ICLR 2026 Papers — Page 16
International Conference on Learning Representations · 5356 papers
Entering the Era of Discrete Diffusion Models: A Benchmark for Schrödinger Bridges and Entropic Optimal Transport
Xavier Aramayo Carrasco (Applied AI Institute), Alexander Korotin (Applied AI Institute)
OptimizationDiffusion modelBenchmarkPhysics Related
🎯 What it does: This paper proposes a benchmark for the Schrödinger Bridge and Entropic Optimal Transport (EOT) in discrete space, utilizing analytical SB solutions to construct discrete distribution pairs, and develops and evaluates three new algorithms (DLightSB, DLightSB-M, α-CSBM) on this benchmark.
Entropic Confinement and Mode Connectivity in Overparameterized Neural Networks
Luca Di Carlo (Princeton University), David J. Schwab (The Graduate Center, CUNY)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: Analyzes the curvature changes along low-loss paths in over-parameterized neural networks and reveals that entropy potential barriers caused by curvature restrict optimization dynamics to local basins.
Entropy Regularizing Activation: Boosting Continuous Control, Large Language Models, and Image Classification with Activation as Entropy Constraints
Zilin Kang (Shanghai Qi Zhi Institute), Huazhe Xu (Shanghai Qi Zhi Institute)
Convolutional Neural NetworkTransformerLarge Language ModelReinforcement LearningImageText
🎯 What it does: Proposed and implemented the ERA (Entropy Regularizing Activation) mechanism, which adds a specially designed activation function at the model's final layer to impose minimal constraints on policy entropy, thereby enhancing performance in continuous control, image classification, and reinforcement learning for large language models.
Entropy-Based Block Pruning for Efficient Large Language Models
Liangwei Yang (Salesforce Ai Research), Shelby Heinecke (Salesforce Ai Research)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes EntroDrop, a block pruning method based on the entropy increase of hidden representations, aimed at reducing the computational and storage requirements of large language models.
Entropy-Guided Dynamic Tokens for Graph-LLM Alignment in Molecular Understanding
Zihao Jing (Western University), Pingzhao Hu (Western University)
Drug DiscoveryGraph Neural NetworkTransformerLarge Language ModelContrastive LearningGraph
🎯 What it does: Developed EDT-Former, a bridge connecting chemical graphs with large language models (LLMs), leveraging entropy-guided subgraph partitioning and dynamic query Transformers to align graphs with LLMs without requiring LLM fine-tuning.
Entropy-Monitored Kernelized Token Distillation for Audio-Visual Compression
Hyoungseob Park (Yale University), Alex Wong (Yale University)
CompressionKnowledge DistillationTransformerMultimodalityAudio
🎯 What it does: Proposes an audio-visual knowledge distillation method based on kernelized token distillation and entropy monitoring, aimed at compressing large audio-visual models while maintaining their performance.
Entropy-preserving reinforcement learning
Aleksei Petrenko (Apple), Philipp Kraehenbuehl (Apple)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed and verified a series of reinforcement learning methods (REPO, ADAPO) that control entropy, ensuring language models maintain sufficient exploration during reinforcement learning training;
EntropyLong: Effective Long-Context Training via Predictive Uncertainty
Junlong Jia (Beihang University), GuoBinghui
Computational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented a method for constructing long-context training data based on model prediction entropy verification, named EntropyLong.
EnvSocial-Diff: A Diffusion-Based Crowd Simulation Model with Environmental Conditioning and Individual-Group Interaction
Bingxue Zhao (Shenzhen University), Hui Huang (Shenzhen University)
GenerationData SynthesisConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkTransformerDiffusion modelTime SeriesSequential
🎯 What it does: Proposes EnvSocial-Diff, a diffusion-based model inspired by social physics for crowd trajectory prediction, which explicitly integrates environmental conditions and individual-group interactions;
Epistemic Uncertainty Quantification To Improve Decisions From Black-Box Models
Sébastien Melo (Inria Saclay), Marine Le Morvan (Inria Saclay)
Explainability and InterpretabilityLarge Language ModelTextTabularBenchmark
🎯 What it does: This paper proposes a binning-free, sample-efficient partitioning estimator for accurately quantifying the empirical uncertainty and decision risk of black-box models;
EquAct: An SE(3)-Equivariant Multi-Task Transformer for 3D Robotic Manipulation
Xupeng Zhu (Northeastern University), Robert Platt (Northeastern University)
Robotic IntelligenceConvolutional Neural NetworkTransformerReinforcement LearningVision-Language-Action ModelTextMultimodalityPoint Cloud
🎯 What it does: Propose a multi-task keyframe manipulation strategy called EquAct, which maps language instructions, point cloud observations, and actions into a shared space through SE(3) equivariant networks, achieving theoretical generalization for 3D pose variations.
Equilibrium Language Models
Yikun Jiang (Huawei Technologies Co Ltd), John C.S. Lui (Chinese University of Hong Kong)
CompressionComputational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Achieve LLM compression and adaptation for edge devices by replacing Transformer layer blocks with fixed-point networks.
Equivariant Splitting: Self-supervised learning from incomplete data
Victor Sechaud (ENS de Lyon), Julián Tachella (ENS de Lyon)
RestorationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: Proposed a self-supervised learning framework called Equivariant Splitting (ES), for image recovery using only a single incomplete measurement.
