ICML 2026 Papers — Page 58
International Conference on Machine Learning · 6554 papers
Tensor Train Diffusion: Leveraging Low-Rank Structures for High-Dimensional Score-Based Sampling
Robert Gruhlke (Freie Universitaet Berlin), Lorenz Richter (Zuse Institute Berlin)
GenerationData SynthesisOptimizationComputational EfficiencyDiffusion modelScore-based ModelTabularTime SeriesSequentialPhysics RelatedStochastic Differential Equation
🎯 What it does: Utilize Functional Tensor Training (FTT) to solve high-dimensional HJB equations, and then achieve high-dimensional probability distribution sampling through a time-reversed diffusion process without samples.
Terminal Dimension Reduction for Time Series with Applications
Alexander Munteanu (TU Dortmund), Chris Schwiegelshohn (Aarhus University)
CompressionOptimizationRepresentation LearningTime Series
🎯 What it does: This paper proposes a new terminal dimensionality reduction technique for compressing the Fréchet distance of multi-dimensional time series (multi-dimensional curves) in high-dimensional Euclidean space, and uses this technique to construct k-clustering core points that are independent of dimensions.
TerraBind: Fast and Accurate Binding Affinity Prediction through Coarse Structural Representations
Matteo Rossi (Terray Therapeutics), John Anthony Parkhill (Terray Therapeutics)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerMixture of ExpertsDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningMultimodalityGraphTabularBiomedical Data
🎯 What it does: Developed TerraBind, a lightweight model based on a multi-modal pre-trained encoder and Pairformer, which achieves fast and high-accuracy ligand binding site conformation and binding affinity prediction using coarse-grained structural representations (protein Cβ and ligand heavy atoms).
Test-Time Anchoring for Discrete Diffusion Posterior Sampling
Litu Rout (University of Texas at Austin), Ira Kemelmacher-Shlizerman (Google)
RestorationSuper ResolutionTransformerDiffusion modelAuto EncoderImageText
🎯 What it does: Propose a training-free discrete diffusion posterior sampling method called APS (Anchored Posterior Sampling), which achieves high-quality inverse problem solving by quantizing expectations and using anchored remasking.
Test-Time Debiasing with Probabilistic Prompts via Wasserstein Distance in Vision-Language Models
Chengye Wang (Zhejiang University), Chaochao Chen (Zhejiang University)
ClassificationRetrievalDomain AdaptationExplainability and InterpretabilityTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Propose a distributed debiasing framework called W4D based on the Wasserstein distance, which achieves fair query embedding through parameter-free updates at test time using probabilistic prompts.
Test-Time Detoxification without Training or Learning Anything
Baturay Saglam (Yale University), Dionysis Kalogerias (Yale University)
GenerationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a detoxification method that is performed only during the inference phase, utilizing zeroth-order gradient descent to optimize input word embeddings, thereby reducing the toxicity of generated text without training the model or modifying its parameters.
Test-time Generalization for Physics through Neural Operator Splitting
Louis Serrano (New York University), Rudy Morel (Polymathic AI)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerMixture of ExpertsDiffusion modelContrastive LearningTabularTime SeriesBenchmarkPhysics Related
🎯 What it does: Propose a strategy that approximates unknown PDE dynamics during testing by combining pre-trained neural operators, achieving zero-shot out-of-distribution generalization without modifying model weights.
Test-Time Graph Search for Goal-Conditioned Reinforcement Learning
Evgenii Opryshko (University of Toronto), Igor Gilitschenski (University of Toronto)
Graph Neural NetworkReinforcement LearningGraphTabularSequential
🎯 What it does: Propose a lightweight graph search (TTGS) during testing for frozen goal-conditioned reinforcement learning (GCRL) policies, which uses offline data and the local geometry of the value function to plan subgoals, thereby completing long-distance tasks.
Test-Time Guidance for Flow-Based Generative Models via Parallel Tempering on Source Distributions
Shih-Hsin Wang (University of Utah), Bao Wang (University of Utah)
GenerationData SynthesisOptimizationReinforcement Learning from Human FeedbackTransformerDiffusion modelScore-based ModelFlow-based ModelImageTextMultimodality
🎯 What it does: Propose a gradient-agnostic test-time steering method based on the source space—Source Parallel Tempering (SPT)—which efficiently explores the source distribution through parallel temperature gradients and pCN sampling, addressing the challenges of traditional methods in overcoming low-probability barriers and handling non-differentiable rewards.
Test-Time Learning of Causal Structure from Interventional Data
Wei Chen (Hong Kong University of Science and Technology), Dongmei Zhang (Microsoft Research)
OptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningGraph Neural NetworkSupervised Fine-TuningMixture of ExpertsContrastive LearningGraphTabularBenchmark
🎯 What it does: Propose a TICL method that utilizes intervention data for self-enhanced training during testing, combining the JCI framework and two-stage supervised learning to accomplish causal structure (I-CPDAG) discovery and intervention target detection.
Test-time Offline Reinforcement Learning on Goal-related Experience
Marco Bagatella (ETH Zurich), Andreas Krause (ETH Zurich)
Recurrent Neural NetworkTransformerSupervised Fine-TuningReinforcement LearningContrastive LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: This paper proposes a method for refining the goal-conditioned reinforcement learning (GCRL) policy trained offline during testing, called GC-TTT. The method selects sub-trajectories that are relevant and optimal to the current state and goal from the pre-trained dataset through self-supervised data screening during testing, and performs a few gradient updates on these data. After that, the updated policy is executed for a period of time and then reset.
