ICLR 2026 Papers — Page 49
International Conference on Learning Representations · 5356 papers
Toward Conservative Planning from Human-AI Preferences in Reinforcement Learning
Huazhong Wang (University of California, Irvine), Wenzhuo Zhou (University of California, Irvine)
Reinforcement Learning from Human FeedbackReinforcement LearningBenchmark
🎯 What it does: This paper proposes a model-driven conservative planning algorithm called MCP for offline reinforcement learning based on human-AI preference signals, learning robust policies under partial data coverage.
Toward Effective Tool-Integrated Reasoning via Self-Evolved Preference Learning
Yifei Chen (Renmin University of China), Zhicheng Dou (Renmin University of China)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes the Tool-Light framework, which significantly improves the tool-integrated reasoning efficiency and accuracy of large language models (LLMs) through entropy-guided sampling and two-stage self-evolution DPO training.
Toward Efficient Exploration by Large Language Model Agents
Dilip Arumugam (Princeton University), Thomas L. Griffiths (Princeton University)
OptimizationTransformerLarge Language ModelReinforcement LearningAgentic AITextChain-of-Thought
🎯 What it does: This paper proposes an approach to explicitly implement the posterior sampling reinforcement learning (PSRL) algorithm using large language models (LLMs), aiming to address the exploration efficiency issues of LLM agents in natural language environments;
Toward Enhancing Representation Learning in Federated Multi-Task Settings
Mehdi Setayesh (Huawei Noah's Ark Lab), Hongliang Li (Huawei Noah's Ark Lab)
Federated LearningRepresentation LearningContrastive LearningMultimodality
🎯 What it does: This paper proposes a new contrastive learning loss (Muscle loss) and designs the FedMuscle algorithm based on it to address the problem of model and task heterogeneity in federated multi-task learning.
Toward Faithful Retrieval-Augmented Generation with Sparse Autoencoders
Guangzhi Xiong (University of Virginia), Aidong Zhang (University of Virginia)
RetrievalAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerAuto EncoderTextRetrieval-Augmented Generation
🎯 What it does: Developed a lightweight retrieval-augmented generation (RAG) hallucination detector called RAGLens based on sparse autoencoders (SAE), utilizing internal activations of large language models (LLMs) to detect the authenticity of generated text and perform interpretable analysis.
Toward Practical Equilibrium Propagation: Brain-inspired Recurrent Neural Network with Feedback Regulation and Residual Connections
Zhuo Liu (University of Science and Technology of China), Tao Chen (University of Science and Technology of China)
ClassificationRecognitionComputational EfficiencyRecurrent Neural NetworkImage
🎯 What it does: Proposed a brain-inspired feedback regulation residual recurrent neural network (FRE-RNN), achieving efficient and stable learning under the equal potential propagation (EP) framework.
Toward Principled Flexible Scaling for Self-Gated Neural Activation
Sudong Cai (Hong Kong Polytechnic University), Bing WANG
Explainability and InterpretabilityConvolutional Neural NetworkTransformerImageText
🎯 What it does: Propose the FleS mechanism to address the non-local tension problem in self-gated activation, providing interpretable flexible scaling activation.
Toward Safer Diffusion Language Models: Discovery and Mitigation of Priming Vulnerability
Shojiro Yamabe (Institute of Science Tokyo), Jun Sakuma (Institute of Science Tokyo)
Safty and PrivacyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningDiffusion modelText
🎯 What it does: Quantitatively and theoretically analyze the 'priming vulnerability' of discrete diffusion language models (MDLM), and propose a novel safety alignment method called Recovery Alignment (RA), enabling the model to recover to safe responses even after encountering harmful confirmation words in intermediate steps.
