ICML 2024 Papers — Page 5
International Conference on Machine Learning · 2610 papers
Code as Reward: Empowering Reinforcement Learning with VLMs
David Venuto (McGill University), Ankit Anand (Google DeepMind)
Robotic IntelligenceTransformerReinforcement LearningPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Utilize a pre-trained Vision-Language Model (VLM) to generate executable reward function code, construct dense rewards, and accelerate reinforcement learning (RL) training.
Codebook Features: Sparse and Discrete Interpretability for Neural Networks
Alex Tamkin (Anthropic), Noah Goodman
Explainability and InterpretabilityTransformerSupervised Fine-TuningText
🎯 What it does: Proposes and trains sparse discrete codebook features to replace the dense continuous hidden states in neural networks, providing more interpretable intermediate representations;
CodeIt: Self-Improving Language Models with Prioritized Hindsight Replay
Natasha Butt (University of Amsterdam), Taco Cohen (Qualcomm AI Research)
GenerationAI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A self-improving language model called CodeIt is proposed, which iteratively improves the model in program generation using backward replay and prioritized sampling.
CogBench: a large language model walks into a psychology lab
Julian Coda-Forno (Max Planck Institute for Biological Cybernetics), Eric Schulz (Max Planck Institute for Biological Cybernetics)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the CogBench benchmark, which systematically evaluates 40 large language models (LLMs) using ten behavioral metrics derived from seven cognitive psychology experiments, and investigates the impact of factors such as RLHF, model size, code training, and prompt-engineering on LLM behavior.
CogDPM: Diffusion Probabilistic Models via Cognitive Predictive Coding
Kaiyuan Chen (Tsinghua University), Mingsheng Long (Tsinghua University)
GenerationData SynthesisOptimizationDiffusion modelTime SeriesSequential
🎯 What it does: A spatiotemporal prediction method called CogDPM has been developed, which combines the Predictive Coding framework with Diffusion Probabilistic Models, and enhances prediction accuracy through an accuracy weighting mechanism.
COLD-Attack: Jailbreaking LLMs with Stealthiness and Controllability
Xingang Guo (University of Illinois), Bin Hu (University of Illinois)
Adversarial AttackLarge Language ModelText
🎯 What it does: This paper proposes a new framework called COLD-Attack for generating controllable attacks on large language models (LLMs), aimed at enhancing the stealth and controllability of the attacks.
Collaborative Heterogeneous Causal Inference Beyond Meta-analysis
Tianyu Guo (University of California Berkeley), Michael Jordan
Federated LearningTabular
🎯 What it does: A collaborative inverse propensity score weighted estimator (CLB-IPW) is proposed, and causal effect estimation across heterogeneous data centers is achieved through federated learning.
Collaborative Learning with Different Labeling Functions
Yuyang Deng (Pennsylvania State University), Mingda Qiao (University of California)
🎯 What it does: This paper studies different label functions in collaborative PAC learning, providing an upper bound on sample complexity under the (k, ε)-realizability assumption and proposing corresponding learning algorithms.
Collage: Light-Weight Low-Precision Strategy for LLM Training
Tao Yu (Amazon), Luke Huan (Amazon)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes COLLAGE, a low-precision LLM training framework based on Multi-Component Floating-point (MCF), which addresses numerical errors and distortions caused by low precision.
Collapse-Aware Triplet Decoupling for Adversarially Robust Image Retrieval
Qiwei Tian (Xi'an JiaoTong University), Chao Shen (Xi'an JiaoTong University)
RetrievalAdversarial AttackImage
🎯 What it does: We propose CA-TRIDE, a framework for adversarial training that combines collapse awareness and triplet decoupling to enhance the robustness of image retrieval models.
Collective Certified Robustness against Graph Injection Attacks
Yuni Lai (Hong Kong Polytechnic University), Kai Zhou (Hong Kong Polytechnic University)
OptimizationAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: The first collective provable robustness scheme against graph injection attacks is proposed, which transforms the problem into a binary quadratic constrained linear programming (BQCLP) and further linearizes it to obtain a solvable LP, ensuring prediction consistency for multiple nodes on a single perturbed graph.
CoLoRA: Continuous low-rank adaptation for reduced implicit neural modeling of parameterized partial differential equations
Jules Berman (New York University), Benjamin Peherstorfer (New York University)
OptimizationComputational EfficiencyTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: For parameterized partial differential equations, a CoLoRA (Continuous Low-Rank Adaptation) framework is proposed: first, a network is pre-trained using a small number of training trajectories, and then low-rank weights are continuously adapted in real-time to quickly predict the solution field under new physical parameters and initial conditions.
Combinatorial Approximations for Cluster Deletion: Simpler, Faster, and Better
Vicente Balmaseda (Texas A&M University), Nate Veldt (Texas A&M University)
OptimizationGraph
🎯 What it does: Two simpler and faster approximation algorithms for Cluster Deletion with an approximation ratio of 3 are proposed, along with their theoretical proofs.