Erase or Hide? Suppressing Spurious Unlearning Neurons for Robust Unlearning
Nakyeong Yang (Seoul National University), Meeyoung Cha (Max Planck Institute for Security and Privacy)
Safty and PrivacyExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper addresses the privacy leakage issue in large language models by studying and proposing a robust unlearning method that can prevent the recovery of target knowledge during subsequent training after its deletion.
Erase to Improve: Erasable Reinforcement Learning for Search-Augmented LLMs
Kang An (SenseTime), Yichao Wu (SenseTime)
RetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Erasable Reinforcement Learning (ERL) framework, leveraging search-enhanced LLMs to detect errors, erase, and regenerate during multi-hop reasoning, significantly improving reasoning robustness.
ERGO: Efficient High-Resolution Visual Understanding for Vision-Language Models
Jewon Lee (Nota Inc), Bo-Kyeong Kim (Nota Inc)
Computational EfficiencyTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Designed and implemented a two-stage coarse-to-fine high-resolution visual reasoning pipeline, training the model via reinforcement learning to autonomously locate task-related regions for fine-grained reasoning under low-resolution inputs.
Error as Signal: Stiffness-Aware Diffusion Sampling via Embedded Runge-Kutta Guidance
Inho Kong (Korea University), Hyunwoo J. Kim (Korea Advanced Institute of Science and Technology)
GenerationDiffusion modelImageTextOrdinary Differential Equation
🎯 What it does: Proposes ERK-Guid, a method that utilizes embedded Runge-Kutta error as a guidance signal to correct local truncation errors in rigid regions during diffusion model sampling, thereby improving sampling quality and efficiency.
Error Feedback for Muon and Friends
Kaja Gruntkowska (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
CompressionOptimizationText
🎯 What it does: Propose EF21-Muon, a compressible, distributed optimizer based on non-Euclidean linear minimax oracle (LMO), integrating stochastic gradient, momentum, and error feedback to achieve communication efficiency for the Muon/Scion/Gluon family.
Error Notebook-Guided, Training-Free Part Retrieval in 3D CAD Assemblies via Vision-Language Models
Yunqing Liu (Fujitsu Research & Development Center), Zhiming Tan (Fujitsu Research & Development Center)
RetrievalTransformerVision Language ModelTextMeshRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a training-free Error Notebook + RAG framework, achieving 3D CAD assembly part retrieval using a two-stage vision-language model (VLM);
ERTACache: Error Rectification and Timesteps Adjustment for Efficient Diffusion
Xurui Peng (ByteDance Inc), Mingbao Lin (Rakuten Asia)
GenerationComputational EfficiencyDiffusion modelFlow-based ModelVideoOrdinary Differential Equation
🎯 What it does: Propose the ERTACache scheme through offline error analysis and online temporal correction, significantly accelerating the inference of diffusion models while maintaining or even enhancing generation quality
ES-dLLM: Efficient Inference for Diffusion Large Language Models by Early-Skipping
Zijian Zhu (Tsinghua University), Kaisheng Ma (Tsinghua University)
Computational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Studied computational redundancy in the inference process of diffusion large language models (dLLM), and proposed a training-agnostic early skipping framework called ES-dLLM to reduce computational cost per iteration.
Escaping Low-Rank Traps: Interpretable Visual Concept Learning via Implicit Vector Quantization
Shujian Gao (Fudan University), Junjun He (Shanghai AI Laboratory)
Explainability and InterpretabilityRepresentation LearningTransformerImage
🎯 What it does: Proposed a concept bottleneck model (CBM) integrating implicit vector quantization (IVQ) and magnet attention, addressing visual feature collapse and enabling many-to-many visual concept learning.
Escaping Model Collapse via Synthetic Data Verification: Near-term Improvements and Long-term Convergence
Bingji Yi (Independent Researcher), Haifeng Xu (University of Chicago)
GenerationData SynthesisData-Centric LearningTransformerLarge Language ModelAuto EncoderImageText
🎯 What it does: This paper investigates the effect of introducing an external verifier for synthetic data filtering during the self-training process of generative models, provides theoretical analysis based on linear regression, and verifies the theoretical predictions through experiments on VAE (MNIST) and LLM (XSUM summaries).
Escaping Policy Contraction: Contraction-Aware PPO (CaPPO) for Stable Language Model Fine-Tuning
Dun Yuan (McGill University), Xue Liu (Mohamed bin Zayed University of Artificial Intelligence)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed Contraction-Aware PPO (CaPPO), addressing the strategy contraction problem in RLHF training through multi-objective gradient updates and entropy scheduling control, and validated on various language models and datasets.
Escaping the Homophily Trap: A Threshold-free Graph Outlier Detection Framework via Clustering-guided Edge Reweighting
Yunhe Zhang (University of Macau), See-Kiong Ng (National University of Singapore)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: Propose the CER-GOD framework, which utilizes a self-discriminative masker and a threshold-free clustering detector to eliminate the homophily trap in graph convolution, thereby achieving unsupervised graph anomaly detection.
