These 1032 ICML 2026 papers come with a code repository. Each shows an AI one-line summary below β get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every ICML 2026 paper, free trial on arXivSub.
$\texttt{PRISM}$:A 3D Probabilistic Neural Representation for Interpretable Shape Modeling
Yining Jiao (University of North Carolina at Chapel Hill), Marc Niethammer (University of California San Diego)
π― What it does: Built PRISM, a probabilistic shape model based on implicit neural representations, which can generate conditional distributions of shapes given continuous covariates (e.g., age) and estimate confidence intervals for individuals' intrinsic developmental time and spatial variation.
π― What it does: Proposed the 3D-DLP model, achieving self-supervised 3D object-centric scene representation learning, which can decompose RGB-D or voxel observations into 3D latent particles and support scene reconstruction and editing.
π― What it does: Lightweight self-supervised fine-tuning of visual foundation models is performed using point tracking, differentiable sorting, and a teacher-student framework in videos, thereby enhancing their 3D spatial consistency and geometric understanding.
CodeRepresentation LearningReinforcement Learning from Human FeedbackReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: Propose the 3ViewSense framework, introducing simulated and reasoned orthogonal projection perspectives into Vision-Language Models (VLM) to address the 'spatial intelligence gap' in spatial reasoning.
π― What it does: Propose Wasserstein Lagrangian Mechanics (WLM) and design an algorithm for learning second-order collective dynamics from time snapshots;
A Cartesian-3j Framework for Machine Learning Interatomic Potentials
Zemin Xu (ShanghaiTech University), Peijun Hu (ShanghaiTech University)
CodeRepresentation LearningGraph Neural NetworkContrastive LearningGraphBenchmarkPhysics Related
π― What it does: This paper constructs a machine learning atomic potential framework based on irreducible Cartesian tensors (ICT), and implements ICT multiplication and contraction on e3nn.
A Computational Framework for Evaluating Human-likeness in LLMs' Open-ended Human Behaviors
Yuxuan Lei (University of Science and Technology of China), Xing Xie (Microsoft Research Asia)
CodeRecommendation SystemData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextSequentialRetrieval-Augmented Generation
π― What it does: A framework based on distributed evaluation was constructed to measure the realism and credibility of LLMs in simulating human behavior using large-scale network behavioral data.
CodeExplainability and InterpretabilityTransformerDiffusion modelAuto EncoderContrastive LearningImage
π― What it does: A distribution perspective for visual mechanism interpretation was constructed, and the KL minimum soft constraint principle and energy-guided diffusion posterior sampling method were proposed;
π― What it does: Proposed a fully first-order based differentiable optimization layer (FFOLayer), which approximates the supergradient through bi-level optimization and active set Lagrangian approximation.
CodeExplainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose a minimalist automated theorem proving agent that supports iterative proof refinement, memory management, and tool calling, facilitating systematic comparisons across different AI reasoners.
A Robust Optimization Guided Pruning Framework for Vision and Large Language Models
Gabriel Afriat (MIT), Rahul Mazumder (MIT)
CodeOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelAuto EncoderContrastive LearningImageText
π― What it does: Proposes a robust optimization guided one-shot pruning framework called RobOP, which can sparsify Vision Transformers and large language models (LLMs) without retraining.
π― What it does: Propose a foundational model that uses simultaneously recorded ECG and respiration signals to generate physiological supervision indicators for multimodal contrastive learning on PPG signals, and performs inference using only PPG after training.
A Studentized Spherical HarmonicsβBased Nonparametric Two-Sample Test for Compositional and Directional Data
Binglin Li (North Carolina A&T State University), Seong-Tae Kim (North Carolina A&T State University)
CodeData SynthesisAnomaly DetectionRepresentation LearningDiffusion modelScore-based ModelContrastive LearningTabularBiomedical DataReview/Survey Paper
π― What it does: A studentized nonparametric two-sample test method based on spherical harmonics is proposed, which can be simultaneously applied to compositional data and directional data, without the need for logarithm ratio transformation or handling of zero values.
