ICML 2026 Papers — Page 9
International Conference on Machine Learning · 6554 papers
CalM: A Self-Supervised Foundation Model for Population Dynamics in Calcium Imaging Data
Xinhong Xu (Tsinghua University), Yuanlong Zhang (Tsinghua University)
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerVision-Language-Action ModelDiffusion modelAuto EncoderContrastive LearningTime SeriesSequentialBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission TomographyUltrasoundDiffusion Tensor Imaging
🎯 What it does: Built and pre-trained CalM — a self-supervised Transformer base model, for neuropopulation dynamics modeling and behavioral decoding of large-scale multi-animal, multi-session calcium imaging data.
CalPro: Prior-Aware Evidential Conformal Prediction with Structure-Aware Sensitivity Bounds for Protein Structures
Ibne Farabi Shihab (Iowa State University), Anuj Sharma (Iowa State University)
Protein Structure PredictionGraph Neural NetworkGraphBiomedical Data
🎯 What it does: Designed CalPro, a post-layer combining a geometry-aware evidence head based on graph neural networks, a differentiable calibration surrogate, and split-conformal methods, to achieve confidence interval generation and calibration for protein structure prediction;
CAMEL: Confidence-Gated Reflection for Reward Modeling
Zirui Zhu (National University of Singapore), Yang You (National University of Singapore)
Explainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose the CAMEL framework, which implements confidence-gated reflection for reward models, first making single-token decisions and then generating reflective text based on confidence levels.
CameraNoise: Enabling Faithful Camera Control in Video Diffusion through Geometry-Flow-Guided Noise Warping
Haoyu Zhao (University of Fudan), Yu-Gang Jiang (University of Fudan)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelOptical FlowVideoStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposed the CameraNoise method, which directly embeds camera pose information into the noise space, achieving camera-controllable generation in video diffusion.
CamGeo: Sparse Camera-Conditioned Image-to-Video Generation with 3D Geometry Priors
Xuanyi Liu (Peking University), Siwei Ma (Peking University)
GenerationData SynthesisPose EstimationDepth EstimationKnowledge DistillationConvolutional Neural NetworkTransformerDiffusion modelContrastive LearningOptical FlowImageVideoPoint Cloud
🎯 What it does: Propose the CamGeo framework, which achieves high-quality image-to-video generation under sparse camera conditions through 3D geometric prior distillation during the training phase.
CAMP: Coherent Alignment of Multimodal Prototypes for Explainable Complementary Learning
Alvaro Lopez Pellicer (JPMorgan AI Research), Saumitra Mishra (JPMorgan AI Research)
ClassificationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringMixture of ExpertsAuto EncoderContrastive LearningImageTextMultimodalityTabular
🎯 What it does: Proposes the CAMP framework, achieving prototype learning and interpretable case reasoning for the complementary modality classification (CMC) task, capable of inference under any subset of modalities.
Can Adaptive Gradient Methods Converge under Heavy-Tailed Noise? A Case Study of AdaGrad
Zijian Liu (New York University)
Optimization
🎯 What it does: Theoretically prove the convergence of AdaGrad under heavy-tailed noise (tail index p ∈ (4/3, 2]), and provide a faster convergence rate for AdaGrad Norm-
Can Agents Generalize to the Open World? Unveiling the Fragility of Static Training in Tool Use
Weiming Wu (Nanjing University), Lan-Zhe Guo (Nanjing University)
Domain AdaptationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper systematically evaluates the robustness of LLM tool usage agents under open-source environments (query, tool, interaction, domain distribution drift) in a controlled sandbox environment, and proposes the Perturbation-Augmented Fine-Tuning (PAFT) method to enhance open-world adaptability.
Can Computational Reducibility Lead to Transferable Models for Graph Combinatorial Optimization?
Semih Cantürk (Universite de Montreal), Guy Wolf (Universite de Montreal)
OptimizationGraph Neural NetworkTransformerSupervised Fine-TuningContrastive LearningGraph
🎯 What it does: This paper proposes a graph neural network encoder based on GCON, and trains the model using an energy-based unsupervised loss function to solve various graph combinatorial optimization tasks. Subsequently, through a pre-training and fine-tuning strategy based on computational reducibility, the paper achieves knowledge transfer and multi-task learning across tasks.
Can I Have Your Order? Monte-Carlo Tree Search for Slot Filling Ordering in Diffusion Language Models
Joshua Ong Jun Leang (Imperial College London), Eleonora Giunchiglia (Imperial College London)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringDiffusion modelScore-based ModelTextSequentialBenchmark
🎯 What it does: Propose MCDIFFUSE, an untrained slot-filling order planning method in Masked Diffusion Models (MDMs) using MCTS, improving the plan-and-infill decoding process.
Can Large Language Models Generalize Procedures Across Representations?
Fangru Lin (University of Oxford), Janet B. Pierrehumbert (University of Oxford)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTextGraphPhysics Related
🎯 What it does: Study the program migration capability of LLMs under three equivalent representations: code, graphs, and natural language
Can LLM Agents Stick to the Script? Modeling Commitment in Interactive Narratives
Yingpeng Ma (University of Macau), Derek F. Wong (University of Macau)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Propose the Narrative Commitment Preservation (NCP) task and the NCP-Bench benchmark to evaluate the logical consistency of large language models in interactive storytelling.