Estimating Dimensionality of Neural Representations from Finite Samples
Chanwoo Chun (Harvard University), Daniel Lee (Flatiron Institute)
Representation LearningTextBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Investigated estimating the global dimension of neural representations from limited samples and noise, proposing an unbiased participation ratio estimator.
Estimating Semantic Alphabet Size for LLM Uncertainty Quantification
Lucas Hurley McCabe (George Washington University), H Howie Huang (George Washington University)
Explainability and InterpretabilityLarge Language ModelText
🎯 What it does: To address uncertainty estimation in large language models, this paper proposes an improved method for estimating the size of the semantic alphabet and uses it to correct discrete semantic entropy (DSE), achieving more accurate entropy estimation and error detection with only a small number of samples.
Estimating Worst-Case Frontier Risks of Open-Weight LLMs
Eric Wallace (OpenAI), Chris Koch (OpenAI)
Safty and PrivacyTransformerSupervised Fine-TuningReinforcement LearningTextBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Study and evaluate the front-line risks of OpenAI gpt-oss in worst-case scenarios, enhance its capabilities in biology and cybersecurity through malicious fine-tuning (MFT), and compare it with existing open-source and closed-source models.
ETGS: Explicit Thermodynamics Gaussian Splatting for Dynamic Thermal Reconstruction
Zhongwen Wang (Nanjing University of Science and Technology), Quansen Sun (Nanjing University of Science and Technology)
Gaussian SplattingVideoTime SeriesPhysics Related
🎯 What it does: Proposed the ETGS method, which achieves dynamic thermal field reconstruction using 3D Gaussian splatting with explicit thermodynamic modeling.
EUBRL: Epistemic Uncertainty Directed Bayesian Reinforcement Learning
Jianfei Ma (National University of Singapore), Wee Sun Lee (National University of Singapore)
Reinforcement LearningBenchmark
🎯 What it does: Propose the EUBRL algorithm, which utilizes Bayesian uncertainty to guide rewards and achieve efficient exploration.
EvA: Evolutionary Attacks on Graphs
Mohammad Sadegh Akhondzadeh, Aleksandar Bojchevski (University of Cologne)
Adversarial AttackGraph
🎯 What it does: Proposed EvA, a black-box graph structure adversarial attack based on genetic algorithm, which can directly optimize accuracy and support multiple objectives.
Evaluating and Improving Cultural Awareness of Reward Models for LLM Alignment
Hongbin Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Developed and released the cross-cultural reward model benchmark CARB, systematically evaluating the cultural awareness of existing reward models and proposing improvement methods.
Evaluating Cross-Modal Reasoning Ability and Problem Characteristics with Multimodal Item Response Theory
Shunki Uebayashi, Koh Takeuchi (CyberAgent)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: This paper proposes Multimodal Multidimensional Item Response Theory (M3IRT) and its simplified version M2IRT to evaluate the cross-modal reasoning capabilities of Multimodal Large Language Models (MLLMs). By decomposing item difficulty and model ability into image, text, and cross-modal dimensions, it achieves automatic construction of high-quality subsets and significantly reduces evaluation costs.
Evaluating Data Influence in Meta Learning
Chenyang Ren (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
OptimizationExplainability and InterpretabilityComputational EfficiencyMeta LearningImage
🎯 What it does: Proposes a dual-level (task-level and instance-level) data importance assessment framework based on influence functions for quantifying and editing the contribution of training data in meta learning.
Evaluating GFlowNet from partial episodes for stable and flexible policy-based training
Puhua Niu (Texas A&M University), Xiaoning Qian (Texas A&M University)
Drug DiscoveryReinforcement LearningGraphSequential
🎯 What it does: Proposes a Sub-EB evaluation function learning method based on sub-trajectory balance to improve the policy gradient training of GFlowNet.
Evaluating Language Models' Evaluations of Games
Katherine M. Collins (University of Cambridge), Thomas L. Griffiths (Princeton University)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Studied the ability of AI systems to assess games, proposing a formal framework for game evaluation, constructing a dataset of 121 novel two-dimensional board games, and collecting approximately 450 human players' evaluations of expected payoff and fun for each game. Subsequently, used multiple language models (including non-reasoning models and chain-of-thought models) as well as explicit simulation benchmarks (such as Intuitive Gamer, MCTS, Expert, etc.) to predict game evaluations, and compared model predictions with human evaluations and approximate game-theoretic optimal values.
Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions
Yuanzhe Hu (University of California, San Diego), Julian McAuley (University of California, San Diego)
Large Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the MemoryAgentBench benchmark to evaluate the memory capabilities of LLM agents in multi-turn interactions, including four core abilities: accurate retrieval, learning during testing, long-range understanding, and selective forgetting.
Evaluating SAE interpretability without generating explanations
Gonçalo Paulo (EleutherAI), Nora Belrose (EleutherAI)
Explainability and InterpretabilityLarge Language ModelSupervised Fine-TuningAuto EncoderText
🎯 What it does: Propose two interpretability evaluation methods for sparse autoencoders (SAEs) that do not require natural language explanations: intruder detection and example embedding scoring.