π― What it does: Studied the distributional simplification bias of diffusion models when learning data statistical distributions, and constructed a analytically tractable mixed cumulative model and a single-neuron denoiser to theoretically analyze the learning dynamics;
A Unifying Relational Perspective on Expressive Lottery Tickets
Lorenz Kummer (University of Vienna), Nils Morten Kriege (University of Vienna)
CodeComputational EfficiencyRepresentation LearningDrug DiscoveryGraph Neural NetworkContrastive LearningGraphTime SeriesBiomedical Data
π― What it does: Studied the impact of sparse parameters on the expressive power of multi-relational and temporal graph neural networks, and proposed the Relational Strong Expressive Lottery Ticket Hypothesis (RSELTH), proving the existence of sparse subnetworks that retain the 1-RWL expressive power.
A3: an Analytical Low-Rank Approximation Framework for Attention
Jeffrey T. H. Wong (Imperial College London), Yiren Zhao (Imperial College London)
CodeCompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelAuto EncoderText
π― What it does: Propose the A3 low-rank approximation framework, which decomposes the Transformer into three functional components: QK, OV, and MLP, and provides closed-form solutions to reduce the hidden dimension and compress the model.
Accordion-Thinking: Self-Regulated Step Summaries for Efficient and Readable LLM Reasoning
Zhicheng Yang (Hong Kong University of Science and Technology (Guangzhou)), Jing Tang (Hong Kong University of Science and Technology)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: Propose the Accordion-Thinking framework, enabling LLMs to dynamically generate and use concise step summaries during the reasoning process, thereby significantly reducing context length and computational cost while maintaining reasoning accuracy.
ACON: Optimizing Context Compression for Long-horizon LLM Agents
Minki Kang (KAIST), Saravan Rajmohan (Microsoft)
CodeCompressionOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
π― What it does: To address the context expansion problem in long-term LLM agents, the ACON framework is proposed, which utilizes natural language optimization compression criteria to achieve context compression.
ActiveScope: Actively Seeking and Correcting Perception for MLLMs
Yajing Wang (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Propose the training-free ActiveScope framework, which enhances the performance of multi-modal large language models in high-resolution fine-grained visual understanding through active localization and self-correction.
AdaHC: Accelerating Multi-Token Prediction with Adaptive Head Chunking with Pipeline Parallelism
Yan Wang (University of Chinese Academy of Sciences), Weile Jia (University of Chinese Academy of Sciences)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose the AdaHC framework, which utilizes adaptive head chunking and pipeline parallelism to accelerate the multi-token prediction (MTP) module in LLM training.
Adalina: Adaptive Linear Approximation for the Shapley Value and Beyond
Weida Li (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
CodeExplainability and InterpretabilityComputational EfficiencyImageTabular
π― What it does: Proposed an adaptive linear approximation algorithm called Adalina, which achieves linear time and linear space approximation for half-values (including Shapley values, Banzhaf values, etc.) under the Ξ(n) space constraint;
π― What it does: This paper proposes an expandable Synergistic Plasticity Unit (SPU) for DNA sequence modeling. The SPU integrates local motifs, global dependencies, and frequency-domain periodic signals through multi-layer plasticity mechanisms, constructing an efficient DNA foundation model (SPU-DNA).
Adaptive Generation of Bias-Eliciting Questions for LLMs
Robin Staab (ETH Zurich), Martin Vechev (ETH Zurich)
CodeGenerationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose a framework for automatically generating realistic open-ended questions based on contrastive variations to trigger biased behaviors in LLMs;
Adaptive Multiscale Binary Expansion Tests for Independence
Yang Yang (University of Illinois Chicago), Ping-Shou Zhong (University of Illinois Chicago)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyContrastive LearningTabularBiomedical DataBenchmark
π― What it does: This paper proposes a novel multi-scale independence test method based on binary expansion (CoBET, dCoBET, wa-dCoBET), which can test the independence between multivariate random variables without relying on kernel functions.
Adaptive Probe-based Steering for Robust LLM Jailbreaking
Junxi Chen (Sun Yat Sen University), Xiaohua Xie (Sun Yat Sen University)
CodeExplainability and InterpretabilityComputational EfficiencyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
π― What it does: This paper proposes an adaptive probe-based steering vector to break aligned LLMs, significantly improving the effectiveness of the attack without additional contrastive prompts or tedious manual parameter tuning.