Can LLMs Reason Like Automated Theorem Provers for Rust Verification? VCoT-Bench: Evaluating via Verification Chain of Thought
Zichen Xie (University of Virginia), Wenxi Wang (University of Virginia)
Explainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study proposes the VCoT-Lift framework and the VCoT-Bench benchmark to elevate the low-level reasoning of SMT solvers into readable verification chains-of-thought, and evaluates the reasoning capabilities of LLMs in the verification of Rust programs.
Can LLMs Reason Structurally? Benchmarking via the lens of Data Structures
Yu He (Stanford University), Ellen Vitercik (Stanford University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed a structure reasoning benchmark called DSR-Bench oriented towards data structures, systematically evaluating the ability of LLMs in structural reasoning, and conducting multi-dimensional evaluations on 13 mainstream LLMs.
Can local learning match self-supervised backpropagation?
Wu S. Zihan (EPFL), Guillaume Bellec (TU Wien)
ClassificationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes a theoretical framework and improves the local self-supervised learning algorithm, making it capable of approximating global backpropagation in deep networks.
Can Microcanonical Langevin Dynamics Leverage Mini-Batch Gradient Noise?
Emanuel Sommer (LMU Munich), David Rügamer (LMU Munich)
OptimizationComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelContrastive LearningImageTextTabularStochastic Differential Equation
🎯 What it does: Proposed a microcanonical Langevin dynamics sampler (SMILE) that can efficiently operate under small-batch gradient noise, and further improved its robustness through gradient noise preprocessing and adaptive scheduling based on energy variance.
Can Muon Fine-tune Adam-Pretrained Models?
Xingyu Qu (MBZUAI), Samuel Horváth
OptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningImageTextMultimodality
🎯 What it does: The study investigates the mismatch problem between the Mu optimizer and Adam during fine-tuning after pre-training, and proposes using LoRA to limit the update magnitude to alleviate this issue.
Can Recommender Systems Teach Themselves? A Recursive Self-Improving Framework with Fidelity Control
Luankang Zhang (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Recommendation SystemTransformerReinforcement LearningAuto EncoderGenerative Adversarial NetworkContrastive LearningTabularSequential
🎯 What it does: Proposes a recursive self-improvement recommendation framework, RSIR, which allows the model to continuously enhance the training set and improve performance by generating high-fidelity interaction sequences without relying on external data or teacher models.
Can Vision Language Models Learn Intuitive Physics from Interaction?
Luca M. Schulze Buschoff (Helmholtz Munich), Eric Schulz (Helmholtz Munich)
Computational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodalityPhysics Related
🎯 What it does: Evaluate the performance and generalization ability of visual language models on intuitive physics reasoning tasks through interactive reinforcement learning and supervised fine-tuning (SFT)
Can VLMs Diagnose and Recover from VLA Manipulation Faults?
Bowen Yan (Shanghai AI Laboratory), Guangtao Zhai (Shanghai AI Laboratory)
Anomaly DetectionRobotic IntelligenceTransformerPrompt EngineeringVision Language ModelVision-Language-Action ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Constructed the VLA-FixBench benchmark, proposed the FaultEval evaluation framework, and designed a VLM-VLA joint closed-loop diagnosis and recovery mechanism for fault diagnosis and recovery in robotic manipulation tasks.
Can We Build a Monolithic Model for Fake Image Detection? SICA: Semantic-Induced Constrained Adaptation for Unified-Yet-Discriminative Artifact Feature Space Reconstruction
Bo Du (Sichuan University), Ji-Zhe Zhou (Sichuan University)
Anomaly DetectionTransformerContrastive LearningImage
🎯 What it does: This paper proposes a single model for detecting forged images across four image forgery subdomains (Deepfake, AIGC, IMDL, Doc);
CANDI: Hybrid Discrete-Continuous Diffusion Models
Patrick Pynadath (Purdue University), Ruqi Zhang (Purdue University)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelTextMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes CANDI, a hybrid model that integrates continuous and discrete diffusion, addressing the performance bottleneck of continuous diffusion on discrete data
CAOS: Conformal Aggregation of One-Shot Predictors
Maja Waldron (University of WisconsinMadison)
ClassificationExplainability and InterpretabilityComputational EfficiencyMeta LearningTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: Propose a conformal prediction framework called CAOS for one-shot prediction, which can aggregate multiple one-shot predictors and achieve effective uncertainty quantification without splitting the data.
CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation
Letian Fu (NVIDIA), Linxi Fan
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsVision-Language-Action ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Propose the CaP-X framework, which includes the CaP-Gym interactive environment, the CaP-Bench benchmark, the CaP-Agent0 untrained agent, and the CaP-RL reinforcement learning module, systematically evaluating and enhancing code-based robotic control models.
Capability Traps in DPO
Marco Pollanen (Trent University)
OptimizationExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Systematically sweep the β parameter in Direct Preference Optimization (DPO) to explore its impact on different capabilities (reasoning, arithmetic, formatting, empathy, etc.), revealing traps such as the decoupling of proxy metrics from true capabilities, path dependence, and probe cluster structures.
Capacitated Fair-Range Clustering: Hardness and Approximation Algorithms
Ameet Gadekar (CISPA Hemholtz Center for Information Security), Suhas Thejaswi (Aalto University)
OptimizationTabularBenchmark
🎯 What it does: The study investigates and proposes a clustering problem that combines capacity constraints with fair range (Capacitated Fair-Range k-Clustering), and provides its theoretical complexity and approximation algorithms.
