without assumptions.
Conformity Score Averaging for Classification
Rui Luo (City University of Hong Kong), Zhixin Zhou (Alpha Benito Research)
ClassificationComputational EfficiencyImage
🎯 What it does: A conformal prediction method is proposed for multi-class classification that performs a weighted average of multiple consistency score functions to enhance the efficiency and information content of the prediction set.
Confounder-Free Continual Learning via Recursive Feature Normalization
Yash Shah (Stanford University), Ehsan Adeli (Stanford University)
TransformerReinforcement LearningImageBiomedical DataAlzheimer's Disease
🎯 What it does: Proposes a Recursive Metadata Normalization (R-MDN) layer to eliminate the influence of covariates on deep network features in continual learning.
ConfPO: Exploiting Policy Model Confidence for Critical Token Selection in Preference Optimization
Hee Suk Yoon (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This work proposes ConfPO, a preference optimization method that utilizes the model's own confidence for key token selection, updating only low-confidence (high information content) tokens, thereby significantly improving human preference consistency in alignment training.
Connecting Thompson Sampling and UCB: Towards More Efficient Trade-offs Between Privacy and Regret
Bingshan Hu (University of British Columbia), Nidhi Hegde (University of Alberta)
OptimizationSafty and PrivacyReinforcement LearningTabular
🎯 What it does: A differential privacy random multi-armed bandit algorithm named DP-TS-UCB is proposed, which combines Thompson Sampling and UCB's exploration mechanism to achieve a tunable trade-off between privacy and reward.
Consensus Based Stochastic Optimal Control
Liyao Lyu (Michigan State University), Jingrun Chen (University of Science and Technology of China)
OptimizationReinforcement LearningTabularTime SeriesStochastic Differential Equation
🎯 What it does: A gradient-free deep reinforcement learning algorithm based on the Consensus-Based Optimization (CBO) framework is proposed to solve high-dimensional, finite-time stochastic optimal control problems.
Consensus Is All You Get: The Role of Attention in Transformers
Álvaro Rodríguez Abella (University of California), Paulo Tabuada (University of California)
TransformerLarge Language ModelText
🎯 What it does: This paper conducts a rigorous mathematical analysis to study the asymptotic behavior of the attention mechanism in Transformers, proving that under various common assumptions, all tokens tend to converge over time (clustering collapse).
Conservative Offline Goal-Conditioned Implicit V-Learning
Kaiqiang Ke (Sun Yat-sen University), Chao Yu (Sun Yat-sen University)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes a conservative offline target conditional implicit value learning algorithm CGCIVL to address the issue of overestimation of state-target pairs across trajectories.
Constant Stepsize Local GD for Logistic Regression: Acceleration by Instability
Michael Crawshaw (George Mason University), Mingrui Liu (George Washington University)
OptimizationTabular
🎯 What it does: This study investigates the convergence of Local Gradient Descent (Local GD) in distributed logistic regression under arbitrary step sizes, proving that it can achieve accelerated convergence without restrictions on step size and communication intervals.
Constrain Alignment with Sparse Autoencoders
Qingyu Yin (Zhejiang University), Linyi Yang (Southern University of Science and Technology)
OptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningAuto EncoderText
🎯 What it does: A feature hierarchical constraint direct preference optimization method (FPO) based on sparse autoencoders is proposed and implemented to align human preferences with large language models while maintaining efficiency and stability.
Constrained Belief Updates Explain Geometric Structures in Transformer Representations
Mateusz Piotrowski (MATS), Adam Shai
TransformerSequential
🎯 What it does: This paper studies how the Transformer achieves constrained Bayesian belief updating in the next word prediction task and reveals the geometric structure of its internal representations.
Constrained Exploitability Descent: An Offline Reinforcement Learning Method for Finding Mixed-Strategy Nash Equilibrium
Runyu Lu (University of Chinese Academy of Sciences), Dongbin Zhao (Chinese Academy of Sciences)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: A novel offline reinforcement learning algorithm named Constrained Exploitability Descent (CED) is proposed to find mixed strategy Nash equilibria under limited datasets.
Constrained Online Convex Optimization with Polyak Feasibility Steps
Spencer Hutchinson (University of California Santa Barbara), Mahnoosh Alizadeh (University of California Santa Barbara)
OptimizationTabular
🎯 What it does: This paper proposes a new online convex optimization method - Polyak Feasibility Steps, aimed at solving online convex optimization problems with fixed convex constraints;
Constrained Pareto Set Identification with Bandit Feedback
Cyrille Kone (University of Lille), Laura Richert (University of Bordeaux)
OptimizationDrug DiscoveryBiomedical Data
🎯 What it does: This paper proposes the problem of identifying the Pareto set under linear constraints in a multi-objective bandit environment and provides a solution with fixed confidence.