Adaptive Time Series Reasoning via Segment Selection
Shvat Messica (Harvard Medical School), Marinka Zitnik (Harvard Medical School)
CodeExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextTime SeriesBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Solve time series reasoning tasks through a controller-reasoner architecture that adaptively selects and reasons about time periods, allowing the model to actively retrieve relevant time periods and generate answers during inference.
AdaS: Adaptive Gradient Descent for Spiking Transformers
Zijian Zhou (University of Electronic Science and Technology of China), Haizhou Li (Shenzhen Loop Area Institute)
CodeOptimizationSpiking Neural NetworkTransformerImageVideoTextBiomedical Data
π― What it does: Proposed AdaS, an adaptive gradient descent optimizer specifically designed for Spiking Transformers, aiming to alleviate the excessive noise problem during the training process with surrogate gradients.
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: Proposed a sparse attention mechanism called ADASPLASH-2 based on Ξ±-entmax, significantly accelerating the forward and backward computations of Transformers.
Advancing Analytic Class-Incremental Learning through Vision-Language Calibration
Binyu Zhao (Harbin Institute of Technology), Ivor Tsang (Agency for Science, Technology and Research)
CodeClassificationComputational EfficiencyRepresentation LearningMeta LearningTransformerReinforcement LearningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: This study proposes a framework called VILA, which is based on pre-trained models and is designed for analytical class-incremental learning, achieving fast and efficient continuous learning through a dual-branch visual-language calibration mechanism.
AdverMCTS: Combating Pseudo-Correctness in Code Generation via Adversarial Monte Carlo Tree Search
Qingyao Li (Shanghai Jiao Tong University), Bo An (Nanyang Technological University)
CodeExplainability and InterpretabilityComputational EfficiencyAdversarial AttackAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextSequentialRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the ADVERMCTS framework, which addresses the issue of pseudo-correctness in code generation through dual MCTS adversarial search by Solver and Attacker.
π― What it does: Proposes a gray-box attack framework MeLA (Meta-Laplacian Attack) for hypergraph neural networks and its corresponding robust training method MeLA-D, which can simultaneously perform low-budget perturbations on the hypergraph structure and node features, and uses the Laplacian operator as the meta-objective.
Adversarial Dual On-Policy Distillation from Expressive Teacher
Zhenglin Wan (National University of Singapore), Yang You (National University of Singapore)
CodeKnowledge DistillationRobotic IntelligenceReinforcement Learning from Human FeedbackSpiking Neural NetworkTransformerReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelRectified FlowGenerative Adversarial NetworkContrastive LearningTabularTime SeriesSequential
π― What it does: Propose a new learning from demonstration method called FA-OPD, which uses Flow Matching (FM) as a co-trainable teacher during the learning process. It employs a dual-channel (reward and action) self-supervised training on a student (lightweight MLP) to accomplish robotic control tasks without requiring real environment rewards.
AesFormer: Transform Everyday Photos into Beautiful Memories
Tianxiang Du (Peking University), Yuxin Peng (Peking University)
CodeImage TranslationRestorationGenerationTransformerLarge Language ModelReinforcement LearningVision Language ModelVision-Language-Action ModelFlow-based ModelImageVideoText
π― What it does: Propose a two-stage framework called AesFormer, which achieves the aesthetic photo reconstruction task by first planning aesthetic actions and then executing structural edits.
π― What it does: This paper proposes an affine equivariant Gaussian kernel space encoding (EKS) for achieving editable and physics-driven neural radiance field rendering.
Agent-Omit: Adaptive Context Omission for Efficient LLM Agents
Yansong Ning (Hong Kong University of Science and Technology), Hao Liu (Hong Kong University of Science and Technology)
CodeComputational EfficiencyKnowledge DistillationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the Agent-Omit framework, which achieves efficient reasoning for LLM agents by adaptively omitting redundant thoughts and observations in multi-round interactions.
AgentExpt: Automating AI Experiment Design with LLM-based Resource Retrieval Agent
Yu Li (Tsinghua University), Yong Li (Tsinghua University)
CodeRecommendation SystemExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This paper constructs the AgentExpt framework, which utilizes a large-scale paper-baseline-dataset knowledge base to automatically recommend experimental baselines and datasets.