Test-Time Reinforcement Learning for Flow Matching
Jili Chen (Zhejiang Normal University), Xiaodi Huang (Charles Sturt University)
GenerationOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningScore-based ModelFlow-based ModelRectified FlowImageTextMultimodalityBenchmarkStochastic Differential Equation
🎯 What it does: Proposes Flow-TTRL, a framework that optimizes flow matching models with reinforcement learning during the inference phase to achieve immediate alignment with human preferences.
Test-Time Training with KV Binding Is Secretly Linear Attention
Junchen Liu (NVIDIA), Ruilong Li (NVIDIA)
OptimizationComputational EfficiencyRepresentation LearningMeta LearningTransformerLarge Language ModelMixture of ExpertsAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Clarifies and proves that the mechanism of using KV binding in test-time training (TTT) is essentially equivalent to learned linear attention, revealing its true non-memorization operation principle, and based on this, proposes structural simplification and parallelization improvements for the model.
TestExplora: Benchmarking LLMs for Proactive Bug Discovery via Repository-Level Test Generation
Steven Liu (Microsoft), Scarlett Li (Microsoft)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Designed the TestExplora benchmark to evaluate the performance of large language models in warehouse-level proactive defect discovery.
Testing For Distribution Shifts with Conditional Conformal Test Martingales
Shalev Shaer (Technion IIT), Yaniv Romano (Technion IIT)
Domain AdaptationAnomaly DetectionContrastive LearningImage
🎯 What it does: This paper proposes an online distribution drift detection method based on Conditional Conformal Test Martingale (CCTM), aiming to avoid the test-time contamination problem that occurs in traditional CTM methods;
TetraJet-v2: Accurate NVFP4 Training for Large Language Models with Oscillation Suppression and Outlier Control
Yuxiang Chen (Tsinghua University), Jianfei Chen (Tsinghua University)
OptimizationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposes TetraJet-v2, an end-to-end 4-bit full quantization training (FQT) method for LLMs, using the NVFP4 format for quantizing activations, weights, and gradients.
TexEditor: Structure-Preserving Text-Driven Texture Editing
Bo Zhao (Nanjing University), Wei Ji (Nanjing University)
Image TranslationRestorationGenerationTransformerReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityMeshBenchmark
🎯 What it does: Propose TexEditor, a model that performs texture editing under text guidance while preserving geometric structure.
Text Before Vision: Staged Knowledge Injection Matters for Agentic RLVR in Ultra-High-Resolution Remote Sensing Understanding
Fengxiang Wang (National University of Defense Technology), Bo Du (Wuhan University)
Representation LearningData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study proposes a staged knowledge injection method that first uses Earth-science text QA for cold-start SFT in the UHR RS task, and then performs Agentic RLVR pre-warming on the same set of hard samples, significantly improving visual reasoning performance.
Text Generation as Continuous Latent Dynamics via Reinforcement Learning
Chen Jia (Independent Researcher)
GenerationTransformerLarge Language ModelReinforcement LearningTextOrdinary Differential Equation
🎯 What it does: Model text generation as a continuous-time latent dynamics process, using neural ODEs to control the flow of latent states, and jointly optimizing the generation policy and dynamics model through reinforcement learning.
Text Has Curvature
Karish Grover (Carnegie Mellon University), Geoffrey J. Gordon (Carnegie Mellon University)
RetrievalCompressionRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a discrete curvature metric called Texture based on the text itself. It first proves through theoretical and empirical analysis that natural language has a non-flat structure in bidirectional reasoning, then defines and calculates this curvature, using it as an untrained control signal to guide long-context compression and retrieval routing.
Text-Conditional JEPA for Learning Semantically Rich Visual Representations
Chen Huang (Apple), Joshua M. Susskind (Apple)
Image TranslationObject DetectionSegmentationRetrievalCompressionAnomaly DetectionFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningReinforcement Learning from Human FeedbackNeural Architecture SearchTransformerVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a text-conditioned feature prediction framework called TC-JEPA, which uses image captions to modulate the masked image completion task, thereby significantly reducing prediction uncertainty and obtaining more semantically rich visual representations.
Text-Driven Fusion for Infrared and Visible Images: Achieving Image Scene Adaptation on Hyperbolic Space
Huan Kang (Jiangnan University), Josef Kittler (University of Surrey)
Image TranslationRestorationDomain AdaptationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposed the TEDFusion framework, which utilizes text-driven hyperbolic space for infrared and visible image fusion, achieving adaptive fusion without text input during inference.
TextAtlas5M: A Large-Scale Dataset for Long Text Image Generation
Dongxing Mao (Central South University), Min Li (Central South University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposed and constructed the TextAtlas5M large-scale long-text image generation dataset, and designed a human evaluation set with 4000 samples called TextAtlasEval.
TextME: Bridging Unseen Modalities Through Text Descriptions
Soyeon Hong (Ajou University), Hyunsouk Cho (Ajou University)
ClassificationRetrievalRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningImageTextMultimodalityPoint CloudBiomedical DataAudio
🎯 What it does: Propose the TextME framework, which utilizes unpaired training with textual descriptions to project multiple pre-trained contrastive encoders into a unified LLM embedding space, achieving multi-modal scalability.
TextMesh4D: Zero-shot Text-to-4D Mesh Generation
Sisi Dai (National University of Defense Technology), Kai Xu (Institute of AI for Industries Chinese Academy of Sciences)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelScore-based ModelGaussian SplattingImageVideoTextMesh
🎯 What it does: Generate continuous 4D dynamic mesh sequences from text prompts, achieving zero-shot text-driven 4D content creation.