Combinatorial Multivariant Multi-Armed Bandits with Applications to Episodic Reinforcement Learning and Beyond
Xutong Liu (University of Massachusetts Amherst), Wei Chen (Microsoft Research)
OptimizationReinforcement Learning
🎯 What it does: A new combinatorial multi-armed bandit framework (CMAB-MT) is proposed, and it is applied to episodic reinforcement learning and the probabilistic maximum coverage problem, providing corresponding optimal or near-optimal theoretical upper bounds on returns.
Combining Experimental and Historical Data for Policy Evaluation
Ting Li (Shanghai University of Finance and Economics), Hongtu Zhu (University of North Carolina at Chapel Hill)
Reinforcement LearningTabularBiomedical Data
🎯 What it does: The study investigates how to integrate experimental data with historical control group data to evaluate the average treatment effect of new strategies, and proposes non-pessimistic and pessimistic weighted estimators.
Community-Invariant Graph Contrastive Learning
Shiyin Tan (Tokyo Institute of Technology), Manabu Okumura (Tokyo Institute of Technology)
Representation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A community-invariant graph contrastive learning framework CI-GCL is proposed, achieving learnable graph enhancement while maintaining community structure.
Compact Optimality Verification for Optimization Proxies
Wenbo Chen (Georgia Institute of Technology), Pascal Van Hentenryck (Georgia Institute of Technology)
OptimizationAdversarial AttackTabular
🎯 What it does: A compact optimality verification framework and a projection gradient attack based on value function approximation (PGA-VFA) are proposed, which are applied to large-scale DC-OPF and Knapsack problems, providing a strict upper bound on the optimality of the optimization agent.
Comparing Graph Transformers via Positional Encodings
Mitchell Black (Oregon State University), Yusu Wang (University of California San Diego)
ClassificationGraph Neural NetworkTransformerGraph
🎯 What it does: This paper studies the distinguishing ability of two types of positional encodings in Graph Transformers (GT) — Absolute Positional Encoding (APE) and Relative Positional Encoding (RPE) — and establishes a theoretical framework to prove that they can be converted into each other while maintaining the same distinguishing power on graphs without node features; it also systematically compares the expressive power of different positional encodings.
CompeteAI: Understanding the Competition Dynamics of Large Language Model-based Agents
Qinlin Zhao (University of Science and Technology of China), Xing Xie (Microsoft Research)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: Construct a competition simulation environment based on GPT-4 to study the competitive behavior of restaurant and customer LLM agents.
Completing Visual Objects via Bridging Generation and Segmentation
Xiang Li (Carnegie Mellon University), Zicheng Liu (Microsoft)
RestorationSegmentationGenerationDiffusion modelImage
🎯 What it does: This paper proposes a model called MaskComp, which achieves complete recovery of partially occluded objects through an iterative process of alternating image generation and segmentation.
Complexity Matters: Feature Learning in the Presence of Spurious Correlations
GuanWen Qiu (University of Pennsylvania), Surbhi Goel (University of Pennsylvania)
Data SynthesisRepresentation LearningConvolutional Neural NetworkSupervised Fine-TuningImageTabular
🎯 What it does: By constructing a synthetic dataset based on Boolean functions, this study systematically investigates the feature learning dynamics of neural networks in the presence of spurious correlations, revealing phenomena such as core feature learning being hindered by simple spurious features, the network being separable into core and spurious feature sub-networks, and spurious features being retained even after training, along with theoretical explanations; it also validates that Last Layer Re-training (LLR) can significantly reduce dependence on spurious features and explains its mechanism for enhancing robustness.
Compositional Capabilities of Autoregressive Transformers: A Study on Synthetic, Interpretable Tasks
Rahul Ramesh (University of Pennsylvania), Hidenori Tanaka (NTT Research)
Data SynthesisExplainability and InterpretabilityTransformerPrompt EngineeringTextSequentialChain-of-Thought
🎯 What it does: This study investigates the compositional capabilities of autoregressive Transformers on synthetic interpretability tasks, exploring how the model learns to apply predefined function sequences to input strings and generate correct outputs.
Compositional Curvature Bounds for Deep Neural Networks
Taha Entesari (Johns Hopkins University), Mahyar Fazlyab (Johns Hopkins University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This work studies and proposes a method to compute the upper bound of the second derivative (curvature) of deep differentiable networks, and based on this upper bound, provides closed-form robustness and attack certificates, while using curvature as a differentiable regularization term in training to enhance the network's adversarial robustness.
Compositional Few-Shot Class-Incremental Learning
Yixiong Zou (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a cognitive-inspired compositional method that uses raw image patches as transferable 'primitives' and completes few-shot class incremental learning classification through the similarity of the primitive set; at the same time, a primitive reuse module is designed to enhance cross-class sharing, avoid forgetting, and improve interpretability.
Compositional Image Decomposition with Diffusion Models
Jocelin Su (Massachusetts Institute of Technology), Yilun Du (Massachusetts Institute of Technology)
GenerationData SynthesisDiffusion modelImageMultimodality
🎯 What it does: An unsupervised image decomposition method called Decomp Diffusion is proposed, which decomposes a single image into several combinable components (such as objects, lighting, shadows, etc.), with each component represented by a diffusion model.