Capacity without Access: Reinterpreting the Mid-Depth Spectral Plateau in LLMs
SeongMin Kang, Yoon-Sik Cho (Chung-Ang University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringDiffusion modelScore-based ModelText
🎯 What it does: This paper analyzes the residual dynamics and hidden state geometry of the Transformer, proposing a perspective that separates 'representation capacity' from 'accessibility', and designs a Deep Perceptual Diagnostic Intervention (DDI) to enhance the accessibility of intermediate layer information to the final prediction.
Capacity-Agnostic Parameter Isolation for Continual Graph Learning
Ye Xiao (University of Sydney), Junbin Gao (University of Sydney)
ClassificationFederated LearningComputational EfficiencyKnowledge DistillationRepresentation LearningGraph Neural NetworkAuto EncoderContrastive LearningGraphTabular
🎯 What it does: Propose a capacity-agnostic parameter isolation framework called CAGNN based on the dendritic neuron mechanism for continual graph learning;
Capacity-Aware Mixture Law Enables Efficient LLM Data Optimization
Jingwei Li (Tsinghua University), Jingzhao Zhang (Tsinghua University)
OptimizationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: Propose CAMEL — a capacity-aware hybrid law, used to efficiently search for data mixing ratios in LLM pre-training and directly predict benchmark performance.
CapBencher: Give Your LLM Benchmark a Built-in Alarm for Test-Set Overfitting
Takashi Ishida (RIKEN), Ikko Yamane (University of Rennes)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose CapBencher, which uses randomized answers to reduce Bayesian accuracy, allowing statistical detection of whether models overfit to the test set when the benchmark is publicly released;
CAPTCHA Solving for Native GUI Agents: Automated Reasoning-Action Data Generation and Self-Corrective Training
Yuxi Chen (University of Illinois Urbana Champaign), Huan Zhang (University of Illinois Urbana Champaign)
Autonomous DrivingRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringVision Language ModelVision-Language-Action ModelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes ReCAP, an agent capable of directly solving modern interactive CAPTCHA in native GUI environments. Based on this, seven types of dynamic CAPTCHA systems, automated data generation, and self-correcting training processes were built, ultimately achieving a significant improvement in CAPTCHA cracking success rates while maintaining general GUI task performance.
Capturing Gaze Shifts for Guidance: Cross-Modal Fusion Enhancement for VLM Hallucination Mitigation
Zheng Qi (AWS AI Labs), Nikolaos Pappas (Oracle AI)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the GIFT method, which generates task-related global saliency maps during the VLM inference phase by tracking the visual attention 'gaze shift' of information-rich query words. During decoding, it simultaneously amplifies both visual and query attention to maintain cross-modal fusion balance, thereby reducing hallucinations.
Caracal: Causal Architecture via Spectral Mixing
BINGZHENG GAN, Tao Yu (Huawei Technologies Co., Ltd.)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Propose the Caracal architecture, replacing global attention with a multi-head Fourier module with O(L log L) complexity, achieving scalable long-sequence modeling.
CARD: Coarse-to-fine Autoregressive Modeling with Radix-based Decomposition for Transferable Free Energy Estimation
Ziyang Yu (ByteDance Seed), Yang Liu (Tsinghua University)
Drug DiscoveryTransformerDiffusion modelScore-based ModelContrastive LearningBiomedical Data
🎯 What it does: Developed a self-attention generative framework called CARD for directly calculating the absolute free energy difference of any molecular system, based on the theoretical foundation of zero free energy proposal;
Cardio-mmFlow: A Gaussian-Prior-Free Physics-Informed Flow Matching Framework for Electrocardiogram to mmWave Radar Synthesis
Ziyang LIU, Kaishun Wu (Hong Kong University of Science and Technology (Guangzhou))
GenerationData SynthesisTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningBiomedical DataElectrocardiogram
🎯 What it does: Propose Cardio-mmFlow, a physics-informed flow matching framework without Gaussian prior, used to synthesize millimeter-wave radar signals from clinical ECG, enabling non-contact electrocardiogram monitoring.
CARE: Adaptive Calibration for Reliable Recommendations
Nitin Bisht (University of Technology), Guandong Xu (Education University)
Recommendation SystemContrastive LearningTabularSequential
🎯 What it does: Propose the CARE framework, which adds an adaptive calibration layer to existing offline-trained recommendation models, generating recommendation sets with variable sizes that have finite-sample performance guarantees in non-stationary user behavior streams.
CARE: Class-Adaptive Expert Consensus for Reliable Learning with Long-Tailed Noisy Labels
Mengke Li (Shenzhen University), Hui Huang (Shenzhen University)
ClassificationTransformerMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed a Class-Adaptive Rectification with Experts (CARE) framework, which uses multi-modal experts (text, image, original labels) to adaptively correct long-tailed noisy labels through a class-adaptive Top-K consensus mechanism and achieve long-tail calibration;
CARE: Confounder-Aware Aggregation for Reliable LLM Evaluation
Jitian Zhao (University of Wisconsin-Madison), Frederic Sala (University of Wisconsin-Madison)
Data-Centric LearningGraph Neural NetworkTransformerLarge Language ModelTextGraphBenchmark
🎯 What it does: Propose the CARE framework, which explicitly models the shared confounding factors between LLM and reviews, and recovers the true quality signals under unsupervised conditions to improve the aggregation quality of multiple LLM reviews.
CAReDiO: Enhancing Cultural Alignment of LLM via Representativeness and Distinctiveness Guided Data Optimization
Jing Yao (Renmin University of China), Xing Xie (Microsoft Research Asia)
OptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Automatically generates culturally aligned data that is both representative and discriminative through the information theory-driven iterative framework CAReDiO.