Evaluating Text Creativity across Diverse Domains: a Dataset and Large Language Model Evaluator
Qian Cao (Renmin University of China), Ruihua Song (Beijing Normal University)
Large Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Propose a context-based dual comparison framework for evaluating text creativity, and develop a large cross-domain dataset CreataSet and the corresponding LLM evaluator CrEval
Event-T2M: Event-level Conditioning for Complex Text-to-Motion Synthesis
Seong-Eun Hong (Korea University), HyeongYeop Kang (Kyung Hee University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelTextMultimodalitySequentialRetrieval-Augmented Generation
🎯 What it does: Propose an event-level conditioned diffusion framework called Event-T2M for generating motion sequences with multiple actions and complex order from natural language.
EventFlash: Towards Efficient MLLMs for Event-Based Vision
Shaoyu Liu (Xidian University), Xiangyang Ji (Tsinghua University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelMultimodalityTime Series
🎯 What it does: Proposed EventFlash, an efficient event-based visual multimodal large language model that achieves fast inference and long sequence processing through spatiotemporal token sparsification.
EVEREST: A Transformer for Probabilistic Rare-Event Anomaly Detection with Evidential and Tail-Aware Uncertainty
Antanas Žilinskas (Imperial College London), Jakub Marecek
Anomaly DetectionTransformerTime SeriesPhysics Related
🎯 What it does: A compact Transformer named EVEREST was constructed for rare event probability prediction in multivariate time series, with multiple auxiliary heads incorporated during training to achieve calibration and tail risk assessment;
Every Language Model Has a Forgery-Resistant Signature
Matthew Finlayson (University of Southern California), Swabha Swayamdipta (University of Southern California)
TransformerLarge Language ModelText
🎯 What it does: Investigate geometric constraints of language model outputs, using high-dimensional elliptical constraints as verifiable model signatures;
Everything in Its Place: Benchmarking Spatial Intelligence of Text-to-Image Models
Zengbin Wang (AMAP, Alibaba Group), Xiangxiang Chu (AMAP, Alibaba Group)
GenerationData SynthesisLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the SpatialGenEval benchmark, which evaluates the spatial intelligence of T2I models with information-dense long prompts and multiple-choice questions across 10 spatial subdomains, and constructed the SpatialT2I dataset for supervised fine-tuning.
Evidence for Limited Metacognition in LLMs
Christopher Ackerman
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Propose two experimental paradigms to evaluate LLM's mastery of internal confidence and self-simulation without relying on self-reporting, thereby quantifying its metacognitive ability.
EVLP: Learning Unified Embodied Vision-Language Planner with Reinforced Supervised Fine-Tuning
Xinyan Cai (Institute of Automation, Chinese Academy of Sciences), Jianye HAO
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision-Language-Action ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposed a unified audio-visual language planning framework EVLP, which simultaneously generates language instructions and visual sub-goals in multi-step object manipulation tasks, helping robots better decompose and execute complex tasks.
Evoking User Memory: Personalizing LLM via Recollection-Familiarity Adaptive Retrieval
Yingyi Zhang (Dalian University of Technology), Xiangyu Zhao (City University of Hong Kong)
RetrievalComputational EfficiencyLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the RF-Mem framework, leveraging the dual processes of familiarity and recall from human memory to construct an adaptive retrieval mechanism;
EvolProver: Advancing Automated theorem proving by Evolving Formalized Problems via Symmetry and Difficulty
Yuchen Tian (Hong Kong Baptist University), Lun Du (Ant Group)
Data-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a three-stage data augmentation pipeline based on symmetry and difficulty (EvolDomain, EvolDifficulty, EvolAST) to generate semantically consistent but structurally diverse formal theorems, and trains a 7B-parameter non-reasoning theorem prover called EvolProver.
Evolution and compression in LLMs: on the emergence of human-aligned categorization
Nathaniel Imel (New York University), Noga Zaslavsky (New York University)
ClassificationCompressionLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper investigates the semantic systems of large language models (LLMs) in color classification tasks. It first evaluates their alignment with human English systems and information bottleneck (IB) efficiency in an English color naming experiment. Subsequently, it simulates cultural evolution through the proposed iterative contextual language learning (IICLL) method, observing how LLMs evolve near-human color category systems without explicit supervision. Preliminary validation is conducted in different semantic domains such as Shepard circles.
Evolution of Concepts in Language Model Pre-Training
Xuyang Ge (OpenMOSS Team, Shanghai Innovation Institute; Fudan University), Xipeng Qiu (OpenMOSS Team, Shanghai Innovation Institute; Fudan University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Track and analyze the evolution of features during the pre-training process of large-scale language models, using cross-snapshot sparse autoencoder (crosscoder) to capture the emergence, rotation, and disappearance of features across different training stages, and further associate micro-level evolution with downstream task performance through feature attribution.
Evolutionary Caching to Accelerate Your Off-the-Shelf Diffusion Model
Anirud Aggarwal (University of Maryland), Matthew Gwilliam (University of Maryland)
GenerationOptimizationComputational EfficiencyTransformerDiffusion modelImageTextMultimodality
🎯 What it does: Propose a cache scheduling search framework based on genetic algorithms for diffusion models, automatically discovering the Pareto frontier of trade-offs between speed and quality without modifying model weights.
Evolving Graph Structured Programs for Circuit Generation with Large Language Models
Yinqi Bai (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringGraphBenchmark
🎯 What it does: This paper proposes CircuitEvo, an evolutionary circuit program generation framework based on large language models, which iteratively evolves circuit programs and achieves functionally accurate and more compact logic synthesis through structure-aware functional optimization.