TextResNet: Decoupling and Routing Optimization Signals in Compound AI Systems via Deep Residual Tuning
Suizhi Huang (Nanyang Technological University), Xiaoxiao Li (University of British Columbia)
OptimizationExplainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextTabularBenchmarkChain-of-Thought
🎯 What it does: Designed the TEXTRESNET framework to address optimization failures caused by gradient signal confusion in composite AI systems.
Textual Stochastic Gradient Descent: Discrete Optimization of External Memory for Reasoning Language Agents
Jian Li (Beijing Normal University), Hua Huang (Beijing Normal University)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a framework called TSGD for discrete optimization on external storage (experience repository), which can continuously improve memory content through add, delete, and modify operations driven by failures under a fixed model.
Textual Supervision Enhances Geospatial Representations in Vision-Language Models
Marcelo Sartori Locatelli (Max Planck Institute for Security and Privacy), Meeyoung Cha (Max Planck Institute for Security and Privacy)
Representation LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Studied the improvement of text supervision on the geospatial representations of vision-language models
TF-FACE: Time-Frequency Fusion Learning via Frequency-Domain Adaptive and Controllable Enhancement for Trajectory Prediction
Dongjian Song (Jilin University), Jiayi Han (Jilin University)
Autonomous DrivingTransformerDiffusion modelAuto EncoderContrastive LearningOptical FlowTime SeriesSequential
🎯 What it does: Propose a trajectory prediction framework called TF-FACE based on time-frequency fusion, which improves traditional time-domain models through frequency-domain adaptive weighting and controllable enhancement.
TFRBench: A Reasoning Benchmark for Evaluating Forecasting Systems
Md Atik Ahamed (Google), Tomas Pfister (Google)
Recommendation SystemAnomaly DetectionAutonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyDrug DiscoveryReinforcement Learning from Human FeedbackNeural Architecture SearchProtein Structure PredictionTransformerLarge Language ModelAgentic AITabularTime SeriesBenchmarkFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the TFRBench benchmark to evaluate the inference capability and numerical accuracy of prediction systems; generate, verify, and improve inference trajectories through multi-agent collaboration.
TFTF: Training-Free Targeted Flow for Conditional Sampling
Qianqian Qu (Tsinghua University), Jun S. Liu (Tsinghua University)
GenerationData SynthesisTransformerFlow-based ModelImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a training-free conditional sampling method called TFTF, which utilizes a pre-trained flow matching model and a model-free likelihood function. By combining importance sampling with SMC resampling, it introduces a tunable noise flow in the intermediate stage of the generation process to generate samples that satisfy given conditions.
TG-RAG: A Retrieval-Augmented Framework for Reasoning Guidance in Specialized Domains
Liang Su (Fudan University), Li Sun (Beijing University of Posts and Telecommunications)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextGraphTabularFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a thought-guided framework based on retrieval augmentation called TG-RAG, which structures standard operating procedures (SOP) using an Expert Procedure Graph (EPG), and dynamically injects step instructions during the reasoning process through an Interrupt-Retrieve-Generate (IRG) mechanism, achieving active control over large reasoning models.
TGPO: Efficient Policy Optimization through Sequence Anchor and Information Gating
Hang Ding (Shanghai Jiao Tong University), Yabiao Wang (Tencent Youtu Lab)
TransformerReinforcement LearningPrompt EngineeringContrastive LearningTextSequentialBenchmark
🎯 What it does: Propose Trust‑Gated Policy Optimization (TGPO), combining sequence anchors with information gating to achieve adaptive balancing of token-level and sequence-level importance sampling updates in reinforcement learning with self-verified rewards (RLVR).
TGV-KV: Text-Grounded KV Eviction for Vision-Language Models
Jizhihui Liu (Harbin Institute of Technology), Yaowei Wang (Harbin Institute of Technology)
Computational EfficiencyTransformerVision Language ModelImageVideoTextMultimodality
🎯 What it does: Propose a text-guided KV cache eviction method called TGV-KV, which dynamically allocates hierarchical budgets using text-visual interaction information, weights and sorts visual KV, and prioritizes retaining text KV, thereby significantly reducing the memory consumption and inference latency of VLM under extremely low budgets.
The (Marginal) Value of a Search Ad: An Online Causal Framework for Repeated Second-price Auctions
Yuxiao Wen (New York University), Zhengyuan Zhou (New York University)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackTabularTime SeriesFinance Related
🎯 What it does: This paper proposes an online causal inference framework for simultaneously estimating the marginal value (treatment effect) of advertisements and performing automatic bidding in repeated second-price auctions (SPAs), and provides an algorithm that achieves asymptotically optimal regret.
The Abstraction Gap in Vision-Language Causal Reasoning
Chinh Hoang (University of Nebraska Lincoln), Mohammad Rashedul Hasan (University of Nebraska Lincoln)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a dual-probe evaluation framework that distinguishes between linguistic feasibility and structural credibility, and defines the abstraction gap (Abstraction Gap) as a metric.
The Accumulation of Score Estimation Error in Diffusion Models
Baoxiang He (Shanghai Jiao Tong University), Bo Jiang (Shanghai Jiao Tong University)
GenerationData SynthesisDiffusion modelScore-based ModelTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper studies the propagation and accumulation of score estimation errors in diffusion models, and provides non-asymptotic Wasserstein error upper bounds.