Compositional Text-to-Image Generation with Dense Blob Representations
Weili Nie (NVIDIA Corporation), Arash Vahdat (NVIDIA Corporation)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: A text-to-image diffusion model called BlobGEN based on dense Blob representation is proposed, enabling compositional image generation and utilizing LLM for Blob generation.
Compress Clean Signal from Noisy Raw Image: A Self-Supervised Approach
Zhihao Li (Nanyang Technology University), Bihan Wen (Nanyang Technology University)
RestorationCompressionAuto EncoderImage
🎯 What it does: A self-supervised raw image compression framework is proposed, which first extracts and removes noise using a physical noise model, and then compresses the denoised clean signal, storing only the noise-free parts in the bitstream, thereby improving coding efficiency and reconstruction quality.
Compressible Dynamics in Deep Overparameterized Low-Rank Learning & Adaptation
Can Yaras (University of Michigan), Qing Qu (University of Michigan)
CompressionDomain AdaptationOptimizationTransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes a compression learning method by utilizing the dynamic weight learning in deep over-parameterized models to achieve a low-dimensional subspace, significantly reducing computational costs while maintaining the advantages of over-parameterization.
Compressing Large Language Models by Joint Sparsification and Quantization
Jinyang Guo (Beihang University), Xianglong Liu (Beihang University)
CompressionTransformerLarge Language ModelText
🎯 What it does: A Joint Sparsification and Quantization (JSQ) framework is proposed for high compression of large-scale language models, eliminating useless outliers through an activation editor before sparsification.
Compression of Structured Data with Autoencoders: Provable Benefit of Nonlinearities and Depth
Kevin Kögler (Institute of Science and Technology Austria), Marco Mondelli
CompressionAuto EncoderImage
🎯 What it does: Theoretical analysis of shallow autoencoders under one-dimensional compression rates is conducted, proving their limitations in handling sparse structured data, and suggesting that the addition of non-linear and multi-layer decoders can significantly enhance compression performance.
Compute Better Spent: Replacing Dense Layers with Structured Matrices
Shikai Qiu (New York University), Andrew Gordon Wilson (New York University)
OptimizationComputational EfficiencyTransformerLarge Language ModelImageText
🎯 What it does: In large-scale foundational models, various structured matrices (such as Monarch, Block Tensor-Train, Kronecker, low-rank, etc.) are used to replace dense linear layers, and the relationship between their initialization, learning rate, and computation/performance is systematically studied.
Concentration Inequalities for General Functions of Heavy-Tailed Random Variables
Shaojie Li (Renmin University of China), Yong Liu (Renmin University of China)
🎯 What it does: This paper proposes a concentration inequality for heavy-tailed random variable functions, extending the classical bounded difference inequality, applicable to all heavy-tailed distributions with finite variance.
Conditional Common Entropy for Instrumental Variable Testing and Partial Identification
Ziwei Jiang (Purdue University), Murat Kocaoglu (Purdue University)
TabularBiomedical Data
🎯 What it does: This study investigates the validity testing of instrumental variables and the partial identification of causal effects using conditional co-entropy under weak confounding assumptions, and provides an approximation algorithm.
Conditional Language Learning with Context
Xiao Zhang (Tsinghua University), Ji Wu (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By appending context in front of the document during fine-tuning, a context-conditioned language model is learned, enabling selective learning of knowledge.
Conditional Normalizing Flows for Active Learning of Coarse-Grained Molecular Representations
Henrik Schopmans (Karlsruhe Institute of Technology), Pascal Friederich (Karlsruhe Institute of Technology)
OptimizationDrug DiscoveryFlow-based ModelTabular
🎯 What it does: An active learning workflow is constructed in low-dimensional coarse-grained (CG) space, utilizing conditional normalizing flow to train an energy function that generates fine-grained distributions, thereby obtaining potential mean force (PMF) without the need for long molecular dynamics (MD) trajectories to achieve high-quality coarse-grained potentials.
Conditionally-Conjugate Gaussian Process Factor Analysis for Spike Count Data via Data Augmentation
Yididiya Y. Nadew (Iowa State University), Christopher John Quinn
Time SeriesSequentialBiomedical Data
🎯 What it does: A conditional conjugate Gaussian process factor analysis (ccGPFA) model is proposed to handle neural spike count data, avoiding the inference difficulties caused by traditional non-conjugate likelihoods;
Confidence Aware Inverse Constrained Reinforcement Learning
Sriram Ganapathi Subramanian (Vector Institute for Artificial Intelligence), Pascal Poupart (University of Waterloo)
Autonomous DrivingOptimizationReinforcement LearningSequential
🎯 What it does: This study investigates confidence estimation in inverse constraint reinforcement learning and proposes the Confidence-Aware Inverse Constraint Reinforcement Learning (CA-ICRL) algorithm, which can learn constraints that are at least as strict as the true constraints based on a specified confidence level and determine whether the number of expert trajectories is sufficient.