EvoTest: Evolutionary Test-Time Learning for Self-Improving Agentic Systems
Yufei He (National University of Singapore), Bryan Hooi (National University of Singapore)
Meta LearningLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the J-TTL benchmark to evaluate AI agents' ability to progressively improve across multiple attempts in the same game, and designed the EvoTest framework to achieve gradient-free self-evolution during testing
Exchangeability of GNN Representations with Applications to Graph Retrieval
Kartik Nair (Carnegie Mellon University), Abir De (IIT Bombay)
RetrievalRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Investigate and exploit the exchangeability of node embeddings along the dimension axis after training with graph neural networks (GNN), and build a unified graph retrieval locality-sensitive hashing (LSH) framework called GRAPHHASH.
Executable Counterfactuals: Improving LLMs' Causal Reasoning Through Code
Aniket Vashishtha (University of Illinois Urbana Champaign), Hao Peng (University of Illinois Urbana Champaign)
Explainability and InterpretabilityAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Propose an operationalizable counterfactual reasoning framework based on executable code (and corresponding mathematical problems), explicitly decomposing reasoning into three steps: abduction, intervention, and prediction;
ExGRPO: Learning to Reason from Experience
Runzhe Zhan (University of Macau), Yu Cheng (Chinese University of Hong Kong)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: In large-scale language model verifiable reward reinforcement learning (RLVR), a hybrid strategy optimization framework based on experience management, ExGRPO, is introduced to improve inference performance and training stability.
Exo-Plore: Exploring Exoskeleton Control Space through Human-aligned Simulation
Geonho Leem (Seoul National University), Jungdam Won (Seoul National University)
OptimizationRobotic IntelligenceReinforcement LearningBiomedical Data
🎯 What it does: Developed the Exo-plore framework based on neuromechanical simulation and deep reinforcement learning to optimize hip exoskeleton control parameters without requiring human experiments.
ExoPredicator: Learning Abstract Models of Dynamic Worlds for Robot Planning
Yichao Liang (University of Cambridge), Kevin Ellis (Cornell University)
Robotic IntelligenceLarge Language ModelPrompt Engineering
🎯 What it does: Propose an end-to-end framework called ExoPredicator for learning abstract world models in robot planning, including symbolic state representations and endogenous/exogenous causal processes, and rapidly adapt to new tasks through online learning.
EXP-Bench: Can AI Conduct AI Research Experiments?
Patrick Tser Jern Kon, Ang Chen (University Of Michigan)
Large Language ModelAgentic AITextBenchmark
🎯 What it does: Propose the EXP-Bench benchmark to evaluate AI agents' ability to complete the full scientific experiment process, from experimental design, implementation, execution, to conclusion.
Expanding Reasoning Potential in Foundation Model by Learning Diverse Chains of Thought Patterns
Xuemiao Zhang (Peking University), Xunliang Cai (Meituan)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: By defining the model's reasoning potential (i.e., the reciprocal of the number of attempts required to solve a problem), a core set containing high-value chain reasoning patterns is constructed, and a dual-granularity algorithm (based on reasoning pattern chains and token entropy) is designed to efficiently select training samples that conform to the core set distribution from massive CoT data, thereby significantly improving the performance of large language models on complex mathematical reasoning tasks.
Expanding the Capability Frontier of LLM Agents with ZPD-Guided Data Synthesis
Xuanzhong Chen (Tsinghua University), Ting Chen (Tsinghua University)
Data SynthesisReinforcement LearningAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: Propose AgentFrontier Engine, a data synthesis framework based on the educational psychology concept of 'Zone of Proximal Development' (ZPD), which automatically generates interdisciplinary, knowledge-integrated frontier tasks and constructs an adaptive evaluation benchmark called ZPD Exam;
Experience-based Knowledge Correction for Robust Planning in Minecraft
Seungjoon Lee (POSTECH), Jungseul Ok (POSTECH)
Knowledge DistillationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the experience-based knowledge correction framework XENON to enhance the planning capabilities of LLMs in complex environments such as Minecraft
Expert Divergence Learning for MoE-based Language Models
Jiaang Li (Alibaba Group), Bo Zheng (Alibaba Group)
Large Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes the Expert Divergence Learning (EDL) pre-training strategy, which introduces a domain label-driven auxiliary loss in MoE models to encourage experts to generate separated routing distributions across different data domains, thereby achieving significant differentiation in expert functions;
Expert Heads: Robust Evidence Identification for Large Language Models
Qi Wu (Soochow University), Zhixu Li (Renmin University of China)
RetrievalTransformerLarge Language ModelText
🎯 What it does: This paper analyzes the attention heads of LLMs in multi-document reasoning tasks, identifying and defining a category of attention heads called 'Expert Heads,' which consistently focus on key evidence and exhibit high stability across different document arrangements.
Expert Merging in Sparse Mixture of Experts with Nash Bargaining
Dung Viet Nguyen, Tan Minh Nguyen
OptimizationComputational EfficiencyMixture of ExpertsImageText
🎯 What it does: This paper proposes a Nash equilibrium-based expert merging method called NAMEx, combining complex momentum to achieve efficient convergence.