Agentic Model Predictive Questioning Control in Visual Design
Kuang-Da Wang (National Yang Ming Chiao Tung University), Shingo Takamatsu (Sony Group Corporation)
CodeGenerationOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringImageTextMultimodalityRetrieval-Augmented Generation
π― What it does: This paper proposes an agent-based model predictive query control (A-MPQC) for visual design, which enhances alignment between design and user intent and reduces cognitive load through multi-round clarifications under a fixed query budget.
AgentVocab: Structure-Aware Vocabulary Adaptation for Efficient LLM Agents
Kai Bian (Harbin Institute of Technology), Xuelong Li (TeleAI of China Telecom)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: This paper proposes the AgentVocab framework, which in the tool calling scenarios of LLM agents, mines structured and semantic fragments from actual calling logs to expand the vocabulary, thereby reducing tokenization fragmentation and improving decoding efficiency.
CodeRecommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AITextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the AgentWebBench benchmark to evaluate user agents in the Agentic Web environment for completing four categories of information retrieval and generation tasks in collaboration with multiple content agents.
Aggregate Models, Not Explanations: Improving Feature Importance Estimation
Joseph Paillard (F. Hoffmann-La Roche Ltd), Bertrand Thirion (Universite Paris-Saclay)
CodeExplainability and InterpretabilityData-Centric LearningTabularBiomedical Data
π― What it does: This study explores the differences between model-level integration (integrating predictors) and explanation-level integration (averaging feature importance from individual models) in estimating feature importance, and validates the performance differences through theoretical derivation and experiments.
AI Engram: In Search of Memory Traces in Artificial Intelligence
Jea Kwon (Max Planck Institute for Security and Privacy), Meeyoung Cha (Max Planck Institute for Security and Privacy)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImageText
π― What it does: Propose a geometric framework based on four principles of neuroscience, defining AI Engram and deriving a closed-form spectral projection estimator to identify and manipulate operable memory traces in deep networks.
AICrypto: Evaluating Cryptography Capabilities of Large Language Models
Yu Wang (Chinese Academy of Sciences), Tianxing He (Tsinghua University)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the AICrypto benchmark to evaluate the capabilities of large language models (LLMs) in cryptography, containing 135 multiple-choice questions, 150 CTF challenges, and 30 proof questions;
AIR-VLA: Vision-Language-Action Systems for Aerial Manipulation
Jianli Sun (Chinese Academy of Sciences), Yonglin Tian (Chinese Academy of Sciences)
CodeRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelDiffusion modelImageVideoTextMultimodalityBenchmark
π― What it does: Proposed the AIR-VLA benchmark, combining a simulation environment, 3000 multimodal data samples, and multidimensional evaluation, specifically targeting aerial manipulation systems.
CodeSafty and PrivacyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed and implemented the AIR (Agent Incident Response) framework, achieving event detection, isolation, recovery, and eradication during the execution of LLM agents. It automatically generates guardrail rules using a DSL to enhance agent security.
AIR: Post-training Data Selection for Reasoning via Attention Head Influence
Jinrui Liu (Beihang University), Chongyang Tao (Beihang University)
CodeExplainability and InterpretabilityKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Proposed an unsupervised, no-training post-training data selection framework called AIR, which selects the most valuable data for inference by examining the causal impact of the retrieval attention head.
π― What it does: Propose a self-supervised method based on flow feature distance to align the trajectory tangent of the consistency model, thereby significantly improving the training convergence speed and generation quality;
AlignedNorm: Prompting VisionβLanguage Models via Coupled Prompt Field
Qi Ma (Nankai University), Deng-Ping Fan (Nankai University)
CodeClassificationRecognitionDomain AdaptationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Proposed the Coupled Prompt Field (CPF) paradigm and applied AlignedNorm to achieve norm alignment in prompt learning for vision-language models such as CLIP, addressing the issue of local optima between baselines and new tasks.
All Circuits Lead to Rome: Rethinking Functional Anisotropy in Circuit and Sheaf Discovery for LLMs
Xi Chen (University of Toronto), Gerald Penn (University of Toronto)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
π― What it does: Proposed and validated the Functional Anisotropy Hypothesis, showing that a single LLM task can be implemented by multiple circuits/fibers with significant structural differences and low overlap;
π― What it does: To address the observation bias problem in imbalanced missing multi-view clustering, the CIMLN framework is proposed for missing view recovery and clustering.