Confidence-aware Contrastive Learning for Selective Classification
Yu-Chang Wu (Nanjing University), Chao Qian (Nanjing University)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: A confidence-weighted contrastive learning method CCL-SC is proposed to explicitly optimize the feature layer of selective classification models, thereby reducing selective risk.
Configurable Mirror Descent: Towards a Unification of Decision Making
Pengdeng Li (Nanyang Technological University), Bo An (Nanyang Technological University)
OptimizationReinforcement Learning from Human FeedbackTabularBenchmark
🎯 What it does: Proposes a Configurable Mirror Descent (CMD) algorithm that uniformly handles single-agent, cooperative, multi-agent zero-sum, general-sum, and mixed decision problems, and constructs a GAMEBENCH benchmark containing 15 academically friendly games.
Conformal Prediction for Deep Classifier via Label Ranking
Jianguo Huang (Southern University of Science and Technology), Hongxin Wei (Southern University of Science and Technology)
ClassificationHyperparameter SearchConvolutional Neural NetworkImage
🎯 What it does: This paper proposes Sorted Adaptive Prediction Sets (SAPS), a non-conformity scoring method in conformal prediction of deep classifiers that retains only the maximum softmax probability while discarding the remaining probability values.
Conformal prediction for multi-dimensional time series by ellipsoidal sets
Chen Xu (Georgia Institute of Technology), Yao Xie (Georgia Institute of Technology)
Anomaly DetectionOptimizationTime SeriesSequential
🎯 What it does: This paper proposes a continuous fractal prediction method for multidimensional time series called MultiDimSPCI, which constructs prediction sets that meet coverage targets using elliptical uncertain regions and sequential conformal inference.
Conformal Prediction Sets Improve Human Decision Making
Jesse C. Cresswell (Layer 6 AI), Noël Vouitsis (Layer 6 AI)
ClassificationImage
🎯 What it does: The study investigated the assistance of conformal prediction sets in human decision-making across three classification tasks.
Conformal Prediction with Learned Features
Shayan Kiyani (University of Pennsylvania), Hamed Hassani (University of Pennsylvania)
ClassificationOptimizationSupervised Fine-TuningImageTabular
🎯 What it does: A partition learning conformal prediction (PLCP) based on learning uncertainty-guided features is proposed to improve the conditional coverage of the conformal prediction set;
Conformal Predictions under Markovian Data
Frédéric Zheng (KTH Royal Institute of Technology), Alexandre Proutiere (KTH Royal Institute of Technology)
TabularTime SeriesSequentialFinance Related
🎯 What it does: This study investigates the split conformal prediction method applied under Markov data, quantifying the coverage gap caused by data correlation, and proposes a new K-split CP method.
Conformal Validity Guarantees Exist for Any Data Distribution (and How to Find Them)
Drew Prinster (Johns Hopkins University), Suchi Saria (Johns Hopkins University)
Drug DiscoveryProtein Structure PredictionTabularBiomedical Data
🎯 What it does: A general theory of Weighted Conformal Prediction is proposed, proving that any joint distribution can achieve valid coverage, and a computable algorithm for Multi-step Feedback Covariate Shift (MFCS) is provided, which is subsequently validated in protein design and active learning experiments.
Conformalized Adaptive Forecasting of Heterogeneous Trajectories
Yanfei Zhou (University of Southern California), Matteo Sesia (University of Southern California)
OptimizationRecurrent Neural NetworkSupervised Fine-TuningTime SeriesSequential
🎯 What it does: A new multi-trajectory time series forecasting method called CAFHT is proposed, which can adaptively generate prediction intervals for trajectories of varying difficulty (heterogeneous) while ensuring finite sample coverage probability.
Conformalized Survival Distributions: A Generic Post-Process to Increase Calibration
Shi-ang Qi (University of Alberta), Russell Greiner (University of Alberta)
Biomedical DataElectronic Health Records
🎯 What it does: A general post-processing framework called CSD is proposed to enhance the calibration of survival models while maintaining unchanged discriminative performance.
Confronting Reward Overoptimization for Diffusion Models: A Perspective of Inductive and Primacy Biases
Ziyi Zhang (Wuhan University), Dacheng Tao (Nanyang Technological University)
OptimizationReinforcement LearningDiffusion modelImage
🎯 What it does: Proposed Temporal Diffusion Policy Optimization (TDPO) and its improved version TDPO-R for reward-based alignment of diffusion models;
Connect Later: Improving Fine-tuning for Robustness with Targeted Augmentations
Helen Qu (University of Pennsylvania), Sang Michael Xie (Stanford University)
Domain AdaptationContrastive LearningTime Series
🎯 What it does: A Fine-Tuning framework called Connect Later is proposed, which uses targeted data augmentation after pre-training to enhance the model's robustness in the target domain.