Cascaded Flow Matching for Heterogeneous Tabular Data with Mixed-Type Features
Markus Mueller (Erasmus University Rotterdam), Dennis Fok (Erasmus University Rotterdam)
GenerationData SynthesisAnomaly DetectionFederated LearningSafty and PrivacyTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowTabularElectronic Health RecordsFinance Related
🎯 What it does: Propose a cascade flow matching framework called TabCascade, which first generates tabular data using low-resolution representations (categories + discretized numerical values), and then builds upon this foundation by using high-resolution flow matching to generate numerical details, thus achieving high-quality synthesis of hybrid features (such as missing values, overflow values).
CAST: Modeling Visual State Transitions for Consistent Video Retrieval
Yanqing Liu (Google), Yan jiao
RetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningVideoTextRetrieval-Augmented Generation
🎯 What it does: Proposes the Consistent Video Retrieval task and the CAST model to address the context and identity consistency issues in video retrieval.
CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMs
Shigeng Wang (Intel Labs China), Anbang Yao (Intel Labs China)
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a post-training ternary quantization method called CAT-Q, which can compress large language models into 1.58-bit weight quantized models without the need for large-scale training data or retraining.
CATArena: Evaluating Evolutionary Capabilities of Code Agents via Iterative Tournaments
Lingyue Fu (Shanghai Jiao Tong University), Yong Yu (Shanghai Jiao Tong University)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the CATArena framework, which evaluates the evolutionary capabilities of code agents through multi-round tournaments and iterative feedback.
Catch-22: On the Fundamental Tradeoff Between Detectability and Robustness in LLM Watermarking
Kuheli Pratihar (Indian Institute of Technology Kharagpur), Debdeep Mukhopadhyay (Indian Institute of Technology Kharagpur)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes an information-theoretic framework to quantify the trade-off between the detectability and robustness of LLM watermarks, and based on this framework, designs a hybrid watermark selection rule with minimal information content.
Categorical Flow Maps
Daan Roos (University of Amsterdam), Jan-Willem van de Meent (University of Amsterdam)
GenerationKnowledge DistillationDrug DiscoveryGraph Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelContrastive LearningImageTextGraphStochastic Differential Equation
🎯 What it does: This study proposes Categorical Flow Maps (CFM), a flow matching method based on endpoint prediction, which can achieve fast generation (in one or two steps) on discrete data (images, molecular graphs, text), and supports self-distillation and test-time guidance;
Categorical Reparameterization with Denoising Diffusion Models
Samson Gourevitch (Ecole polytechnique), Yazid Janati (Mohamed Bin Zayed University of AI)
OptimizationDiffusion modelScore-based ModelImageTextTabular
🎯 What it does: Proposed a discrete reparameterization method called REDGE based on denoising diffusion models for gradient estimation under discrete distributions.
CatFlow: Co-generation of Slab-Adsorbate Systems via Flow Matching
Minkyu Kim (Korea Advanced Institute of Science and Technology), Sungsoo Ahn (Korea Advanced Institute of Science and Technology)
GenerationDrug DiscoveryTransformerDiffusion modelFlow-based ModelGraphBenchmark
🎯 What it does: This paper proposes a CATFLOW framework based on flow matching, which jointly generates catalyst surface profiles and adsorbate structures;
CauchyNet: Compact and Data-Efficient Learning using Holomorphic Activation Functions
Hong-Kun Zhang (Great Bay University), Zhihong Jeff Xia
Computational EfficiencyData-Centric LearningAuto EncoderContrastive LearningOptical FlowTabularTime SeriesSequentialBenchmarkStochastic Differential Equation
🎯 What it does: Propose a single-hidden-layer complex-valued network called CauchyNet, which utilizes a multivariate Cauchy kernel activation function, specifically designed for low-data, high-peak regression tasks such as near-singular peaks and missing value interpolation.
Causal Attention with Lookahead Keys
Zhuoqing Song (Princeton University), Quanquan Gu (ByteDance Seed)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed CASTLE (Causal Attention with Lookahead Keys) mechanism in autoregressive language models, which allows keys to be updated incrementally with context while maintaining the autoregressive structure, and provided efficient parallel training and inference algorithms.
Causal Dependency-Aware Unsupervised Routing for Large Reasoning Models
Jiacheng Liu (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningTextBenchmarkChain-of-Thought
🎯 What it does: Developed an unsupervised router called ReasoningRouter to address the causal dependency and length imbalance between thinking and answers in the outputs of large-scale reasoning models.
Causal Detection of Multi-Step LLM Agent Attacks
Viraaji Mothukuri (Kennesaw State University), Reza M. Parizi (Kennesaw State University)
Anomaly DetectionExplainability and InterpretabilityAdversarial AttackGraph Neural NetworkTransformerLarge Language ModelTextGraphBenchmark
🎯 What it does: Propose the CausalTrace framework, which utilizes causal inference to detect multi-step injection attacks in LLM agents and can restore the complete attack chain during post-audit.
Causal Direct Preference Optimization for Distributionally Robust Generative Recommendation
Chu Zhao (Northeastern University), Guibing Guo (Northeastern University)
Domain AdaptationRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningTabular
🎯 What it does: This paper proposes CausalDPO, a direct preference optimization framework that introduces causal invariance learning into LLM recommendation systems, aiming to eliminate the interference of environmental confounding factors and enhance the model's generalization ability on out-of-distribution data.