Expert Merging: Model Merging with Unsupervised Expert Alignment and Importance-Guided Layer Chunking
Dengming Zhang (Huawei Technologies), Xinghao Chen (Huawei Technologies)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsTextMultimodalityBenchmark
🎯 What it does: Proposed Expert Merging and its improved version Expert Merging++, achieving unsupervised fusion of multi-domain expert models by learning hierarchical (or block-wise) coefficients and aligning hidden states and logits on unlabeled calibration data.
Expertise Can Be Helpful for Reinforcement Learning-based Macro Placement
Chengrui Gao (Nanjing University), Zhi-Hua Zhou (Nanjing University)
OptimizationReinforcement Learning from Human FeedbackConvolutional Neural NetworkReinforcement LearningBenchmark
🎯 What it does: Proposed a macro placement framework EXPlace based on reinforcement learning, which utilizes expert knowledge to guide learning and achieves post-training temporal fine-tuning through preference optimization.
ExpertLongBench: Benchmarking Language Models on Expert-Level Long-Form Generation Tasks with Structured Checklists
Jie Ruan (University of Michigan), Lu Wang (University of Michigan)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Created the expert-level long-text generation benchmark ExpertLongBench, which includes 11 interdisciplinary tasks and proposed the CLEAR structured checklist-based evaluation framework.
ExpGuard: LLM Content Moderation in Specialized Domains
Minseok Choi (KAIST), Jungmin Son (KakaoBank Corp)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextFinance RelatedChain-of-Thought
🎯 What it does: Proposed a content safety model for LLMs in professional fields such as finance, healthcare, and law called EXPGUARD, and constructed a large-scale professional safety dataset named EXPGUARDMIX.
Explain in Your Own Words: Improving Reasoning via Token-Selective Dual Knowledge Distillation
Minsang Kim (Korea University), Seung Jun Baek (Korea University)
Knowledge DistillationTextChain-of-Thought
🎯 What it does: This paper proposes a bidirectional knowledge distillation framework called TSD-KD based on token selection, aiming to enhance the Chain-of-Thought generation capability of small models in complex reasoning tasks.
Explainable $ K $-means Neural Networks for Multi-view Clustering
Yalan Qin (Shanghai University), Guorui Feng (Shanghai University)
OptimizationExplainability and InterpretabilityImageTextMultimodality
🎯 What it does: Proposed an interpretable K-means neural network (EKNN), which decomposes multi-view clustering into three layers of optimization: linear clustering, non-linear clustering, and view fusion, balancing effectiveness, efficiency, completeness, and consistency.
Explainable LLM Unlearning through Reasoning
Junfeng Liao (University of Technology Sydney), Zhen Fang (RIKEN Center for Advanced Intelligence Project)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose a reasoning-based unlearning target and a method named TRU that combines a supervised reasoning loss with a gradient ascent (GA) loss, aiming to specifically remove undesirable knowledge from large language models while preserving other model capabilities.
Explainable Mixture Models through Differentiable Rule Learning
Matthias Wilms (CISPA Helmholtz Center for Information Security), Jilles Vreeken (CISPA Helmholtz Center for Information Security)
Explainability and InterpretabilityFlow-based ModelTabular
🎯 What it does: This paper proposes an Explainable Mixture Model (XMM), which pairs each mixture component with interpretable rules based on descriptive features (e.g., interval rules), to simultaneously achieve precise modeling of the target distribution and intuitive explanations for subpopulations.
Explainable Token-level Noise Filtering for LLM Fine-tuning Datasets
Yuchen Yang (Zhejiang University), Kui Ren (Zhejiang University)
Explainability and InterpretabilityData-Centric LearningLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Studied token-level noise in LLM fine-tuning datasets and proposed the XTF framework for interpretability-based token-level noise filtering.
Explaining Grokking and Information Bottleneck through Neural Collapse Emergence
Keitaro Sakamoto (University of Tokyo), Issei Sato (University of Tokyo)
OptimizationExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: By analyzing the 'neural collapse' phenomenon in the internal representations of neural networks, this paper provides a unified explanation for two late-stage training phenomena in deep learning: grokking (sudden improvement in generalization performance after training) and the information bottleneck (IB) compression phase.
Exploiting Low-Dimensional Manifold of Features for Few-Shot Whole Slide Image Classification
Conghao Xiong (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
ClassificationRepresentation LearningMeta LearningImageBiomedical Data
🎯 What it does: To address few-shot pathological whole-slide image classification, the authors propose a geometry-aware Manifold Residual (MR) block, which fixes random projections to preserve the low-dimensional manifold structure of pre-trained features and learns task-specific adaptations through low-rank residual learning, significantly enhancing model generalization.
Exploration vs Exploitation: Rethinking RLVR through Clipping, Entropy, and Spurious Reward
Peter Chen (Columbia), Tianyi Lin (Columbia)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper investigates the impact of random rewards and clipping on the exploration-exploitation balance, policy entropy, and model performance in large language models (LLMs) through theoretical analysis and large-scale experiments, revealing the implicit constraints of clipping on policy entropy and the potential mechanisms of random rewards under different model and data difficulty conditions.