The ACUTE Protocol: Operationalizing Language Model Activations for Better Calibration, Utility, and Trust
Nishant Subramani (Carnegie Mellon University), Hamid Palangi (Google)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: Proposed the ACUTE protocol, which uses internal activations of language models to estimate the confidence of outputs, thereby improving the calibration and practicality of confidence estimation.
The Appeal and Reality of Recycling LoRAs with Adaptive Merging
Haokun Liu (University of Toronto), Colin Raffel (University of Toronto)
Computational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: Evaluated the impact of using adaptive merging methods on multi-task performance across approximately 1000 LoRA pools contributed by real users.
The Art of Interrogation: Consistency Amplifies Factuality in Spatial Reasoning
Théo Uscidda, Leonidas Guibas (Google DeepMind)
Autonomous DrivingRepresentation LearningReinforcement Learning from Human FeedbackReinforcement LearningVision Language ModelContrastive LearningImageTextPoint CloudChain-of-Thought
🎯 What it does: Improving the accuracy of pre-trained vision-language models on spatial reasoning tasks through self-supervised consistency verifiers and reinforcement learning;
The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models
Christina Lu (University of Oxford), Jack Lindsey (Anthropic)
Safty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper defines and quantifies the 'assistant axis' through linear analysis of the activation space of large language models, investigates its impact on model personality drift, and proposes an activation pruning method to stabilize the default assistant role.
The Axiomatic Value of Regularization in AI Alignment from Human Preferences
Ezgi Korkmaz
Reinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningText
🎯 What it does: Analyze and prove the impact of regularization on the social choice axiomatic properties of RLHF algorithms in AI alignment.
The benefits of full data shuffle, now with optimal I/O cost: $k$-wise independence and matrix transposition to the rescue
Peyman Afshani (Aarhus University), Mariafiore Tognon (University of Padova)
OptimizationFederated LearningComputational EfficiencyData-Centric LearningContrastive LearningTabularBenchmark
🎯 What it does: Designed and implemented two I/O-level linear algorithms that achieve 2-wise and approximate k-wise random permutations (GEN-2-WISE-IND-PERM and IO SHUFFLE), significantly reducing the I/O cost of data reordering while maintaining the fast convergence of SGD (O(1/T²)).
The Bridge-Garden Dilemma in LLM Distillation: Why Mixing Hard and Soft Labels Works
Guanghui Wang (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Studied the effectiveness of combining hard labels and soft labels in training during the distillation of large language models (LLMs), and proposed the Bridge‑Garden explanation framework and corresponding hybrid distillation method.
The Catastrophic Failure of *the* k-Means Algorithm in High Dimensions, and How Hartigan's Algorithm Avoids It
Roy R Lederman, Tamir Bendory (Tel Aviv University)
OptimizationContrastive LearningGaussian SplattingImageTextTabularBenchmark
🎯 What it does: This paper proves that in high-dimensional high-noise environments, the Lloyd k-means algorithm almost all partitions are fixed points, leading the algorithm to make almost no improvements; whereas Hartigan k-means does not exhibit this phenomenon under the same conditions and can converge to the correct clustering.
The Consistency Dilemma in LLMs: Generator-Evaluator Agreement and Vulnerability to Mistakes
Marina Mancoridis (Massachusetts Institute of Technology), Zoe Hitzig (Harvard Society of Fellows)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented a generation-evaluation self-consistency testing framework, automatically extracting concepts and evaluating 491 concepts across 10 state-of-the-art LLMs, then associating self-consistency with medical validation errors, revealing a consistency dilemma.
The Convergent Representation of Contrastive Vision-Language Models: Geometry, Modality Gap and Shared Space Alignment
Lingjie Yi (Stony Brook University), Chao Chen (Stony Brook University)
ClassificationRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper reveals the geometric origin of modality gaps caused by dimensional collapse through theoretical analysis of the convergence optimal representation (COR) in multimodal contrastive learning.
The Cost of Commitment in Option-Based Hierarchical RL
Randy Lefebvre (Universite Laval), Audrey Durand (Universite Laval)
Reinforcement LearningTabularBenchmark
🎯 What it does: Proposed and quantified the 'commitment loss' to evaluate the trade-off between option duration, decision cost, and model error in hierarchical reinforcement learning.
The Cost of Information: Phase Transitions in Contextual Bandits with Paid Observations
Xueping Gong (Xiamen University), Jiheng Zhang (Hong Kong University of Science and Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the contextual multi-armed bandit model where observations are paid for in each round, and proposes an optimal algorithm in adversarial environments while analyzing the impact of information cost on minimizing total regret.
The Cost of Learning Under Multiple Change Points
Tomer Gafni (Columbia University), assaf zeevi
Anomaly DetectionOptimizationFederated LearningComputational EfficiencyTime Series
🎯 What it does: Studied the multi-change-point scenario in online learning environments, and proposed an algorithm to track the mean without prior information.
The Crowded Embedding Space: A Mean-Field Mechanism for Emergent Marginalization in Retrieval-Augmented Agents
Shwan Ashrafi (Oracle), Dan Roth (Oracle)
RetrievalRepresentation LearningData-Centric LearningTransformerContrastive LearningImageTextMultimodalityRetrieval-Augmented GenerationStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper investigates the mechanism by which minority user interests are marginalized in shared embedding spaces due to target collisions in retrieval-augmented generation (RAG) agents, and provides theoretical and experimental validation.
The Cylindrical Representation Hypothesis for Language Model Steering
Lang Gao (Mohamed bin Zayed University of Artificial Intelligence), Xiuying Chen (Mohamed bin Zayed University of Artificial Intelligence)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Proposed and verified the cylindrical representation hypothesis (CRH) based on cylindrical geometry to explain the sample dependence and instability of large language models during the activation regulation (steering) process.