Toward Universal and Transferable Jailbreak Attacks on Vision-Language Models
Kaiyuan Cui (University of Melbourne), Hanxun Huang (University of Melbourne)
Safty and PrivacyAdversarial AttackPrompt EngineeringMultimodalityBenchmark
🎯 What it does: Proposes a vision-based, general, and transferable attack method named UltraBreak that undermines the security of Vision-Language Models (VLMs);
Towards a Certificate of Trust: Task-Aware OOD Detection for Scientific AI
Bogdan Raonic (ETH Zurich), Samuel Lanthaler (ETH Zurich)
Anomaly DetectionDiffusion modelScore-based ModelImageTime SeriesBiomedical DataPhysics RelatedOrdinary Differential Equation
🎯 What it does: Investigated an OOD detection method for regression tasks, utilizing the likelihood estimation of the joint input/output distribution as a task-aware reliability score to perform zero-shot validation of predictive reliability in scientific AI;
Towards a Foundation Model for Crowdsourced Label Aggregation
Hao Liu (East China Normal University), Xiaofeng Hou (Shanghai Jiao Tong University)
ClassificationData-Centric LearningGraph Neural NetworkTextGraph
🎯 What it does: Proposes a foundational model for crowdsourcing label aggregation called CrowdFM, achieving label inference through pre-trained bilateral graph neural networks without requiring dataset training.
Towards a Sharp Analysis of Offline Policy Learning for $f$-Divergence-Regularized Contextual Bandits
Qingyue Zhao (University of California Los Angeles), Quanquan Gu (University of California Los Angeles)
Reinforcement LearningImageTabularSequential
🎯 What it does: Studies offline policy learning under f-divergence regularization (particularly inverse KL and strongly convex f-divergence) in contextual and adversarial bandits, providing matching theoretical results for optimal sample complexity and weakest data coverage conditions (single-policy concentration), and proposes corresponding algorithms.
Towards a Theoretical Understanding of In-context Learning: Stability and Non-I.I.D Generalisation
Yingjie Wang, Dacheng Tao
Explainability and InterpretabilityMeta LearningTransformerLarge Language ModelTabular
🎯 What it does: Theoretical analysis of the generalization performance of large-scale Transformers in ICL scenarios, constructing a generalization error upper bound based on algorithmic stability and distribution mismatch quantification, and verifying the theoretical predictions through experiments.
Towards a Transferable Acceleration Method for Density Functional Theory
Zhe Liu (Bytedance Seed), Wen Yan (Bytedance Seed)
OptimizationComputational EfficiencyGraph Neural NetworkGraphBenchmarkPhysics Related
🎯 What it does: Predict electron density coefficients in auxiliary basis expansions using E(3)-equivariant neural networks, directly construct the Kohn-Sham Hamiltonian from these predictions, and generate high-quality initial guesses, significantly accelerating SCF convergence in DFT.
Towards All-Atom Foundation Models for Biomolecular Binding Affinity Prediction
Liang Shi (Peking University), Jian Tang (HEC Montréal)
Drug DiscoveryTransformerDiffusion modelBiomedical Data
🎯 What it does: This paper proposes an Atom-level Diffusion Transformer (ADiT) based on AlphaFold3, which can serve as a general-purpose foundational model to predict the binding affinity of multiple types of biomolecules.
Towards Anomaly-Aware Pre-Training and Fine-Tuning for Graph Anomaly Detection
Yunhui Liu (Nanjing University), Tieke He (Hong Kong University of Science and Technology (Guangzhou))
Anomaly DetectionGraph Neural NetworkGraphTabularBenchmarkFinance Related
🎯 What it does: This paper proposes a two-phase framework APF for graph anomaly detection, which first performs anomaly-aware pre-training through unsupervised Rayleigh quotient subgraph sampling and dual spectral filtering, and then significantly improves the anomaly detection effect by adopting node and dimension adaptive fusion and anomaly-aware regularization during the fine-tuning phase.
Towards Better Branching Policies: Leveraging the Sequential Nature of Branch-and-Bound Tree
Ce Zhang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences), Guoliang Fan (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences)
OptimizationContrastive LearningSequential
🎯 What it does: Propose a sequential branching strategy called Mamba-Branching based on the Mamba network, which utilizes contrastive learning for pre-training embeddings and models the branching paths of B&B trees in an autoregressive manner to achieve more efficient branching decisions.
Towards Better Optimization For Listwise Preference in Diffusion Models
Jiamu Bai (Penn State University), Yu Wang (TikTok Inc)
GenerationOptimizationDiffusion modelImageText
🎯 What it does: Propose the Diffusion-LPO framework, which aligns diffusion models with human preferences using Plackett-Luce list preference optimization.