Exploratory Causal Inference in SAEnce
Tommaso Mencattini (Institute of Science and Technology), Francesco Locatello (Institute of Science and Technology)
Explainability and InterpretabilityRepresentation LearningTransformerAuto EncoderImageVideo
🎯 What it does: Propose an unsupervised exploratory causal inference framework based on pre-trained models and sparse autoencoders, and design a recursive testing method called Neural Effect Search (NES) to identify treatment effects in high-dimensional data;
Exploratory Diffusion Model for Unsupervised Reinforcement Learning
Chengyang Ying (Tsinghua University), Jun Zhu (Tsinghua University)
Reinforcement LearningDiffusion modelScore-based Model
🎯 What it does: Propose a framework called Exploratory Diffusion Model (ExDM), which utilizes diffusion models to estimate the state distribution density in reward-free environments, generating score-based intrinsic rewards that guide agents to explore efficiently and pretrain diverse policies.
Exploratory Memory-Augmented LLM Agent via Hybrid On- and Off-Policy Optimization
Zeyuan Liu (Microsoft Research), Yuqing Yang (Microsoft Research)
OptimizationKnowledge DistillationTransformerLarge Language ModelReinforcement LearningAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Propose EMPO 2, a RL framework integrating memory retrieval with a hybrid update mechanism combining On-Policy and Off-Policy approaches, aimed at enhancing exploration and adaptation capabilities of large language model agents.
Explore-on-Graph: Incentivizing Autonomous Exploration of Large Language Models on Knowledge Graphs with Path-refined Reward Modeling
Shiqi Yan (Zhongguancun Laboratory), Yunqi Zhang (Zhongguancun Laboratory)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackGraph Neural NetworkLarge Language ModelSupervised Fine-TuningReinforcement LearningGraphBenchmarkChain-of-Thought
🎯 What it does: Propose a framework called EoG that encourages large language models to autonomously explore and discover new reasoning paths on knowledge graphs through reinforcement learning
Exploring Cross-Modal Flows for Few-Shot Learning
Ziqi Jiang (Hong Kong University of Science and Technology), Long Chen (Hong Kong University of Science and Technology)
ClassificationMeta LearningFlow-based ModelMultimodality
🎯 What it does: This paper proposes a multi-step feature alignment framework based on flow matching (Flow Matching Alignment, FMA) for cross-modal alignment and classification in few-shot vision-language models.
Exploring Diverse Generation Paths via Inference-time Stiefel Activation Steering
Dongxuan Zhu (Chinese University of Hong Kong), Viet Anh Nguyen (Chinese University of Hong Kong)
GenerationData SynthesisOptimizationTransformerTextBenchmark
🎯 What it does: This paper proposes a technique called STARS, which schedules orthogonal activation vectors across multiple parallel generation paths during inference, achieving diverse generation by maximizing the volume of the activation space on the Stiefel manifold;
Exploring Interpretability for Visual Prompt Tuning with Cross-layer Concepts
Yubin Wang (Microsoft Research Asia), Cairong Zhao (Tongji University)
ClassificationExplainability and InterpretabilityTransformerPrompt EngineeringImage
🎯 What it does: Introduce interpretable cross-layer concept prototypes in visual prompt tuning to generate interpretable visual prompts corresponding to human-understandable concepts.
Exploring Knowledge Purification in Multi-Teacher Knowledge Distillation for LLMs
Ruihan Jin (Tsinghua University), Jianhua Tao (Tsinghua University)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelReinforcement LearningTextBiomedical Data
🎯 What it does: Study knowledge purification methods in multi-teacher knowledge distillation, proposing five purification strategies including aggregation, three LLM routing methods (Plackett-Luce, PLM classifier, similarity routing), and reinforcement learning-based teacher selection, and verify their effectiveness on multiple tasks.
Exploring Mode Connectivity in Krylov Subspace for Domain Generalization
Aodi Li (University of Science and Technology of China), Shafei Wang (Peng Cheng Laboratory)
Domain AdaptationOptimizationConvolutional Neural NetworkTransformerPrompt EngineeringImageBenchmark
🎯 What it does: Propose the Billiard Optimization Algorithm (BOA), which leverages mode connectivity to search for models with better domain generalization performance in low-loss channels.
Exploring Real-Time Super-Resolution: Benchmarking and Fine-Tuning for Streaming Content
Evgeney Bogatyrev (Lomonosov Moscow State University), Dmitriy S. Vatolin (Lomonosov Moscow State University)
Super ResolutionComputational EfficiencyConvolutional Neural NetworkSupervised Fine-TuningVideoBenchmark
🎯 What it does: Proposed a StreamSR dataset containing 5,200 YouTube videos specifically for streaming compressed video, conducted benchmark testing of 11 real-time super-resolution models on this dataset, and subsequently designed and implemented the EfRLFN model, significantly improving the quality and efficiency of real-time super-resolution.
Exploring Specular Reflection Inconsistency for Generalizable Face Forgery Detection
Hongyan Fei (Peking University), Jie Zhou (Tencent Inc)
Anomaly DetectionConvolutional Neural NetworkTransformerImageVideo
🎯 What it does: Propose a deepfake detection framework SRI-Net based on specular reflection inconsistency, utilizing illumination separation and Retinex texture extraction to achieve facial forgery identification.