Context is Key: A Benchmark for Forecasting with Essential Textual Information
Andrew Robert Williams, Alexandre Drouin (ServiceNow Research)
Large Language ModelPrompt EngineeringTextTime SeriesBenchmark
🎯 What it does: The 'Context is Key (CiK)' benchmark is proposed, specifically to evaluate the ability to utilize necessary natural language context in time series forecasting.
Context Matters: Query-aware Dynamic Long Sequence Modeling of Gigapixel Images
Zhengrui Guo (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
ClassificationComputational EfficiencyTransformerImageBiomedical Data
🎯 What it does: A query-aware dynamic long sequence modeling framework called Querent is proposed for efficient global context attention on gigapixel WSIs.
Context-Informed Neural ODEs Unexpectedly Identify Broken Symmetries: Insights from the Poincaré–Hopf Theorem
In Huh (Purdue University), Muhammad Alam
Time SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: The researchers propose using context-aware neural ordinary differential equations (NODE) to identify and reconstruct bifurcation behaviors that break symmetry, using only symmetric, pre-bifurcation data.
ConText: Driving In-context Learning for Text Removal and Segmentation
Fei Zhang (Shanghai Jiao Tong University), Ya Zhang (Shanghai Jiao Tong University)
RestorationSegmentationPrompt EngineeringImage
🎯 What it does: This paper proposes and implements ConText, a visual context learning framework for simultaneously performing text removal and text segmentation tasks.
Contextual Bandits for Unbounded Context Distributions
Puning Zhao (Shenzhen Campus of Sun Yat-sen University), Tianhang Zheng (Zhejiang University)
Reinforcement LearningImage
🎯 What it does: This paper studies the non-parametric contextual multi-armed bandit problem, providing an optimal asymptotic theoretical lower bound in the case where the context distribution is unbounded and potentially heavy-tailed. It proposes two nearest neighbor-based UCB algorithms: fixed k kNN-UCB and adaptive k Adaptive kNN-UCB, achieving approximately optimal expected cumulative loss.
Contextual Linear Bandits with Delay as Payoff
Mengxiao Zhang (University of Iowa), Haipeng Luo (University of Southern California)
Reinforcement LearningTabular
🎯 What it does: This paper studies the learning problem in contextual linear bandits with delays equivalent to rewards, proposing a phased arm elimination algorithm based on volumetric spanners, and provides a theoretical upper bound on cumulative risk due to delays; it then extends to the case of dynamic action sets using the reduction from context to non-context proposed by Hanna et al.
Contextual Online Decision Making with Infinite-Dimensional Functional Regression
Haichen Hu (Massachusetts Institute of Technology), David Simchi-Levi (Massachusetts Institute of Technology)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a unified framework that utilizes infinite-dimensional function regression to learn the cumulative distribution function (CDF) corresponding to context-action pairs, and based on this, designs a decision-making algorithm for low-frequency calls, achieving optimal or near-optimal performance for multi-class online decision problems (such as contextual gambling, hypothesis testing, and risk control).
Contextual Optimization Under Model Misspecification: A Tractable and Generalizable Approach
Omar Bennouna (Massachusetts Institute of Technology), Asuman E. Ozdaglar
OptimizationTabular
🎯 What it does: A new Integrated Learning and Optimization (ILO) framework is proposed, specifically addressing contextual optimization problems under model misspecification, utilizing manageable surrogate loss functions to tackle the challenges posed by model misspecification.
Contextures: Representations from Contexts
Runtian Zhai (Carnegie Mellon University), Pradeep Kumar Ravikumar
Representation LearningContrastive LearningTabular
🎯 What it does: Proposes the 'contexture theory', unifying various representation learning methods as learning representations through the association between input X and context A, asserting that the optimal representation corresponds to the context-induced expected operator Top-d singular function;
Continual Generalized Category Discovery: Learning and Forgetting from a Bayesian Perspective
Hao Dai (University College London), Jagmohan Chauhan (University College London)
ClassificationRepresentation LearningImage
🎯 What it does: A Bayesian framework for continuous generalized category discovery, VB-CGCD, is proposed to learn new categories in continuous data streams while preventing forgetting.
Continual Reinforcement Learning by Planning with Online World Models
Zichen Liu (Sea AI Lab), Min Lin (Sea AI Lab)
Robotic IntelligenceReinforcement LearningWorld ModelSequentialBenchmark
🎯 What it does: Designed and implemented an online reinforcement learning agent (OA) that learns a sparse shallow world model for online Follow-The-Leader (FTL) and uses planning and model predictive control (MPC) to achieve continuous multi-task learning under unified environmental dynamics.