Connecting the Dots: Collaborative Fine-tuning for Black-Box Vision-Language Models
Zhengbo Wang (University of Science and Technology of China), Tieniu Tan (Institute of Automation, Chinese Academy of Sciences)
ClassificationDomain AdaptationOptimizationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: A collaborative fine-tuning framework called CraFT is proposed for black-box vision-language models (VLM), which can adapt to downstream tasks solely through input prompts and output predictions.
Connecting the Dots: Is Mode-Connectedness the Key to Feasible Sample-Based Inference in Bayesian Neural Networks?
Emanuel Sommer (LMU Munich), David Rügamer (LMU Munich)
Tabular
🎯 What it does: This paper studies sample-based inference (SBI) in Bayesian Neural Networks (BNN), exploring the multimodal structure of the network parameter space and its relationship with over-parameterization and prior information, and proposes a deep ensemble initialization-based MCMC (DEI-MCMC) method to overcome sampling difficulties.
Consistent Adversarially Robust Linear Classification: Non-Parametric Setting
Elvis Dohmatob (Meta)
ClassificationOptimizationAdversarial AttackTabular
🎯 What it does: This paper proposes a consistent estimator for robust linear classification in binary classification tasks under non-parametric conditions, and provides its statistical convergence rate.
Consistent Diffusion Meets Tweedie: Training Exact Ambient Diffusion Models with Noisy Data
Giannis Daras (University of Texas at Austin), Constantinos Costis Daskalakis (Massachusetts Institute of Technology)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A complete framework based on a dual Tweedie formula and consistency loss is proposed, which can learn a diffusion model that is completely consistent with the original distribution using training samples that only contain noise.
Consistent Long-Term Forecasting of Ergodic Dynamical Systems
Vladimir R Kostic, Massimiliano Pontil
Time SeriesSequentialPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a DLI (Deflate-Learn-Inflate) framework based on Koopman operator regression for achieving sustainable long-term predictions.
Consistent Submodular Maximization
Paul Duetting, Morteza Zadimoghaddam (Google Research)
OptimizationGraph
🎯 What it does: A submodular function maximization algorithm that maintains consistency in a streaming environment is proposed, which can change only a limited number of elements with each insertion while maintaining an approximately optimal solution.
Constrained Ensemble Exploration for Unsupervised Skill Discovery
Chenjia Bai (Shanghai Artificial Intelligence Laboratory), Xuelong Li (Shanghai Artificial Intelligence Laboratory)
Robotic IntelligenceReinforcement Learning
🎯 What it does: An unsupervised skill learning framework called CeSD is proposed, which can automatically discover diverse and comprehensive skills without using external rewards.
Constrained Exploration via Reflected Replica Exchange Stochastic Gradient Langevin Dynamics
Haoyang Zheng (Purdue University), Guang Lin (Purdue University)
ImageStochastic Differential Equation
🎯 What it does: A reflection-based replica exchange SGLD (r2SGLD) is proposed for constrained non-convex exploration, addressing the issue of excessive diffusion in high-temperature chains.
Constrained Reinforcement Learning Under Model Mismatch
Zhongchang Sun (University at Buffalo), Shaofeng Zou (University at Buffalo)
OptimizationReinforcement Learning
🎯 What it does: A robust constrained reinforcement learning (RCPO) algorithm is proposed to address the issue of constraint violation caused by model mismatch, ensuring reward improvement and constraint satisfaction in each iteration.
Contamination-Resilient Anomaly Detection via Adversarial Learning on Partially-Observed Normal and Anomalous Data
Wenxi Lv (Sun Yat-sen University), Wenchao Xu (Hong Kong Polytechnic University)
Anomaly DetectionGenerative Adversarial NetworkImage
🎯 What it does: Proposes a GAN-based anomaly detection method that utilizes the learning of constrained normal and abnormal samples in contaminated data.
Context-Guided Diffusion for Out-of-Distribution Molecular and Protein Design
Leo Klarner (University of Oxford), Yee Whye Teh (University of Oxford)
Drug DiscoveryDiffusion modelBiomedical DataStochastic Differential Equation
🎯 What it does: This paper proposes the Context-Guided Diffusion (CGD) method, which incorporates regularization based on unlabeled data and smoothness constraints into the conditional sampling process of diffusion models, significantly enhancing the generalization ability of the guiding model in out-of-distribution (OOD) scenarios, thereby generating better compounds, materials, and protein sequences.
Contextual Feature Selection with Conditional Stochastic Gates
Ram Dyuthi Sristi (University of California San Diego), Hadas Benisty (Technion Israel Institute of Technology)
ClassificationOptimizationSupervised Fine-TuningTabularBiomedical Data
🎯 What it does: A context feature selection method based on Conditional Stochastic Gates (c-STG) is proposed, and it is extended to a weighted version to achieve feature selection and prediction under different contexts (such as age, gender, time, rotation angle, taste, etc.) using a single model.
ConTextual: Evaluating Context-Sensitive Text-Rich Visual Reasoning in Large Multimodal Models
Rohan Wadhawan (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
RecognitionRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper presents a new CONTEXTUAL dataset for evaluating the capabilities of large multimodal models in text-rich image reasoning tasks that require contextual awareness.