The Data Manifold under the Microscope
Marios Koulakis (Independent Researcher), Constantin Seibold (University Clinic Heidelberg)
GenerationData SynthesisRepresentation LearningConvolutional Neural NetworkAuto EncoderGenerative Adversarial NetworkImageBenchmark
🎯 What it does: Proposes a reproducible benchmark framework that uses densely axis-aligned sampled low-dimensional image datasets to accurately estimate geometric quantities of data manifolds (volume, curvature, reach), and evaluates the observability of generative models and theoretical estimation errors using this framework;
The Decrypto Benchmark for Multi-Agent Reasoning and Theory of Mind
Andrei Lupu (FAIR at Meta), Jakob Nicolaus Foerster (FAIR at Meta)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and implemented DECRYPTO, a multi-agent reasoning and theory of mind (ToM) benchmark based on the board game 'Decrypto,' which can evaluate the collaborative, competitive, and ToM capabilities of large language models (LLMs) in interactive game environments.
The Deterministic Horizon: When Extended Reasoning Fails and Tool Delegation Becomes Necessary
Dongxin Guo (University of Hong Kong), SM Yiu
Explainability and InterpretabilityAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: The paper explores the performance degradation in extended reasoning tasks for deterministic state tracking and proposes the necessity of tool delegation.
The Devil is in the Condition Numbers: Why is GLU Better than non-GLU Structure?
Xingyu Lyu (State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences), Qingming Huang (School of Computer Science and Technology, University of Chinese Academy of Sciences)
OptimizationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkTransformerDiffusion modelScore-based ModelFlow-based ModelNeural Radiance FieldAuto EncoderGenerative Adversarial NetworkContrastive LearningImageTextTabularTime SeriesSequential
🎯 What it does: This paper systematically analyzes the differences between Gated Linear Unit (GLU) and non-GLU structures in two-layer networks through the Neural Tangent Kernel (NTK) framework and random matrix theory, and empirically verifies the optimization speed advantage of GLU.
The Devil is in the Spectrum: Mitigating Representation Collapse in LLMs via Topologically Regularized Side-Path
Yiheng Tao (Peking University), Jie Chen (Harbin Institute Of Technology)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a method called Topologically Regularized Side-Path (TRSP) to alleviate the representation collapse problem in large language models (LLMs), especially in terms of long context performance.
The Differences Between Direct Alignment Algorithms are a Blur
Alexey Gorbatovski (T Tech), Daniil Gavrilov (T Tech)
OptimizationHyperparameter SearchData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a unified two-stage training framework that places all direct alignment algorithms (DAA) within the same hyperparameter space, and introduces a β regulation factor to modify one-stage methods (such as ORPO, ASFT).
The Double Dilemma in Multi-Task Radiology Report Generation: A Gradient Dynamics Analysis and Solution
Erjian Zhang (Xinjiang University), Zhiqing Guo (Xinjiang University)
GenerationOptimizationData-Centric LearningTransformerDiffusion modelScore-based ModelTextBiomedical DataComputed TomographyReview/Survey PaperStochastic Differential Equation
🎯 What it does: Propose an optimizer called CAME-Grad for multi-task radiology report generation, addressing the gradient conflict issues caused by traditional linear scalarization.
The Double-Edged Nature of the Rashomon Set for Trustworthy Machine Learning
Ethan Hsu (Duke University), Lesia Semenova (Rutgers University)
Information TheorySafty and PrivacyExplainability and InterpretabilityAdversarial AttackTabularBenchmarkFinance Related
🎯 What it does: This paper investigates the double-edged effect of the Rashomon set of machine learning models on trustworthiness, exploring how diversity simultaneously enhances reactive robustness and exacerbates information leakage.
The Easy, the Hard, and the Learnable: Confidence and Difficulty-Adaptive Policy Optimization for LLM Reasoning
Zhanke Zhou (Hong Kong Baptist University), Sanmi Koyejo (Stanford University)
OptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the Confidence and Difficulty-Adaptive Policy Optimization (CoDaPO) method for reinforcement learning after training of large language models on verifiable reasoning tasks. By estimating the confidence and difficulty of each question, CoDaPO achieves adaptive reweighting of update weights and resampling of high-value questions, thereby improving computational efficiency and enhancing reasoning accuracy.
The Efficiency Gap in Byte Modeling
Celine Lee (Google DeepMind), Ruoxi Wang (Google DeepMind)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelDiffusion modelContrastive LearningText
🎯 What it does: Studied the computational efficiency gap between byte-level language models without tokenizers or sequential constraints and traditional subword-level models under autoregressive (AR) and masked diffusion modeling (MDM) objectives, using compute-matched scaling experiments to quantify performance overhead.
The Entropic Signature of Class Speciation in Diffusion Models
Florian Handke (Ghent University), Luca Ambrogioni (Radboud University)
GenerationData SynthesisExplainability and InterpretabilityDiffusion modelScore-based ModelImageText
🎯 What it does: This paper proposes using class-conditional entropy and its time derivative to identify the differentiation window of semantic decisions in diffusion models, and implements its operational estimation during model training.
The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Level
Jeremy Herbst (University of Hamburg), Jae Hee Lee (University of Hamburg)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsContrastive LearningTextSequential
🎯 What it does: This paper investigates the interpretability of Mixture-of-Experts (MoE) language models, comparing the polysemy of MoE experts and dense FFN neurons using k-sparse probes, and demonstrates that the expert level can serve as an interpretable module through LLM-generated expert function labels and causal validation.