Causal Discovery for Irregularly Time Series with Consistency Guarantees
Weihong Li (Zhejiang University), Kun Kuang (Zhejiang University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTabularTime Series
🎯 What it does: Proposed a joint missing value imputation and causal structure learning framework called ReTimeCausal based on the EM algorithm, which is used to recover interpretable lagged causal graphs from irregularly sampled time series with high missing rates.
Causal discovery for time series with endogenous context variables
Oana-Iuliana Popescu (University of Potsdam), Jakob Runge (University of Potsdam)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTime SeriesPhysics Related
🎯 What it does: This paper addresses the problem of causal discovery in time series with endogenous context variables, proposing two adaptive testing algorithms based on PCMCI+ (PAC-PCMCI+ and SAC-PCMCI+), which can recover context-specific causal graphs without assuming context exogeneity.
Causal Disentangled Anchor Learning for Scalable Fair Multi-view Clustering
Suyuan Liu (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
Federated LearningSafty and PrivacyComputational EfficiencyRepresentation LearningAuto EncoderContrastive LearningImageTabularBenchmark
🎯 What it does: Propose an expandable fair multi-view clustering framework CDAL, which physically separates the semantic subspace from the sensitive subspace through a dual anchor point mechanism, and ensures statistical independence between the two by utilizing the linearized HSIC constraint, thus achieving fair clustering.
Causal Effect Identifiability in the Presence of Latent Confounders Without Auxiliary Variables
Xiu-Chuan Li (University of Sydney), Tongliang Liu (University of Sydney)
OptimizationExplainability and InterpretabilityComputational EfficiencyTabularReview/Survey Paper
🎯 What it does: This paper conducts a complete identifiability analysis on whether the joint distribution of the outcome variable and intervention variable can uniquely determine the causal effect under the condition of no auxiliary variables in linear structural causal models (SCM);
Causal Feature Learning via Generalized Rayleigh Quotients
Liang Cao (University of British Columbia), Weide Liu (Jiangxi University of Finance and Economics)
Anomaly DetectionRepresentation LearningAuto EncoderContrastive LearningTime Series
🎯 What it does: By constructing time-delayed features and using a generalized Rayleigh quotient optimization, the method jointly maximizes the correlation with the target variable and the invariance of environmental mean, extracting causal features from time series.
Causal Fine-Tuning under Latent Confounded Shift
Jialin Yu (University of Oxford), Ricardo Silva (University College London)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose a Causal Fine-Tuning (CFT) framework that decomposes representations into stable causal components and environment-sensitive bias components through fine-tuning guided by structural causal models, thereby achieving robust prediction under distribution shifts with implicit confounding.
Causal Flow Q-Learning for Robust Offline Reinforcement Learning
Mingxuan Li (Columbia University), Elias Bareinboim (Columbia University)
Reinforcement LearningFlow-based ModelImage
🎯 What it does: Propose a causal inference-based offline reinforcement learning objective, combine it with flow matching methods to learn robust continuous action policies, and provide the corresponding CFQL algorithm.
Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation
Hongzhou Zhu (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelScore-based ModelVideoOrdinary Differential Equation
🎯 What it does: Propose a new autoregressive video diffusion model distillation process called Causal Forcing. First, train an autoregressive teacher model using teacher forcing. Then, perform ODE distillation with this teacher to ensure frame-level injectivity. Finally, further improve the model performance through asymmetric DMD.
Causal Identification from Counterfactual Data: Completeness and Bounding Results
Arvind Raghavan (Columbia University)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningMeta LearningDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTabularTime SeriesSequentialBiomedical DataElectronic Health RecordsReview/Survey Paper
🎯 What it does: This paper proposes the CTFIDU+ algorithm, which can achieve point identification for any non-nested log queries when obtaining any realizable layer 3 (log layer) data (including log-distributed data sampled via ctf-rand()), and provides a completeness proof for this algorithm. It also clarifies the theoretical limits of realizable log data and identifiability, and derives a tighter analytical bound for the number of non-identifiable quantities based on this. Finally, it verifies through simulation and real data that using log data can significantly narrow the uncertainty interval.
Causal Matrix Completion under Multiple Treatments via Mixed Synthetic Nearest Neighbors
Minrui Luo (Tsinghua University), Zhiheng Zhang (Shanghai University of Finance and Economics)
Recommendation SystemOptimizationData-Centric LearningTabularBenchmark
🎯 What it does: Proposed the Mixed Synthetic Nearest Neighbor (MSNN) method to complete the causal matrix at multiple processing levels, especially for scenarios with missing not at random (MNAR) and scarce processing level data;
Causal Modeling of Selection in Evolution
Haoyue Dai (Carnegie Mellon University), Kun Zhang (Carnegie Mellon University)
OptimizationFederated LearningExplainability and InterpretabilityGraph Neural NetworkDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningGraphTabularBenchmark
🎯 What it does: Proposed and verified an evolutionary selection model, clarifying methods for causal discovery and identification under evolutionary conditions;
Causal Preference Elicitation
Edwin V. Bonilla (CSIRO), Daniel M. Steinberg (CSIRO)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningDrug DiscoveryReinforcement Learning from Human FeedbackGraph Neural NetworkMixture of ExpertsGraphTabularBiomedical DataBenchmark
🎯 What it does: Propose a Bayesian framework named CaPE that actively approximates the DAG posterior distribution by utilizing step-by-step feedback from experts on local causal relationships.