Contextualized Policy Recovery: Modeling and Interpreting Medical Decisions with Adaptive Imitation Learning
Jannik Deuschel (Carnegie Mellon University), Eric P. Xing (Carnegie Mellon University)
Explainability and InterpretabilityDrug DiscoveryRecurrent Neural NetworkBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: This paper studies a framework called Contextualized Policy Recovery (CPR) for recovering interpretable and contextualized policies from observational data in medical decision-making.
Continuous Treatment Effects with Surrogate Outcomes
Zhenghao Zeng (Carnegie Mellon University), Edward Kennedy
Tabular
🎯 What it does: In the case of missing primary outcomes with only a small number of observed values, a double robust estimator for continuous treatment effects is proposed using completely observed proxy outcomes and unlabeled samples.
ContPhy: Continuum Physical Concept Learning and Reasoning from Videos
Zhicheng Zheng (Tsinghua University), Chuang Gan (Massachusetts Institute of Technology)
RecognitionObject DetectionExplainability and InterpretabilityLarge Language ModelVideoTextBenchmarkPhysics Related
🎯 What it does: This study investigates the learning and reasoning of continuous physical concepts, proposing the ContPhy dataset and the ContPRO model, aimed at evaluating and enhancing AI's reasoning ability regarding physical common sense in continuous physical scenarios such as soft bodies, liquids, ropes, and fabrics.
Contrasting Multiple Representations with the Multi-Marginal Matching Gap
Zoe Piran (Apple), marco cuturi
Domain AdaptationRepresentation LearningContrastive LearningImageMultimodalityTime Series
🎯 What it does: This paper proposes a novel multi-view/multi-modal representation learning loss—Multiple Boundary Matching Gap (M3G), which globally aligns embeddings of k≥3 views through Multi-Boundary Optimal Transport (MM-OT).
Contrastive Learning for Clinical Outcome Prediction with Partial Data Sources
Meng Xia (Duke University), Ricardo Henao (Duke University)
ClassificationRepresentation LearningContrastive LearningTabularBiomedical DataElectronic Health Records
🎯 What it does: A clinical outcome prediction framework CLOPPS is constructed, which utilizes multi-source electronic medical record data during the training phase and only uses single-source data during the inference phase.
Contrastive Predict-and-Search for Mixed Integer Linear Programs
Taoan Huang (University of Southern California), Bistra Dilkina (University of Southern California)
OptimizationGraph Neural NetworkContrastive LearningTabularBenchmark
🎯 What it does: A hybrid integer linear programming prediction and search framework called ConPaS is proposed, which learns to distinguish and predict high-quality solutions from positive samples (high-quality solutions) and negative samples (infeasible or low-quality solutions), and then uses the prediction results to accelerate the solving process in constraint search.
Contrastive Preference Optimization: Pushing the Boundaries of LLM Performance in Machine Translation
Haoran Xu (Johns Hopkins University), Young Jin Kim (Microsoft)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Fine-tuning medium-scale (7B/13B) large language models for machine translation, a method called Contrastive Preference Optimization (CPO) is proposed, enabling the model to learn to distinguish between excellent and suboptimal translations, thereby improving translation quality.
Contrastive Representation for Data Filtering in Cross-Domain Offline Reinforcement Learning
Xiaoyu Wen (Northwestern Polytechnical University), Zhen Wang (Northwestern Polytechnical University)
Reinforcement LearningContrastive LearningTabular
🎯 What it does: A mutual information (MI) gap measurement method based on contrastive learning is proposed, utilizing the MI gap of source domain data to filter and select the most useful transition data, thereby improving the sample efficiency of cross-domain offline reinforcement learning.
Controllable Prompt Tuning For Balancing Group Distributional Robustness
Hoang Phan (New York University), Qi Lei (New York University)
TransformerPrompt EngineeringImage
🎯 What it does: A Controllable Prompt Tuning (CPT) method is proposed to achieve robust and balanced learning across multiple distributions, avoiding model reliance on shortcut features.
Controlled Decoding from Language Models
Sidharth Mudgal (Google DeepMind), Ahmad Beirami (Google DeepMind)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A modular reasoning control framework called Controlled Decoding (CD) is proposed, which utilizes a prefix scorer to learn the reward value function. It achieves reward-driven control of model outputs through token-wise or block-wise sampling while keeping the underlying language model frozen.
Controlling Behavioral Diversity in Multi-Agent Reinforcement Learning
Matteo Bettini (University of Cambridge), Amanda Prorok (University of Cambridge)
Reinforcement Learning
🎯 What it does: A method named DiCo is proposed, which introduces shared and individual components in the policy network and scales them according to the target diversity ratio, thereby precisely controlling the behavioral diversity of multi-agent systems.