The Expressive Power of Low Precision Softmax Transformers with (Summarized) Chain-of-Thought
Moritz Brösamle (University of Tübingen), Stephan Eckstein (University of Tübingen)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proves that under low-precision activation and attention weight rounding, the standard softmax Transformer can achieve Turing completeness through Chain-of-Thought and Summarized Chain-of-Thought, and provides specific bounds on how depth and width grow logarithmically with context length.
The Extra Tokens Matter: Disentangled Representation Learning with Vision Transformers
Maofeng Tang (University of Tennessee), Hairong Qi (University of Tennessee)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningTransformerAuto EncoderContrastive LearningImage
🎯 What it does: This paper introduces learnable additional factor tokens into Vision Transformer, and achieves unsupervised semantic decomposition through minimal volume constraint and multi-stage aggregation;
The Fairness Hierarchy: A viewpoint from causal inference
ChengBo Zhang (Dalian University of Technology), Changcheng Li (Dalian University of Technology)
Federated LearningSafty and PrivacyExplainability and InterpretabilityContrastive LearningTextTabularReview/Survey PaperBenchmark
🎯 What it does: A causal fair prediction framework that depends on multiple sensitive paths is proposed by selectively intervening on sensitive attribute paths in structural causal models.
The First Drop of Ink: Nonlinear Impact of Misleading Information in Long-Context Reasoning
Muhan Gao (Texas A&M University), Kuan-Hao Huang (Texas A&M University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Investigate the nonlinear impact of misleading information on performance in long context reasoning, discovering the 'First Drop of Ink' effect.
The Fisher Dimension: Instance-Dependent Complexity for Causal Discovery
Luong Doan (N2TP Technology Solutions JSC), Tuan Do (N2TP Technology Solutions JSC)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningContrastive LearningGraphTabularBenchmark
🎯 What it does: Propose and study the Fisher dimension — an instance-specific complexity measure for recovering Markov equivalence classes (MEC) in linear Gaussian structural equation models (SEM).
The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models
Zanlin Ni (Tsinghua University), Gao Huang (Tsinghua University)
Computational EfficiencyAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningDiffusion modelText
🎯 What it does: This paper investigates whether diffusion-based large language models (dLLM) require arbitrary order generation for reasoning tasks, and proposes that using standard autoregressive (AR) order for reinforcement learning can enhance reasoning capabilities, significantly improving performance while maintaining the advantages of parallel decoding.
The Forgetting-Retention Dilemma: Certified Unlearning Theory in Continual Learning
Yiting Hu (Singapore University of Technology and Design), Qian Zhang (Singapore University of Technology and Design)
Federated LearningSafty and PrivacyContrastive LearningGaussian SplattingImage
🎯 What it does: Proposed and implemented a certified machine forgetting method in a continual learning environment, constructed a theoretical framework, and provided an upper bound on post-forgetting risks.
The Generalization Spectrum: A Chromatographic Approach to Evaluating Learning Algorithms
Jinghan Zhang (ByteDance Seed), Tianle Cai (ByteDance Seed)
Data-Centric LearningAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes the Generalization Spectrum framework, which constructs five-level test variants ranging from exact recall to distant transfer for each training sample, to visualize the generalization radius of learning algorithms.
The Geometric Mechanics of Contrastive Representation Learning: Alignment Potentials, Entropic Dispersion, and Cross-Modal Divergence
Yichao Cai (Adelaide University), Javen Qinfeng Shi (Adelaide University)
RetrievalRepresentation LearningTransformerAuto EncoderContrastive LearningGaussian SplattingImageTextMultimodality
🎯 What it does: Propose a geometric analysis of InfoNCE contrastive learning based on a measure-theoretic framework, revealing deterministic energy functions and their gradient consistency in the large-batch limit; explain the unique Gibbs equilibrium of single-modal alignment through intrinsic free energy; introduce negative symmetric KL cross-coupling in the multi-modal case to explain and predict persistent separation between modalities (modality gap).
The Geometric Origin of Grokking: Accelerating Generalization via Active Structural Reorganization
Kefei Tao (National University of Defense Technology), Xiaojun Duan (Hong Kong Baptist University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningTextSequential
🎯 What it does: Analyze the geometric evolution of internal representations during the grokking process in Transformers, and propose the R2G (Repel-to-Grokking) loss, which actively promotes the model's transition from the memorization phase to the generalization phase by applying geometric repulsion between Value vectors in hidden layers.
The Geometric Reasoner: Manifold-Informed Latent Foresight Search for Long-Context Reasoning
Ren Zhuang (Hangzhou Normal University), Shuifa Sun (Hangzhou Normal University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningAI Code AssistantTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextSequentialRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a training-free geometric reasoner TGR, which enhances coverage quality in long-context reasoning by utilizing manifold-aware latent look-ahead search;
The Geometry of Narrow Fine-Tuning Degradation: Trajectory Lock-in and Spectral Bifurcation
Jia Liu (South China University of Technology), Min Chen (South China University of Technology)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper investigates the trajectory locking and spectral bifurcation phenomena that occur during narrow task fine-tuning of large language models, explores the impact of update directions on the degradation of broad capabilities, and proposes a diagnostic method based on geometric probes and SVD regularization intervention;
The Geometry of Projection Heads: Conditioning, Invariance, and Collapse
Faris Chaudhry (Imperial College London)
Representation LearningConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes a geometric theory of the projection head in self-supervised learning, viewing the projection head as a learnable Riemannian metric, and systematically analyzes its roles in three aspects: whitening, invariance, and collapse. Furthermore, it proves that linear projection heads achieve subspace whitening, while non-linear projection heads can adaptively shape local metrics and dispel collapse through a negative curvature mechanism. Finally, it explains the geometric reasons behind the 'guillotine effect' (when the projection head is discarded) and the degradation of information entropy.