Continuous Bayesian Model Selection for Multivariate Causal Discovery
Anish Dhir (Imperial College London), Mark van der Wilk (University of Oxford)
Graph Neural NetworkGraphTabular
🎯 What it does: A scalable continuous Bayesian model selection method is proposed for multivariate causal structure learning, utilizing GP-CDE for continuous relaxation of causal graphs and maximizing posterior probability.
Continuous Semi-Implicit Models
Longlin Yu (Peking University), Cheng Zhang (Peking University)
GenerationKnowledge DistillationDiffusion modelImage
🎯 What it does: A continuous semi-implicit model (CoSIM) is proposed, which transforms the hierarchical semi-implicit model into a continuous time framework to achieve rapid distillation and generation of multi-step diffusion models.
Continuous Visual Autoregressive Generation via Score Maximization
Chenze Shao (Tencent Inc), Jie Zhou (Tencent Inc)
GenerationData SynthesisTransformerScore-based ModelImage
🎯 What it does: A continuous visual autoregressive generation framework (Continuous VAR) is proposed, achieving direct generation without quantization through a strictly proper scoring rule (energy score);
Continuous-Time Analysis of Heavy Ball Momentum in Min-Max Games
Yi Feng (Shanghai University of Finance and Economics), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationGenerative Adversarial NetworkTabularOrdinary Differential Equation
🎯 What it does: Conduct a continuous time analysis of the Heavy Ball momentum in two-player games, studying its local convergence and implicit gradient regularization.
Continuously Updating Digital Twins using Large Language Models
Harry Amad (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
TransformerLarge Language ModelPrompt EngineeringBiomedical Data
🎯 What it does: A digital twin framework CALMDT based on large language models is proposed, which can continuously update the model through context learning during inference.
Contour Integration Underlies Human-Like Vision
Ben Lonnqvist (Ecole Polytechnique Federale de Lausanne), Martin Schrimpf (Ecole Polytechnique Federale de Lausanne)
RecognitionLarge Language ModelImage
🎯 What it does: Designed and conducted a large-scale human experiment with 1,038 pre-trained and self-trained deep networks in a contour fragmentation object recognition task, systematically evaluating the models' recognition performance at different levels of fragmentation.
Contract Design Under Approximate Best Responses
Francesco Bacchiocchi (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
Optimization
🎯 What it does: This paper studies the hidden action principal-agent problem where the agent can only execute approximately optimal actions. It proposes a polynomial-time algorithm to solve the optimal robust contract given a tolerance level δ, and presents a no-regret algorithm within an online learning framework.
Contradiction Retrieval via Contrastive Learning with Sparsity
Haike Xu (Massachusetts Institute of Technology), Piotr Indyk (Massachusetts Institute of Technology)
RetrievalContrastive LearningText
🎯 What it does: This paper proposes the SparseCL method for contradiction retrieval, which combines cosine similarity and sparsity measurement to efficiently retrieve and query contradictory sentences in large-scale text databases.
Contrastive Learning with Simplicial Convolutional Networks for Short-Text Classification
Huang Liang, Kelin Xia (Nanyang Technological University)
ClassificationConvolutional Neural NetworkTransformerContrastive LearningText
🎯 What it does: A short text classification framework C-SCN based on simple composite networks and contrastive learning is proposed;
Contrastive Localized Language-Image Pre-Training
Hong-You Chen (Apple), Zhe Gan (Apple)
Object DetectionRetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Based on CLIP, we introduce region-text contrastive loss and a lightweight prompter to perform region-level pre-training on massive amounts of image-text data, enhancing CLIP's localization capabilities, allowing it to directly replace the original CLIP as the visual encoder for multimodal large language models.
Contrastive Private Data Synthesis via Weighted Multi-PLM Fusion
Tianyuan Zou (Tsinghua University), Ya-Qin Zhang
Data SynthesisSafty and PrivacyLarge Language ModelPrompt EngineeringContrastive LearningGaussian SplattingText
🎯 What it does: Utilizing multi-model (PLM) collaboration to generate differential privacy synthetic data, addressing issues of private sample scarcity, synthetic noise, and model selection uncertainty.
Contrastive Visual Data Augmentation
Yu Zhou (University of California Los Angeles), Nanyun Peng (University of Illinois Urbana-Champaign)
ClassificationRecognitionGenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes a Contrastive Visual Data Augmentation (CoDA) method, which first identifies target concepts and easily confused concepts through 0-shot inference of LMM. It then extracts text and visual features from the differences between the two, filtering out distinguishable and generable features. A text-to-image generation model is used to synthesize high-quality images, which are then used to fine-tune the LMM using LoRA or other updating methods, significantly enhancing the model's ability to recognize novel, low-resource, and confusing concepts.
Control and Realism: Best of Both Worlds in Layout-to-Image without Training
Bonan Li (University of Chinese Academy of Sciences), Xinchao Wang (National University of Singapore)
Object DetectionGenerationDiffusion modelImage
🎯 What it does: A training-free layout-to-image generation method called WinWinLay is proposed, which can accurately control object positions while maintaining generation quality.