Convergence and Complexity Guarantee for Inexact First-order Riemannian Optimization Algorithms
Yuchen Li (University of Wisconsin Madison), Hanbaek Lyu (University of Wisconsin Madison)
Optimization
🎯 What it does: A general Tangential Block Majorization-Minimization (tBMM) framework is proposed for solving multi-block Riemannian constrained optimization problems, proving convergence to ε-stationary points under approximate subproblem solutions, with an iterative complexity of O(ε⁻²);
Convergence and Trade-Offs in Riemannian Gradient Descent and Riemannian Proximal Point
David Martínez-Rubio (Zuse Institute Berlin), Sebastian Pokutta (Zuse Institute Berlin)
Optimization
🎯 What it does: Analyzed the convergence of Riemannian gradient descent and Riemannian proximal point algorithms, quantified their convergence rates, and proposed different variants to achieve various trade-offs.
Convergence Guarantees for the DeepWalk Embedding on Block Models
Christopher Harker (University of Utah), Aditya Bhaskara (University of Utah)
OptimizationRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: The theoretical analysis of the convergence of low-dimensional deep random walk (DeepWalk) embeddings on stochastic block model (SBM) graphs is conducted, proving that gradient descent can recover community structure with small random initialization.
Convergence of Online Learning Algorithm for a Mixture of Multiple Linear Regressions
Yujing Liu (Academy of Mathematics and Systems Science Chinese Academy of Sciences), Lei Guo (Academy of Mathematics and Systems Science Chinese Academy of Sciences)
Time SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes an online EM algorithm for parameter estimation and data clustering in multi-submodel mixed linear regression (MLR);
Convergence of Some Convex Message Passing Algorithms to a Fixed Point
Vaclav Voracek (University of Tuebingen), Tomas Werner (Czech Technical University in Prague)
Optimization
🎯 What it does: This paper studies the convergence of coordinate descent-based MAP inference algorithms (such as maximum sum diffusion and maximum edge average) in convex message passing, and provides an upper bound on their iterative error. It also proves that under a specific coordinate selection rule, coordinate descent can converge to a fixed point, while another 'midpoint' rule can lead to cycles.
Converting Transformers to Polynomial Form for Secure Inference Over Homomorphic Encryption
Itamar Zimerman (IBM Research), Lior Wolf (Tel Aviv University)
Safty and PrivacyTransformerImageTextFinance Related
🎯 What it does: The first polynomial Transformer for homomorphic encryption is proposed, enabling secure inference on encrypted data.
Convex and Bilevel Optimization for Neural-Symbolic Inference and Learning
Charles Andrew Dickens, Lise Getoor (University of California)
OptimizationTabular
🎯 What it does: This paper proposes a general parameter learning framework for neural-symbolic (NeSy) systems based on convex optimization and bilevel optimization, implemented on NeuPSL, significantly improving inference and learning efficiency.
Convex Relaxations of ReLU Neural Networks Approximate Global Optima in Polynomial Time
Sungyoon Kim (Stanford University), Mert Pilanci (Stanford University)
OptimizationConvolutional Neural NetworkImageTabular
🎯 What it does: This study investigates the optimality gap between two-layer ReLU networks and their convex relaxations, proving that under random Gaussian data, the convex relaxation obtained through random sampling differs from the original problem by only O(√log n), and provides a polynomial-time approximation algorithm along with convergence guarantees for gradient methods.
ConvNet vs Transformer, Supervised vs CLIP: Beyond ImageNet Accuracy
Kirill Vishniakov (MBZUAI), Zhuang Liu (Meta AI Research)
ClassificationRecognitionConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: A series of performance comparisons 'beyond ImageNet' are conducted for two mainstream model architectures, ConvNeXt and ViT, under the two training paradigms of supervised learning and CLIP.
convSeq: Fast and Scalable Method for Detecting Patterns in Spike Data
Roman Koshkin (Okinawa Institute of Science and Technology), Tomoki Fukai (Okinawa Institute of Science and Technology)
Anomaly DetectionOptimizationComputational EfficiencyConvolutional Neural NetworkSpiking Neural NetworkTime SeriesSequential
🎯 What it does: This paper proposes an unsupervised method called convSeq, which utilizes backpropagation to optimize 2D filters for the rapid detection of spatiotemporal patterns in neural spike data.
Cooperative Graph Neural Networks
Ben Finkelshtein (University of Oxford), Ismail Ilkan Ceylan (University of Oxford)
Graph Neural NetworkGraph
🎯 What it does: A cooperative graph neural network (CO-GNN) is proposed, allowing each node to adaptively choose from four behaviors: 'standard (listen + broadcast)', 'listen', 'broadcast', and 'isolate' at each layer, thereby achieving dynamic and asynchronous message passing.
COPAL: Continual Pruning in Large Language Generative Models
Srikanth Malla (Samsung Semiconductor), Chiho Choi (Samsung Semiconductor)
GenerationCompressionDomain AdaptationTransformerLarge Language ModelText
🎯 What it does: The COPAL algorithm is proposed, achieving domain adaptation of large language generation models through continuous pruning without fine-tuning.