The Geometry of Reasoning: Self-Evaluation via Layerwise Trajectory Evolution
Jinhe Bi (Huawei Heisenberg Research Center), Yunpu Ma (LMU Munich)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a white-box self-evaluation framework called GeoR, based on the hierarchical evolution of hidden states, to assess the correctness of outputs from large-scale reasoning models.
The Geometry of Representational Failures in Vision Language Models
Daniele Savietto (Universita di Torino), Alan Perotti (Intesa Sanpaolo AI Research)
Explainability and InterpretabilityRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Studied the errors of vision-language models (VLMs) in multi-object tasks, and analyzed their internal geometric structure by extracting concept vectors, proving that geometric overlap leads to binding failure.
The Geometry of Sequential Learning: Lie-Bracket Prediction of Transfer Order
John Sweeney (Sideplane AI)
OptimizationFederated LearningHyperparameter SearchMeta LearningLarge Language ModelSupervised Fine-TuningDiffusion modelTextSequential
🎯 What it does: Proposes using the Lie bracket (Lie-Bracket) of gradient updates to predict the optimal training order for multi-source datasets during sequential fine-tuning, addressing the factorial-level curriculum search problem.
The Geometry of Updates: Fisher Alignment at Vocabulary Scale
John Sweeney (Sideplane AI)
Domain AdaptationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningImageTextTabular
🎯 What it does: In large language models with shared vocabularies, the authors propose a task similarity metric based on Fisher alignment, addressing the limitations of representation similarity metrics in the 'activation dark' scenario.
The Heterogeneous Safety Impacts of Benign Multilingual Fine-Tuning
Will Hawkins (Oxford Internet Institute, University of Oxford), Chris Russell (Oxford Internet Institute, University of Oxford)
Safty and PrivacyExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper systematically evaluates the impact of non-attackive data translation on model safety behavior by performing LoRA fine-tuning on the Llama-3.2, Qwen3, and Gemma-3 models across multiple languages.
The Hidden Cost of Structured Generation in LLMs: Draft-Conditioned Constrained Decoding
Avinash Reddy (University of Central Florida), Amrit Singh Bedi (University of Central Florida)
GenerationComputational EfficiencyData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a two-step reasoning process—Draft-Conditioned Constrained Decoding (DCCD), which first generates an unconstrained draft and then performs conditional constrained decoding to improve the accuracy of structured generation.
The Hidden Link between RLHF and Contrastive Learning
Xufei Lv (Tsinghua University), Houde Liu (Tsinghua University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningContrastive LearningText
🎯 What it does: This paper interprets the alignment objectives of RLHF and DPO as mutual information maximization, and proposes using the Jensen-Shannon mutual information estimator to replace DV/MINE, forming a new method called MIO;
The Hidden Risk: Membership Inference Attacks on Multimodal Federated Learning via Modality Imbalance
chang Ma, Yiyang Ni (Jiangsu Second Normal University)
Anomaly DetectionFederated LearningSafty and PrivacyTransformerAuto EncoderGenerative Adversarial NetworkContrastive LearningVideoMultimodalityAudio
🎯 What it does: Study membership inference attacks caused by modality imbalance in multi-modal federated learning, and propose a modality-based attack framework.
The Hippocampal Place Field Gradient: A Bio-inspired Framework Building Multiscale Representation for Better Sample Efficiency
ZHOU Shujun, Guozhang Chen (Peking University)
Computational EfficiencyKnowledge DistillationRepresentation LearningData-Centric LearningTransformerLarge Language ModelPrompt EngineeringDiffusion modelContrastive LearningTextTabularRetrieval-Augmented Generation
🎯 What it does: This paper proposes a unified theoretical framework that explains how the hippocampus achieves continuous place field gradients through frequency-dependent grid cell projection attenuation, and based on this, proposes a multi-scale position encoding called HIPE, significantly improving the sample efficiency of Transformers under data-limited conditions.
The Illusion of Generalization in Tabular Language Models
Aditya Gorla (University of California, Los Angeles), Ratish Puduppully (IT University of Copenhagen)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTabularBenchmarkFinance Related
🎯 What it does: A systematic re-evaluation of the Tabula-8B table language model using the UniPredict dataset of 165 datasets to assess its performance on three tasks: binary classification, categorical classification, and quartile classification, revealing three main findings: task type bias, data leakage, and instruction tuning dominate the results.
The Implicit Bias of Adam and Muon on Smooth Homogeneous Neural Networks
Eitan Gronich (Weizmann Institute), Gal Vardi (Weizmann Institute)
OptimizationAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper investigates the implicit preferences of momentum optimizers (Adam, Muon, and their combinations) on smooth homogeneous neural networks, proving that these algorithms tend to maximize the corresponding norm margins under decaying learning rates.