Exploring State-Space Models for Data-Specific Neural Representations
Jinsung Lee (Pohang University of Science and Technology), Suha Kwak (Pohang University of Science and Technology)
RestorationCompressionRepresentation LearningImageVideoMesh
🎯 What it does: This paper proposes integrating state-space models (SSM) into data-specific neural representations and designs a structured state-space kernel (S3K) to achieve multi-dimensional compression and downsampling;
Exploring Synthesizable Chemical Space with Iterative Pathway Refinements
Seul Lee (Korea Advanced Institute Of Science And Technology), Arash Vahdat (NVIDIA)
Drug DiscoveryTransformerFlow-based ModelBiomedical Data
🎯 What it does: Proposes the ReaSyn framework, achieving a synthetically feasible projection of a synthetically accessible chemical space through iterative bottom-up and top-down path generation combined with global editing.
Exploring the Design Space of Transition Matching
Uriel Singer (FAIR at Meta), Yaron Lipman (FAIR at Meta)
GenerationTransformerFlow-based ModelImage
🎯 What it does: This paper systematically evaluates various design choices for the backend head in continuous-time Transition Matching (TM), conducting comprehensive ablation experiments on 56 1.7B models to investigate the impact of head architecture, size, sequence expansion, batch size, and time weights on generation quality and efficiency;
Exploring the Potential of Encoder-free Architectures in 3D LMMs
Yiwen Tang (Shanghai AI Laboratory), Xuelong Li (Tele AI)
ClassificationRecognitionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPoint Cloud
🎯 What it does: This paper introduces an encoder-free architecture into 3D multimodal large models, proposes strategies such as LLM embedded semantic encoding and hierarchical geometric aggregation, and implements the ENEL model, completing the full training process from pre-training to instruction tuning.
ExPO-HM: Learning to Explain-then-Detect for Hateful Meme Detection
Jingbiao Mei (University of Cambridge), Bill Byrne (University of Cambridge)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Proposed an explain-then-detect method for hateful meme detection called ExPO-HM.
EXPO: Stable Reinforcement Learning with Expressive Policies
Perry Dong (Stanford University), Chelsea Finn (Stanford University)
Reinforcement LearningDiffusion modelFlow-based ModelTabularSequential
🎯 What it does: This paper proposes EXPO, an algorithm that performs online reinforcement learning fine-tuning on expression strategies (e.g., diffusion, flow matching).
Exponential-Wrapped Mechanisms: Differential Privacy on Hadamard Manifolds Made Practical
Yangdi Jiang (Nanyang Technological University Singapore), Bei Jiang (University of Alberta)
Safty and PrivacyImageBiomedical Data
🎯 What it does: Propose Exponential-Wrapped Laplace and Gaussian mechanisms to achieve ε-DP, (ε,δ)-DP, GDP, and RDP on Hadamard manifolds without requiring MCMC;
Exposing and Defending the Achilles' Heel of Video Mixture-of-Experts
Songping Wang (Nanjing University), Caifeng Shan (Nanjing University)
ClassificationAdversarial AttackMixture of ExpertsVideo
🎯 What it does: Propose temporal Lipschitz-guided attacks (TLGA, J-TLGA) targeting the router and expert components in video Mixture-of-Experts (MoE), revealing their independent and collaborative vulnerabilities, and design hierarchical joint adversarial training (J-TLAT) based on these weaknesses to enhance robustness.
Exposing Mixture and Annotating Confusion for Active Universal Test-Time Adaptation
Jiayao Tan (Tianjin University), Rui Yao (China Unviersity of Mining and Technology)
Domain AdaptationContrastive LearningImage
🎯 What it does: Proposed the Active Universal Test-Time Adaptation (AUTTA) framework and implemented the EMAC method for real-time model adaptation using limited human annotations in dual domain/category drift scenarios.
Exposing Weaknesses of Large Reasoning Models through Graph Algorithm Problems
Qifan Zhang (Hong Kong University of Science and Technology), Jia Li (Hong Kong University of Science and Technology)
Large Language ModelGraphBenchmark
🎯 What it does: Proposed the GRALGOBENCH benchmark, which evaluates the reasoning capabilities of large reasoning models (LRMs) using graph algorithm problems, with a particular focus on long-context reasoning and over-reasoning phenomena.
Expressive Power of Implicit Models: Rich Equilibria and Test-Time Scaling
Jialin Liu (University of Central Florida), Wotao Yin (Alibaba)
OptimizationImageTextTabularPhysics Related
🎯 What it does: Investigate the expressive power of implicit models, proving that they can achieve any local Lipschitz mapping through iteration.
Expressive yet Efficient Feature Expansion with Adaptive Cross-Hadamard Products
Xuyang Zhang (Beijing Institute of Technology), Qingshan Guo (Beijing Institute of Technology)
ClassificationObject DetectionComputational EfficiencyNeural Architecture SearchConvolutional Neural NetworkTransformerImage
🎯 What it does: Proposed an adaptive cross-Hadamard (ACH) module, achieving parameter-independent feature reuse and efficient nonlinear fusion, and constructed Hadaptive-Net via gradient NAS, achieving higher accuracy and lower computational cost on tasks such as image classification, object detection, and Transformers;