AlphaRouter: Token-level Routing Between SLM and LLM with Reinforcement Learning and Tree Search
Siteng Liao (Beijing Normal University), Tian Wang (Beijing Normal University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Designed a Token-Level routing framework called AlphaRouter based on Small Language Models (SLM) and Large Language Models (LLM), which learns the optimal collaborative inference path using search and reinforcement learning.
Ivri Hikri (Technion - Israel Institute of Technology), Nir Rosenfeld (Technion - Israel Institute of Technology)
CodeClassificationSafty and PrivacyExplainability and InterpretabilityReinforcement LearningContrastive LearningTabularBenchmark
π― What it does: This paper studies how to influence users' strategic behavior and model performance by controlling the uncertainty (ambiguity) of classifiers, in scenarios where the learning system can only disclose partial classifier information.
π― What it does: Propose a fuzzy task for 3D visual localization, and build an interactive multi-round question-answering framework to eliminate referential ambiguity and accurately locate the target object.
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Proposed AMDP, an asynchronous multi-direction pipeline parallel training framework, which limits the parameter mismatch between forward and backward passes to one step, significantly improving model training throughput.
Amortized Variational Inference for Partial-Label Learning: A Probabilistic Approach to Label Disambiguation
Tobias Fuchs (Karlsruhe Institute of Technology), Nadja Klein (Karlsruhe Institute of Technology)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerAuto EncoderContrastive LearningImageTextTabular
π― What it does: Propose the VILLP framework, modeling label uncertainty in partial label learning as a variational inference problem, using neural networks to predict variational parameters, achieving a direct approximation of the true label posterior.
π― What it does: Propose an EM algorithm based on Polya-Gamma data augmentation to jointly estimate user ability ΞΈ and item difficulty Ξ² in the Rasch model, addressing the inconsistency or slow convergence issues of traditional JMLE, MMLE, and CMLE on sparse data.
π― What it does: Proposes the definition of deep learning model cloning and a detection method called MCDetector based on dual similarity (structural and weight similarity).
Fabian Fumagalli (LMU Munich), R. Teal Witter (Claremont McKenna College)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerContrastive LearningImageTextTabular
π― What it does: Proposed an OddSHAP, a Shapley value approximation estimator based on odd subspaces, which achieves consistent and efficient estimation by utilizing paired sampling and Fourier basis functions.
Analytic Bijections for Smooth and Interpretable Normalizing Flows
Mathis Gerdes (University of Amsterdam), Miranda C. N. Cheng (University of Amsterdam)
CodeExplainability and InterpretabilityComputational EfficiencyDiffusion modelScore-based ModelFlow-based ModelRectified FlowContrastive LearningImageTabularTime SeriesSequentialBiomedical DataPhysics Related
π― What it does: Designed and implemented three globally smooth, closed-form invertible scalar bijections (cubic rational, sinh, cubic conjugation), and constructed interpretable radial flows based on them as pluggable invertible transformations for coupling and other normalization flows.
π― What it does: This study systematically evaluates 14,664 experiments to explore the impact of plastic interventions in DRL (Deep Reinforcement Learning) on backdoor attacks injected later, and reveals its internal mechanisms through pathological feature analysis.
CodeAnomaly DetectionExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMultimodalityTime SeriesChain-of-Thought
π― What it does: Propose the ANOMSEER framework, which utilizes reinforcement learning to enable multi-modal large language models to perform fine-grained reasoning in time series anomaly detection and generate interpretable diagnostic reports.
π― What it does: Propose the Any2Any framework to achieve arbitrary cross-modal translation of remote sensing multi-modal images, and construct a million-scale RST-1M training set.
Xin Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
CodeRetrievalRepresentation LearningTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: Proposes AOEB (Agent-Oriented Embedding Benchmark), a multi-task, multi-modal evaluation benchmark specifically designed for the retrieval needs of LLM agents;
π― What it does: Propose a lightweight prompt-tuning framework named AOEPT, aimed at overcoming the implicit modal reduction bottleneck (IMR) in multi-modal Transformers under missing modal conditions, and restoring the inference scope of missing modalities by injecting modal contextualized prompts (MCP) into the model.