Causal Representation Learning with Optimal Compression and Complex Treatments
Wanting Liang (Shanghai University of Finance and Economics), Zhiheng Zhang (Shanghai University of Finance and Economics)
OptimizationRepresentation LearningHyperparameter SearchData-Centric LearningAuto EncoderGenerative Adversarial NetworkContrastive LearningImageGraphTabular
🎯 What it does: Propose an 'optimal compression' framework to address the hyperparameter tuning challenge in causal representation learning with multiple processors, and design three balancing strategies (Pairwise, OVA, Aggregation) as well as the generative model Multi-Treatment CausalEGM;
Causal Structure Learning for Sparse Matrix Fill-in Reduction
Ziwei Li (Chinese Academy of Sciences), Huiyuan Li (Chinese Academy of Sciences)
OptimizationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkPrompt EngineeringAuto EncoderContrastive LearningGraphTabularBenchmark
🎯 What it does: Propose a sparse matrix reordering framework based on causal structure learning (CTS), which significantly reduces fill-in by identifying and blocking causal triplets that lead to fill-in during LU factorization.
Causal-Adapter: Taming Text-to-Image Diffusion for Faithful Counterfactual Generation
Lei Tong (Centre for AI, DS and AI, Astrazeneca), Chen Jin (Centre for AI, DS and AI, Astrazeneca)
GenerationData SynthesisExplainability and InterpretabilityTransformerPrompt EngineeringDiffusion modelContrastive LearningImageBiomedical DataAlzheimer's Disease
🎯 What it does: Proposes the Causal-Adapter framework, which achieves high-fidelity causal counterfactual generation by utilizing a frozen text-to-image diffusion model through causal adapters;
Causal-aware Anomaly Detection for Tabular Data
Dang Nguyen (Deakin University), Sunil Gupta (Deakin University)
Anomaly DetectionGenerative Adversarial NetworkContrastive LearningTabular
🎯 What it does: Propose an unsupervised table anomaly detection method called CausalAno, which uses causal GAN to learn the causal structure of normal data, and performs anomaly scoring based on Mahalanobis distance in the discriminator's latent space.
Causal-EPIG: Causally Aligned Active CATE Estimation
Erdun Gao (Australian Institute for Machine Learning), Dino Sejdinovic (Australian Institute for Machine Learning)
OptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningContrastive LearningGaussian SplattingTabularBiomedical DataBenchmark
🎯 What it does: Propose an active learning framework called Causal-EPIG based on causal outcome alignment (COA), which is used for efficiently estimating the conditional average treatment effect (CATE) and reducing the cost of outcome measurements.
Causal-JEPA: Learning World Models through Object-Level Latent Masking
Heejeong Nam (Brown University), Randall Balestriero (Brown University)
Autonomous DrivingRepresentation LearningRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelAuto EncoderContrastive LearningWorld ModelImageVideoSequential
🎯 What it does: Propose C-JEPA, an object-level masked joint embedding prediction (JEPA) world model, which learns dynamic representations without using reconstruction loss by leveraging frozen object-aware encoders (such as VideoSAUR / SAVi).
CausalArmor: Efficient Indirect Prompt Injection Guardrails via Causal Attribution
Minbeom Kim (Google Cloud AI Research), Tomas Pfister (Google Cloud AI Research)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes CausalArmor, a selective defense framework based on causal attribution, designed to defend against indirect prompt injection attacks targeting tool-calling LLM agents.
CausalGame: Benchmarking Causal Thinking of LLM Agents in Games
Zhenhao Chen (MBZUAI), Kun Zhang (MBZUAI)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringWorld ModelTabularBenchmark
🎯 What it does: Created an interactive benchmark called CausalGame, where LLM agents actively conduct experiments, collect data, and propose final designs and explanations in a simulated drone design and deployment environment to evaluate their causal reasoning abilities.
Causally Evaluating the Learnability of Formal Language Tasks
Vésteinn Snæbjarnarson (ETH Zürich), Ryan Cotterell (ETH Zürich)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningSequential
🎯 What it does: This paper conducts a causal evaluation of the learnability of language models (Transformer and LSTM) in learning probabilistic finite automata (PFA) under a controlled formal language environment, revealing the errors of traditional correlation-based methods.
CausalProfiler: Generating Synthetic Benchmarks for Rigorous and Transparent Evaluation of Causal Machine Learning
Panayiotis Panayiotou (University of Bath), Özgür Şimşek (University of Bath)
Data SynthesisExplainability and InterpretabilityDiffusion modelScore-based ModelFlow-based ModelTabularBenchmark
🎯 What it does: Proposes CausalProfiler, a data set generator for randomly generating synthetic causal benchmarks, used to evaluate causal machine learning methods within a diverse space of causal models;
CausalRAG2: Hierarchical Causal Knowledge Graph Design for RAG
Nengbo Wang (Case Western Reserve University), Vipin Chaudhary (Case Western Reserve University)
RetrievalExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose CausalRAG2, a framework that improves schema retrieval augmented generation (RAG) through a hierarchical causal gating structure;
CausalX: A Unified and Causally-Interpretable Plug-and-Play Model for Multi-modal Spatio-Temporal Forecasting
Shiqi Zhang (Zhejiang University of Technology), Cong Bai (Zhejiang University of Technology)
OptimizationExplainability and InterpretabilityComputational EfficiencyRecurrent Neural NetworkGraph Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningMultimodalityGraphTime Series
🎯 What it does: Propose the CausalX model, construct a dynamic causal heuristic graph, and achieve interpretability and performance improvement in multi-modal spatiotemporal prediction.