The Implicit Bias of Depth: From Neural Collapse to Softmax Codes
Connall Garrod (University of Oxford), Christos Thrampoulidis (University of British Columbia)
ClassificationOptimizationRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper investigates how depth induces implicit low-rank bias in deep unconstrained feature models (UFM) when trained without regularization under cross-entropy loss, leading to neural collapse (NC) no longer being the unique optimum, and revealing that low-rank structures correspond to softmax encoding;
The Implicit Bias of Steepest Descent with Mini-batch Stochastic Gradient
Jichu Li (Renmin University of China), Difan Zou (University of Hong Kong)
ClassificationOptimizationImageTabular
🎯 What it does: Analyzes and reveals the implicit bias of the steepest descent induced by different norms in multi-class linear models under the mini-batch stochastic gradient descent (SGD) environment, exploring the impact of batch size, momentum, and variance reduction on convergence speed and the limiting maximum margin solution.
The Information Geometry of Softmax: Probing and Steering
Kiho Park (University of Chicago), Victor Veitch (University of Chicago)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageText
🎯 What it does: Propose the Dual Steering method based on information geometry, which is used to robustly manipulate representations in the softmax model to achieve changes in target concepts;
The Interplay Between Interpolation and Aggregation in Regression: Optimal Sample Complexity
Mikael Møller Høgsgaard (Aarhus University), Liang-Yu Zou (Aarhus University)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningData-Centric LearningAgentic AIPrompt EngineeringMixture of ExpertsContrastive LearningReview/Survey Paper
🎯 What it does: Studied the interaction between interpolation and aggregation in regression, proved that the γ-graph dimension determines the sample complexity of aggregation learning, and proposed an optimal median aggregation algorithm with three complementary interpolators.
The Invisible Lottery: How Subtle Cues Steer Algorithm Choice in LLM Code Generation
Akanksha Narula (Max Planck Institute for Software Systems), Laurent Bindschaedler (Max Planck Institute for Software Systems)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studies how prompt context influences large language models' selection of different algorithms in code generation, leading to performance and safety differences in the implementation of the same functionality.
The Label Horizon Paradox: Rethinking Supervision Targets in Financial Forecasting
Chen-Hui Song (E Fund Management Co., Ltd.), Liyuan Chen (E Fund Management Co., Ltd.)
OptimizationRecurrent Neural NetworkTransformerContrastive LearningTime SeriesFinance Related
🎯 What it does: This paper rethinks the supervised objective in financial forecasting, finding that alignment between training labels and inference targets is not always optimal, and proposes a bi-level optimization framework with adaptive label duration to find the best supervisory signal.
The Labyrinth and the Thread: Rethinking Regularizations in Sequential Knowledge Editing for Large Language Models
Zheng Wang (Bosch Center for Artificial Intelligence), Xiaonan Lu (Bosch Center for Artificial Intelligence)
OptimizationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextSequentialBenchmark
🎯 What it does: This paper studies sequential knowledge editing in large language models, proposes a theoretical framework that establishes the equivalence between single-edit and sequential editing, reveals the fundamental reasons behind editing stability, and systematically evaluates common regularization techniques.
The Latent Color Subspace: Emergent Order in High-Dimensional Chaos
Mateusz Pach (Technical University of Munich), Zeynep Akata (Technical University of Munich)
GenerationRepresentation LearningTransformerDiffusion modelFlow-based ModelAuto EncoderImage
🎯 What it does: Analyze the color in the latent space of FLUX VAE and propose a training-free color control method based on the Latent Color Subspace.
The Latent Guardian: Defending Collaborative Perception via Feature-Level Consistency Verification
Zhuangzhuang Zhang (City University of Hong Kong), Jianping Wang (City University of Hong Kong)
Anomaly DetectionAutonomous DrivingFederated LearningSafty and PrivacyAdversarial AttackConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningOptical FlowImageVideoPoint CloudBenchmark
🎯 What it does: Propose the Cerberus framework, which detects and defends against adversarial attacks by performing multi-dimensional consistency verification on intermediate feature maps within collaborative perception (CP) systems.
The Lie We Tell: Correcting the Euclidean Fallacy in Vision Language Action Policies via Score Matching on Tangent Space
Bing-Cheng Chuang (National Taiwan University), Chun-Yi Lee (National Taiwan University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerVision-Language-Action ModelDiffusion modelScore-based ModelImageVideoTextMultimodalityStochastic Differential Equation
🎯 What it does: Proposed the Lie Diffuser Actor, which achieves intrinsic diffusion-based action generation on SE(3) through tangent space score matching, correcting traditional Euclidean errors;
The Loss Is Not Enough: Sampling Conditions and Inductive Bias in Contrastive Representation Learning
Justinas Zaliaduonis (Technical University of Munich), Sergios Gatidis (Stanford University)
Representation LearningConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderContrastive LearningImage
🎯 What it does: This paper proposes a diversity condition in contrastive learning. Under this framework, it analyzes the impact of the InfoNCE loss on the identifiability of the latent space geometry when the positive sample sampling satisfies or violates this condition, and presents a variant of InfoNCE with support correction to restore orthogonal (isometric) recovery; meanwhile, the theoretical predictions are validated through synthetic spherical data and CIFAR-10 experiments.
The Mechanistic Emergence of Symbol Grounding in Language Models
Shuyu Wu (University of Michigan), Joyce Chai (University of Michigan)
Explainability and InterpretabilityRepresentation LearningRecurrent Neural NetworkSpiking Neural NetworkTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: By constructing a minimized experimental environment (splitting words into environmental representation ⟨ENV⟩ and language representation ⟨LAN⟩) and training Transformer, Mamba-2, and LSTM on this basis, we systematically investigate how language models establish symbolic-meaning associations through internal mechanisms without explicit multimodal supervision; the conclusions are subsequently validated on real visual dialogue data (Visual Dialog).