Coprocessor Actor Critic: A Model-Based Reinforcement Learning Approach For Adaptive Brain Stimulation
Michelle Pan (University of California Berkeley), Anca Dragan (University of California Berkeley)
Reinforcement Learning
🎯 What it does: Proposes the Coprocessor Actor Critic (CopAC), a model-based reinforcement learning framework for learning closed-loop neural coprocessor stimulation strategies for stroke patients.
Copula-Nested Spectral Kernel Network
Jinyue Tian (Southeast University), Pengfei Fang (Southeast University)
ClassificationTabular
🎯 What it does: A spectral kernel network combined with copula (CokeNet) is constructed to model the spectral density of non-independent features.
Copyright Traps for Large Language Models
Matthieu Meeus (Imperial College London), Yves-Alexandre de Montjoye (Imperial College London)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper proposes injecting specially designed copyright trap sentences into text to achieve document-level membership inference of training data for large language models.
Coresets for Multiple $\ell_p$ Regression
David Woodruff, Taisuke Yasuda (Carnegie Mellon University)
🎯 What it does: Constructed a strong and weak coreset for multi-response ℓp regression, where the size of the core sample is independent of the number of responses m, and provided near-optimal sample complexity; also proposed applications of this method in Euclidean power means and ℓp subspace approximation.
Correcting Diffusion-Based Perceptual Image Compression with Privileged End-to-End Decoder
Yiyang Ma (Peking University), Jiaying Liu (Peking University)
CompressionConvolutional Neural NetworkDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: This paper proposes an image compression method (CorrDiff) that combines diffusion models with a privileged end-to-end decoder, enhancing perceptual quality while maintaining a certain level of distortion by transmitting a small amount of correction information at the decoder end.
Correlation-Induced Label Prior for Semi-Supervised Multi-Label Learning
Biao Liu (Southeast University), Xin Geng (Southeast University)
ClassificationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a semi-supervised multi-label learning framework named PCLP, which utilizes structural causal models to learn the relevant priors of labels and guides the generation of pseudo-labels, thereby improving the model's performance on limited labeled data.
CosPGD: an efficient white-box adversarial attack for pixel-wise prediction tasks
Shashank Agnihotri (University of Mannheim), Margret Keuper (Max-Planck Institute for Informatics)
SegmentationAdversarial AttackImage
🎯 What it does: A new white-box pixel-level adversarial attack called CosPGD is proposed, which uses cosine alignment to balance pixel errors and improve attack efficiency.
Counterfactual Image Editing
Yushu Pan (Columbia University), Elias Bareinboim (Columbia University)
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Proposes an image counterfactual editing framework based on the Augmented Structural Causal Model (ASCM), and provides feasible estimation methods and training algorithms.
Counterfactual Metarules for Local and Global Recourse
Tom Bewley (J.P. Morgan), Manuela Veloso (J.P. Morgan)
OptimizationExplainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: A model-agnostic algorithm T-CREx is proposed, which learns interpretable counterfactual rules and their meta-rules using tree-based surrogate models, providing both individual-level counterfactuals and actionable global solutions.
Counterfactual Reasoning for Multi-Label Image Classification via Patching-Based Training
Ming-Kun Xie (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A multi-label image classification framework based on causal reasoning is proposed, which eliminates the negative mediating effects caused by label co-occurrence through patch-based training and inference.
Covert Malicious Finetuning: Challenges in Safeguarding LLM Adaptation
Danny Halawi (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method called 'Covert Malicious Finetuning' by uploading encrypted training data to the finetuning interface of closed-source LLMs, bypassing detection mechanisms to alter model security and output harmful content.
Craftax: A Lightning-Fast Benchmark for Open-Ended Reinforcement Learning
Michael Matthews (University of Oxford), Jakob Nicolaus Foerster
Recurrent Neural NetworkReinforcement LearningBenchmark
🎯 What it does: A high-performance, open-ended learning benchmark called Craftax has been built based on JAX, and new environments with complex mechanisms including multi-layer dungeons, combat, attributes, potions, etc., have been expanded upon it; at the same time, two evaluation tasks, 1B interactions and 1M interactions, have been defined.
Creative Text-to-Audio Generation via Synthesizer Programming
Manuel Cherep (Massachusetts Institute of Technology), Jessica Shand (Massachusetts Institute of Technology)
GenerationOptimizationContrastive LearningTextAudio
🎯 What it does: This paper proposes a text-to-audio generation method based on a virtual modular synthesizer (with only 78 interpretable parameters), achieving the mapping from text prompts to high-quality abstract audio through iterative parameter optimization.
Criterion Collapse and Loss Distribution Control
Matthew J. Holland (SANKEN Osaka University)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the phenomenon of 'criterion collapse' among different learning criteria in machine learning, analyzing whether optimizing a certain criterion necessarily leads to optimality in error probability.
Critical feature learning in deep neural networks
Kirsten Fischer (Jülich Research Centre), Moritz Helias (RWTH Aachen University)
Representation LearningTabular
🎯 What it does: This paper derives the feature learning theory of finite-width deep networks from a Bayesian perspective and splits the network prior into a Gaussian process superposition, obtaining the forward-backward propagation equations for the maximum posterior kernel.