APE-Bench: Evaluating Automated Proof Engineering for Formal Math Libraries
Huajian Xin (ByteDance Seed), Wenda Li (University of Edinburgh)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed the Automated Proof Engineering (APE) framework, which includes automated task extraction (APE-Bench), a unified execution and evaluation infrastructure (APE-Harness), and a multi-version content deduplication and retrieval service; conducted a systematic evaluation of proof engineering tasks at the repository level on the formal mathematics library (Mathlib).
Approximate Equivariance via Projection-Based Regularisation
Torben Berndt (Heidelberg Institute for Theoretical Studies), Jan StΓΌhmer (Heidelberg Institute for Theoretical Studies)
CodeClassificationRestorationRepresentation LearningConvolutional Neural NetworkContrastive LearningImagePoint CloudBiomedical DataComputed TomographyReview/Survey Paper
π― What it does: This paper proposes a projection-based regularization framework for learning approximate equivariant neural networks. By performing an orthogonal decomposition of the network weights, the equivariant and non-equivariant components are regularized separately, thus gradually guiding the network toward equivariance while maintaining model flexibility.
Ben Hagag (Carnegie Mellon University), Sarah Scheffler
CodeSafty and PrivacyTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
π― What it does: Experimentally evaluate the impact of multi-agent system architectures on security, systematically studying the effects of role configuration, communication topology, and memory visibility on attack success rates and task performance.
ArenaRL: Scaling RL for Open-Ended Agents via Tournament-based Relative Ranking
Qiang Zhang (University of Science and Technology of China), Jiawei Liu (University of Science and Technology of China)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringContrastive LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose ArenaRL, a reinforcement learning framework for open-ended tasks, which abandons traditional point-to-point scalar rewards and instead performs process-aware pairwise evaluation among generated trajectories within the same group, and uses a tournament for relative ranking, thereby providing more stable advantage signals; simultaneously, a complete open-ended agent evaluation benchmark, Open-Travel and Open-DeepResearch, is constructed.
Artemis: Structured Visual Reasoning for Perception Policy Learning
Wei Tang (Nanjing University of Science and Technology), Zechao Li (Nanjing University of Science and Technology)
CodeObject DetectionAutonomous DrivingOptimizationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageMultimodality
π― What it does: Proposed a visual perception strategy learning framework called Artemis, which employs structured visual reasoning, outputting (label, bounding box) pairs at intermediate steps to achieve verifiable spatial reasoning; and uses verifiable rewards (GRPO) in reinforcement learning to guide the model's learning.
ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models
Jiahui Guang (Harbin Institute of Technology), Zhaoquan Gu (Harbin Institute of Technology)
CodeSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodality
π― What it does: Propose a controllable multimodal large model learning framework ASRU, which achieves fine-grained control over forgotten knowledge through activation steering and reinforcement learning, maintaining generation quality and model utility.
Michael Shalyt (Technion Israel Institute of Technology), Ido Kaminer (Technion Israel Institute of Technology)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkPhysics RelatedRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes a high-resolution symbolic mathematics operation benchmark called ASyMOB, containing 35,368 university-level symbolic integration, limit, differential equation, series, and hypergeometric problems, and generates a large number of variants through systematic symbolic, numerical, and equivalent transformations, specifically designed to evaluate the true reasoning ability of LLMs in symbolic reasoning rather than memorization patterns.
π― What it does: Studied an attention-based multi-layer fusion method (ALF) that dynamically fuses CLS and average pooling (AP) features from all layers in a visual Transformer, and performed linear probing on frozen pre-trained models.
Auditing Sybil: Explaining Deep Lung Cancer Risk Prediction Through Generative Interventional Attributions
Bartlomiej Sobieski (University of Warsaw), Przemyslaw Biecek (University of Warsaw)
CodeAnomaly DetectionExplainability and InterpretabilityDiffusion modelScore-based ModelAuto EncoderImageBiomedical DataComputed Tomography
π― What it does: Propose the S(H)NAP framework to perform causal intervention-based auditing on the Sybil lung cancer risk prediction model, using 3D diffusion bridge to generate lung nodule insertion/deletion, and construct SHNAP and SNAP explanation methods.