CausalXRL: Explainable Reinforcement Learning through Causal Graph Reasoning
Yanming Zhang (Stony Brook University), Klaus Mueller (Stony Brook University)
Explainability and InterpretabilityGraph Neural NetworkReinforcement LearningContrastive LearningGraphTabular
🎯 What it does: Propose the CausalXRL framework, which uses causal graph reasoning to provide interpretability for model-free reinforcement learning;
CauScale: Neural Causal Discovery at Scale
Bo Peng (Shanghai Jiao Tong University), Chaochao Lu (Shanghai Artificial Intelligence Laboratory)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerScore-based ModelAuto EncoderContrastive LearningGraphTabularBenchmark
🎯 What it does: Proposes CauScale, a neural network architecture capable of efficiently performing causal structure learning on large-scale graphs with thousands of nodes.
CauSciBench: Can LLMs Automate Causal Inference in Real-World Scientific Research?
Sawal Acharya (Jinesis Lab, University of Toronto & Vector Institute), Zhijing Jin (Jinesis Lab, University of Toronto & Vector Institute)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes CauSciBench, a benchmark for evaluating large language models across the complete causal inference workflow (variable selection, method decision-making, implementation, and result interpretation).
CauseCollab: Causal Unified and Modality-Agnostic Network for Heterogeneous Collaborative Perception
Weize Li (Beijing University of Posts and Telecommunications), Jinglin Li (Beijing University of Posts and Telecommunications)
Autonomous DrivingComputational EfficiencyRepresentation LearningData-Centric LearningConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningOptical FlowImageMultimodalityPoint CloudStochastic Differential Equation
🎯 What it does: Propose a causal unified, modality-agnostic collaborative perception framework (CauseCollab), which achieves efficient collaboration across different sensor modalities by learning semantically consistent representations in the protocol space and performing local semantic reconstruction based on this.
Causes and Consequences of Representational Similarity in Machine Learning Models
Zeyu Michael Li (Duke University), Emily Wenger (Duke University)
Representation LearningAdversarial AttackConvolutional Neural NetworkTransformerMixture of ExpertsContrastive LearningImageText
🎯 What it does: Studied the causal impact of dataset overlap and task overlap in training data on the representational similarity of machine learning models, and systematically evaluated the similarity and susceptibility to attacks across different models (such as ResNet, ViT, nanoGPT, Llama, UNet, etc.).
CB-SLICE: Concept-Based Interpretable Error Slice Discovery
Yael Konforti (University of Cambridge), Mateja Jamnik (University of Cambridge)
Anomaly DetectionExplainability and InterpretabilityImageTabular
🎯 What it does: This study proposes CB-SLICE, an error slice discovery method based on the Concept Bottleneck Model (CBM);
CBV: Clean-label Backdoor Attacks on Vision Language Models via Diffusion Models
Ji Guo (University of Electronic Science and Technology of China), Wenbo Jiang (University of Electronic Science and Technology of China)
Adversarial AttackTransformerVision Language ModelDiffusion modelScore-based ModelImageTextMultimodality
🎯 What it does: Proposes a clean-label backdoor attack method based on diffusion models (CBV), which can generate natural-looking malicious samples that trigger Vision-Language Models without altering the text labels.
CCLRec: Consensus-driven Contrastive Learning for LLM-enhanced Graph Recommendation
Ting Guo (North University of China), Pinle Qin (North University of China)
Recommendation SystemGraph Neural NetworkTransformerLarge Language ModelContrastive LearningMultimodalityGraph
🎯 What it does: Propose the CCLRec framework, which deeply integrates large language models (LLMs) with graph neural networks (GNNs), leveraging the consistency between structural and semantic views to generate high-confidence positive and negative samples;
CE$^4$L: Continual Ego, Exo, and Ego-Exo Learning
Hongwei Yan (Tsinghua University), Liyuan Wang (Tsinghua University)
Computational EfficiencyRepresentation LearningMeta LearningTransformerVision-Language-Action ModelAuto EncoderContrastive LearningVideoBenchmark
🎯 What it does: Studied multi-perspective video continual learning, proposing the CE4L benchmark and the VISTA method.
CELL: A Causal Perspective for Fairness-aware Graph Adaptation
Hourun Li (Peking University), Ming Zhang (Peking University)
Domain AdaptationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: Propose the CELL framework to achieve fair graph domain adaptation without target-sensitive labels, leveraging causal graphs to decouple tasks from sensitive information, enhancing target domain supervision through pseudo-labels and group-unbiased learning, and finally mitigating domain differences by using a fairness bilateral graph alignment.
CellBRIDGE: Learning Cellular Trajectories via Interaction-Aware Alignment
Silas Ruhrberg Estévez (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Drug DiscoveryGraph Neural NetworkTransformerDiffusion modelFlow-based ModelContrastive LearningOptical FlowTabularTime SeriesBiomedical DataStochastic Differential Equation
🎯 What it does: Propose the CellBRIDGE method, which integrates ligand-receptor communication structures with gene expression, and solves cross-timepoint cell correspondence through multi-channel Fused Gromov–Wasserstein (FGW), thereby generating more accurate cross-snapshot matching that can serve as prior information for subsequent continuous time trajectory learning.
Cello: A Universal Cell-wise Feature Aggregation framework for Reliable Pathology Images Analysis
Hengrui Lou (Zhejiang University), Zunlei Feng (Zhejiang University)
ClassificationImage TranslationSegmentationData SynthesisAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningDrug DiscoveryConvolutional Neural NetworkGraph Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringMixture of ExpertsAuto EncoderGenerative Adversarial NetworkContrastive LearningImageBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission TomographyReview/Survey Paper
🎯 What it does: Propose the Cell-first framework Cello, which injects cell-level evidence into panoptic image modeling by learning cell-level features through protein signal supervised learning and aggregating them in fully convolutional layers.