Controllable Data Generation with Hierarchical Neural Representations
Sheyang Tang (University of Waterloo), Zhou Wang (University of Waterloo)
GenerationData SynthesisMixture of ExpertsDiffusion modelImageMultimodality
🎯 What it does: A generative model based on Hierarchical Implicit Neural Representation (CHINR) is proposed to achieve controllable generation of multimodal data.
Controlled Generation with Equivariant Variational Flow Matching
Floor Eijkelboom (University of Amsterdam), Jan-Willem van de Meent (University of Amsterdam)
GenerationDrug DiscoveryFlow-based ModelGraph
🎯 What it does: In the framework of Variational Flow Matching (VFM), a controllable generative target and equivariant flow matching method is proposed, allowing for the maintenance of structural symmetry while satisfying specific constraints in molecular generation tasks.
Controlling Large Language Model with Latent Action
Chengxing Jia (Nanjing University), Yang Yu (Nanjing University)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This study investigates learning a compact latent action space in large language models and proposes the CoLA framework to enhance RL control and exploration capabilities.
Controlling Neural Collapse Enhances Out-of-Distribution Detection and Transfer Learning
Md Yousuf Harun (Rochester Institute of Technology), Christopher Kanan (University of Rochester)
Domain AdaptationAnomaly DetectionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Under ID training, the internal neural collapse (NC) of the neural network is controlled to simultaneously improve OOD detection and OOD transfer learning performance.
Controlling Underestimation Bias in Constrained Reinforcement Learning for Safe Exploration
Shiqing Gao (Shanghai Jiao Tong University), Xinbing Wang (Shanghai Jiao Tong University)
OptimizationSafty and PrivacyReinforcement LearningSequential
🎯 What it does: The MICE method is proposed to control the underestimation of the cost value function through memory-driven intrinsic cost estimation, thereby reducing constraint violations and enhancing safety.
Convergence Analysis of Policy Gradient Methods with Dynamic Stochasticity
Alessandro Montenegro (Politecnico di Milano), Alberto Maria Metelli (Politecnico di Milano)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper studies the global last iteration convergence of policy gradient methods under dynamic noise levels and proposes a Phase Exploration Scheduling (PES) algorithm;
Convergence of Consistency Model with Multistep Sampling under General Data Assumptions
Yiding Chen (Cornell University), Wen Sun (Cornell University)
Stochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The paper studies the convergence properties of consistency models under multi-step sampling and provides theoretical upper bounds on the Wasserstein and TV distances.
Convergence of Mean-Field Langevin Stochastic Descent-Ascent for Distributional Minimax Optimization
Zhangyi Liu (Tsinghua University), Shuang Li (Chinese University of Hong Kong)
OptimizationAdversarial AttackGenerative Adversarial NetworkStochastic Differential Equation
🎯 What it does: This paper studies the convergence properties of the Mean Field Langevin Stochastic Gradient Descent-Ascent (MFL-SDA) algorithm in distributed minimax optimization.
Convergence of Policy Mirror Descent Beyond Compatible Function Approximation
Uri Sherman (Tel Aviv University), Yishay Mansour (Tel Aviv University)
OptimizationReinforcement Learning
🎯 What it does: In the policy optimization problem of discrete discounted MDPs, this paper proves that Policy Mirror Descent (PMD) can achieve optimal convergence rates for any convex policy class without requiring the complete closure condition; it introduces a new concept of local smoothness and transforms PMD into a non-Euclidean Bregman proximal point algorithm, thereby obtaining an optimal convergence rate of 1/K²⁄³.
Convex Markov Games: A New Frontier for Multi-Agent Reinforcement Learning
Ian Gemp (Google DeepMind), Georgios Piliouras
Reinforcement Learning
🎯 What it does: Proposes a convex Markov game (cMG) framework that supports arbitrary convex preferences for occupancy measures and proves the existence of pure Nash equilibria;
Cooperation of Experts: Fusing Heterogeneous Information with Large Margin
Shuo Wang (University of Electronic Science and Technology of China), zhao kang
ClassificationOptimizationGraph Neural NetworkMixture of ExpertsMultimodalityGraph
🎯 What it does: This paper proposes the Cooperation of Experts (CoE) framework, which integrates multimodal and multi-relational information through a multi-layer heterogeneous network and an expert collaboration mechanism for node/multimodal classification.
Copilot Arena: A Platform for Code LLM Evaluation in the Wild
Wayne Chi (Carnegie Mellon University), Ameet Talwalkar (Carnegie Mellon University)
AI Code AssistantLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Copilot Arena has been built, a VSCode extension that can generate code completions for multiple large language models (LLMs) in a real development environment and instantly collect developers' preferences for two types of completions.