π― What it does: Propose the ACC-MARL framework in multi-task, multi-agent reinforcement learning, using DFA to represent tasks, and achieve solutions for historical dependency, credit assignment, and representation bottleneck through DFA progress, potential reward shaping, and pre-trained RAD embeddings, thereby learning decentralized policies that can perform optimal task allocation at test time.
Automatic Layer Selection for Hallucination Detection
Xinpeng Wang (University of Virginia), Zhe Zeng (University of Virginia)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAuto EncoderContrastive LearningText
π― What it does: This paper studies automatically selecting suitable layers for hallucination detection under the hidden state probing framework, and proposes two new methods, FEPoID and FST;
Automatic Pruning Discovery for Large Language Models
Haidong Kang (Northeastern University), Hao Wang (Xidian University)
CodeOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Automatically prune models without expert knowledge by generating and optimizing sparsification rules using large language models themselves.
AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning
Jiaru Zou (Princeton University), Mengdi Wang (Princeton University)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the AutoTool framework, enabling large language models to dynamically select and integrate tools during the reasoning process, and supporting the evolution of the toolset over time.
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: This paper proposes an agent framework based on large language models (LLMs) to address the dead-end problem in knowledge graph question answering caused by mismatches between queries and graph structures. It introduces a 'Backjump' mechanism, allowing LLMs to retreat to historical nodes and re-explore alternative paths during reasoning.
CodeOptimizationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsAuto EncoderContrastive LearningTextBenchmark
π― What it does: This paper studies the problem of different optimal point condition numbers caused by LoRA over-parameterization, and proposes Balanced LoRA (BaLoRA) to accelerate convergence by projecting to the balanced manifold after each iteration.
BARRED: Synthetic Training of Custom Policy Guardrails via Asymmetric Debate
Arnon Mazza (Plurai Inc), Elad Levi (Plurai Inc)
CodeClassificationSafty and PrivacyExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: By leveraging task descriptions and a few unannotated samples, high-quality synthetic training data is automatically generated using dimension decomposition and multi-agent debate to train custom guardian models.
Base Models Know How to Reason, Thinking Models Learn When
Constantin Venhoff (University of Oxford), Neel Nanda
CodeExplainability and InterpretabilityKnowledge DistillationTransformerSupervised Fine-TuningReinforcement LearningAuto EncoderTextChain-of-Thought
π― What it does: This paper proposes an unsupervised method that uses sparse autoencoders to extract the reasoning mechanisms of thinking language models, and explains the content learned under different training paradigms (RL vs. SFT-distillation) by constructing model differences (category vectors + heuristic of when they are activated).
Peng Cui (Mohamed bin Zayed University of Artificial Intelligence), Lijie Hu (Mohamed bin Zayed University of Artificial Intelligence)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningContrastive LearningImage
π― What it does: Propose Bayesian Gated Non-negative Contrastive Learning (BayesNCL), which dynamically suppresses shared background features through a variational Bayesian gating mechanism, addressing optimization conflicts in traditional contrastive learning, thus achieving interpretable sparse representations.
Bayesian Rain Field Reconstruction using Commercial Microwave Links and Diffusion Model Priors
Badr MOUFAD, Eric Moulines (MBZUAI)
CodeRestorationTransformerDiffusion modelScore-based ModelTabularTime Series
π― What it does: This paper proposes using commercial microwave links (CML) for rainfall field reconstruction, employing a Bayesian inverse problem framework combined with a diffusion model prior.
Bayesian-LoRA: Probabilistic Low-Rank Adaptation of Large Language Models
Moule Lin (Trinity College Dublin), Goetz Botterweck (Trinity College Dublin)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningFlow-based ModelText
π― What it does: In the parameter-efficient fine-tuning of large-scale language models, the deterministic low-rank updates of LoRA are transformed into a probabilistic low-rank adaptation framework. By utilizing variational inference with sparse Gaussian processes, uncertainty is injected into the low-rank subspace of LoRA, and self-consistent calibrated learning is achieved through regularization and flow transformations.