Censoring with Plausible Deniability: Asymmetric Local Privacy for Multi-Category CDF Estimation
Qirui Hu (Shanghai University of Finance and Economics), Yi Liu (York University)
Safty and PrivacySupervised Fine-TuningContrastive LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: Propose an asymmetric local differential privacy mechanism based on suspicious denial (ACRR) for multi-class cumulative distribution function estimation, achieving privacy-preserving data collection without the need for predefined sensitive intervals.
CentaurEval: Benchmarking Human-in-the-Loop Value in Agentic Coding
Hanjun Luo (New York University Abu Dhabi), Hanan Salam (New York University Abu Dhabi)
AI Code AssistantLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Built and released the CentaurEval benchmark, specifically designed to quantify the value of human-AI collaboration in coding tasks.
Cerebellar-Inspired Residual Control for Fault Recovery: From Inference-Time Adaptation to Structural Consolidation
Nethmi Jayasinghe (University of Illinois Chicago), Amit Ranjan Trivedi (University of Illinois Chicago)
Robotic IntelligenceSpiking Neural NetworkTransformerReinforcement LearningMixture of ExpertsAuto EncoderContrastive LearningTime SeriesSequential
🎯 What it does: Propose a cerebellar residual control-based adaptation framework during inference, adding an online residual path on a frozen reinforcement learning policy to achieve fault recovery.
Cert-LAS: Toward Certified Model Ownership Verification for Text-to-Image Diffusion Models via Layer-Adaptive Smoothing
Leyi Qi (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
GenerationData SynthesisSafty and PrivacyDiffusion modelScore-based ModelImageTextMultimodality
🎯 What it does: Proposes a certifiable model ownership verification method called Cert‑LAS for text-to-image diffusion models.
Certain Head, Uncertain Tail: Expert-Sample for Test-Time Scaling in Fine-Grained MoE
Yuanteng Chen (Institute of Automation, Chinese Academy of Sciences), Jian Cheng (Institute of Automation, Chinese Academy of Sciences)
Computational EfficiencyKnowledge DistillationRepresentation LearningTransformerPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: Propose the Expert-Sample method in fine-grained MoE models, which retains high-confidence expert heads and introduces random sampling in the uncertain tail, thereby improving diversity without compromising stability.
Certificate-Guided Pruning for Stochastic Lipschitz Optimization
Ibne Farabi Shihab (Iowa State University), Anuj Sharma (Iowa State University)
OptimizationTabularTime SeriesSequentialBenchmark
🎯 What it does: Study black-box optimization of Lipschitz functions under noise assessment, proposing the Certificate-Guided Pruning (CGP) method, which explicitly maintains the active set A_t and constructs confidence certificates based on the Lipschitz UCB envelope, providing computable progress indicators and stopping criteria.
Certificates for Complex-Compatible Learned Cochain Laplacians
Nivar Anwer (Georgia Institute of Technology), David Elizondo (De Montfort University)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningDiffusion modelScore-based ModelFlow-based ModelRectified FlowAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowMeshGraphTabularTime SeriesSequentialBiomedical DataPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a computable compatibility certificate for evaluating and correcting inconsistencies in the learned chain conformal Laplacian operator, and provides a closed-form projection onto the nearest compatible complex.
Certified Circuits: Stability Guarantees for Mechanistic Circuits
Alaa Anani (Max Planck Institute for Informatics), Jonas Fischer (Max Planck Institute for Informatics)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImageText
🎯 What it does: Proposes the Certified Circuits framework, which provides provable stability guarantees for 'circuits' in neural networks under edits on concept datasets, resulting in more reliable and interpretable subnetworks.
Certified Robustness under Heterogeneous Perturbations via Hybrid Randomized Smoothing
Blaise Delattre (Institute of Science Tokyo), Yang Cao (Institute of Science Tokyo)
Safty and PrivacyComputational EfficiencyAdversarial AttackVision Language ModelImageTextMultimodality
🎯 What it does: Proposes a unified randomized smoothing framework that can provide robustness guarantees against both discrete and continuous perturbations in multi-modal models;
Certifying Capabilities from Finite Tests: When Is It Possible?
Changlong Wu (University of Arizona), Wojciech Szpankowski (Purdue University)
Information TheoryOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningContrastive LearningReview/Survey PaperBenchmark
🎯 What it does: This paper proposes a 'capability evaluation' framework centered on reasoning, systematically studying how to reliably demonstrate a model's capabilities across an entire task family under limited testing.
Certifying Graph Neural Networks Against Label and Structure Poisoning
Lukas Gosch (Technical University of Munich), Stephan Günnemann
ClassificationFederated LearningExplainability and InterpretabilityAdversarial AttackGraph Neural NetworkSupervised Fine-TuningAuto EncoderContrastive LearningGraph
🎯 What it does: Propose a semi-supervised self-training graph partition aggregation framework named ST-GPA, which provides verifiable robustness guarantees for graph neural networks under node label and structural poisoning attacks.
CFPO: Counterfactual Policy Optimization for Multimodal Reasoning
ZhangYuan Yu, Qicheng Lao (Beijing University of Posts and Telecommunications)
OptimizationExplainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: Proposed a cross-modal causal counterfactual policy optimization (CFPO), which forces the model to generate causal dependencies on visual evidence during inference by performing counterfactual interventions on visual attention within the reinforcement learning loop.