BEDTime: A Unified Benchmark for Automatically Describing Time Series
Medhasweta Sen (University of Virginia), Thomas Hartvigsen (University of Virginia)
CodeRecognitionGenerationAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelTextMultimodalityTime SeriesBenchmarkChain-of-Thought
π― What it does: Proposed the BEDTime benchmark to evaluate models on tasks involving identification, discrimination, and generation of structural descriptions for univariate time series, and conducted systematic comparisons using a unified multi-task, cross-modal evaluation framework.
Being More Lightweight and Practical: Mini-sized Contrastive Learning Pre-trained Models for Fine-grained Traffic Task
Shuhao Li (Fudan University), Fan Zhang (Guangzhou University)
CodeAutonomous DrivingComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerContrastive LearningGraphTime Series
π― What it does: Proposes MiniTraffic, a lightweight pre-training framework that pre-trains using road-level data and transfers to lane-level fine-grained traffic prediction tasks;
Benchmarking and Enhancing VLM for Compressed Image Understanding
Zifu Zhang (Tsinghua University), Yan Wang (Tsinghua University)
CodeCompressionComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerPrompt EngineeringVision Language ModelImageMultimodalityBenchmark
π― What it does: Constructed a VLM benchmark covering over 1M compressed images, 11 types of bitstreams, and 7 evaluation metrics, and proposed a lightweight adapter to enhance the understanding ability of VLMs for compressed images.
Benchmarking and Evolving Reason-Reflect-Rectify for Reflective Visual Generation
Junjie Wang (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
π― What it does: Propose R3-Bench benchmark and R3-Refiner framework to evaluate and improve the Reason-Reflect-Rectify reflection-correction loop in visual generation
π― What it does: Constructed a large-scale reproducible ICRL4AHT benchmark to evaluate the adaptability of Transformer-based ICRL in Ad-Hoc team collaboration.
Beyond Additive Decompositions: Interpretability Through Separability
Jinyang Liu (University of Copenhagen), Munir Hiabu (University of Copenhagen)
CodeExplainability and InterpretabilityTabular
π― What it does: Proposed and implemented Tensor Separation Learning (TSL), an interpretable regression model based on separable product differentiation.
Beyond Blind Noising: Disentangled Visual Rectification for Hallucination Mitigation in MLLMs
Yujia Chen (University of Science and Technology of China), Wenzhang SUN (Tsinghua University)
CodeRestorationAnomaly DetectionExplainability and InterpretabilityRepresentation LearningTransformerPrompt EngineeringVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
π― What it does: To address the image hallucination problem in multi-modal large language models (MLLM), the authors propose a training-free, plug-and-play dual-stream framework called Disentangled Visual Rectification (DVR). The framework utilizes the response differences of LIP (e.g., CLIP, SigLIP) and SSL (e.g., DINOv3) encoders under noise perturbations. It first adaptively suppresses or enhances the original features in the visual encoding layer, and then further weakens the residual hallucination-inducing components through a contrastive mechanism in the decoding layer.
π― What it does: This paper proposes the Unbalanced SchrΓΆdinger Bridge (USB) framework, which utilizes the branching SchrΓΆdinger bridge theory and unbalanced score matching to simultaneously infer trajectories of stochasticity and imbalanced mass changes in single-cell snapshot data, and supports discrete cell birth/apoptosis simulations.
Beyond Correctness: Distance-Based Social Dynamics of Multi-Agent Debate
Seungwoong Ha (Santa Fe Institute), Melanie Mitchell (Santa Fe Institute)
CodeExplainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Studied the microdynamics of answers in multi-agent debate systems, analyzing distance-aware revisions and convergence behaviors in social interactions using the ConceptARC two-dimensional grid task.
π― What it does: Proposed a general semi-supervised learning framework called SAGE, which is designed for unknown, arbitrarily distributed, and extremely few labeled samples. The core idea is to replace distribution estimation with structural reasoning;
Beyond Drift: Stabilizing Subjective LLM Evaluation with Information-Theoretic Rubrics
Wang Xu (HeFei University of Technology), Qian Wan (Central China Normal University)
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: Construct a questionnaire-based evaluation framework based on Expected Information Gain (EIG) to address the dimension drift problem in subjective evaluation of LLMs.
π― What it does: Proposes a learnable density control framework called LeGS based on reinforcement learning for 3D Gaussian Splatting scene reconstruction.