Towards Bridging the Gap between Large-Scale Pretraining and Efficient Finetuning for Humanoid Control
Weidong Huang (State Key Laboratory of General Artificial Intelligence, BIGAI), Jingwen Zhang (State Key Laboratory of General Artificial Intelligence, BIGAI)
Robotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: Proposes the LIFT framework, combining large-scale pre-trained SAC with physics-informed world models to achieve zero-shot deployment and efficient fine-tuning from simulation to real-world humanoid robots.
Towards Cognitively-Faithful Decision-Making Models to Improve AI Alignment
Cyrus Cousins (Duke University), Walter Sinnott-Armstrong (Duke University)
Explainability and InterpretabilityTabular
🎯 What it does: Propose a two-stage rule model based on cognitive processes to learn and simulate individual decision-making in pairwise comparisons, and validate its effectiveness in kidney transplant allocation tasks.
Towards Dynamic Interleaving Optimizers
Yile Chen (South China University Of Technology), Jin Huang (South China University Of Technology)
OptimizationImageTextTabular
🎯 What it does: Propose a dynamic optimizer interleaving training method called DOIT, which can switch between different optimizers in real-time based on the current training status to improve model convergence speed and accuracy.
Towards Efficient Constraint Handling in Neural Solvers for Routing Problems
Jieyi Bi (Nanyang Technological University), Cathy Wu (MIT)
OptimizationComputational EfficiencyTransformerReinforcement Learning
🎯 What it does: Propose the Construct-and-Refine (CaR) framework, which jointly trains a neural constructor and refiner to achieve explicit feasibility correction, applicable to complex-constrained vehicle routing problems (VRP).
Towards Efficient Optimizer Design for LLM via Structured Fisher Approximation with a Low-Rank Extension
Wenbo Gong (Microsoft Research), Edward Meeds
OptimizationComputational EfficiencyLarge Language ModelText
🎯 What it does: Propose an optimizer design framework based on structured Fisher information matrix (FIM) approximation, and under this framework introduce two efficient LLM optimizers: RACS (Row-and-Column-Scaled SGD) and Alice (OSOAP with Low-Rank Extension).
Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models
Jiaqi Liu (Jilin University), Bo Yang (Jilin University)
TransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Developed the Perceptually-Grounded Geospatial Chain-of-Thought (Geo-CoT) framework, and implemented the RSThinker Vision-Language model, which can generate verifiable step-by-step reasoning chains on remote sensing images, providing both final answers and visual evidence for each step.
Towards Generalizable PDE Dynamics Forecasting via Physics-Guided Invariant Learning
Siyang Li (Hong Kong University Of Science And Technology), Hui Xiong (Alibaba Cloud)
Mixture of ExpertsTime SeriesPhysics Related
🎯 What it does: Propose the iMOOE framework to achieve generalization of PDE dynamics prediction under zero-shot scenarios.
Towards Greater Leverage: Scaling Laws for Efficient Mixture-of-Experts Language Models
Changxin Tian (Ant Group), JUN ZHOU
Computational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Studied the computational efficiency of Mixture-of-Experts (MoE) language models and proposed the Efficiency Leverage (EL) metric to quantify the computational advantages of MoE over equivalent dense models.
Towards High Data Efficiency in Reinforcement Learning with Verifiable Reward
Xinyu Tang (Renmin University of China), JUN ZHOU
Reinforcement LearningText
🎯 What it does: This work proposes DEPO, a data efficiency enhancement method for RLVR that combines offline multi-objective sample selection with online explorability filtering.
Towards Improved Sentence Representations using Token Graphs
Krishna Sri Ipsit Mantri (University of Bonn), Moshe Eliasof (University of Cambridge)
Computational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Propose a lightweight pooling module called GLOT based on token graphs, utilizing a frozen large language model to generate sentence-level representations
Towards Interpretable Visual Decoding with Attention to Brain Representations
Pinyuan Feng (Columbia University), Nikolaus Kriegeskorte (Columbia University)
GenerationExplainability and InterpretabilityRepresentation LearningTransformerDiffusion modelImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose NeuroAdapter, which directly injects fMRI signals into cross-attention Stable Diffusion for visual reconstruction through partitioned linear mapping, bypassing the intermediate feature bottleneck.
Towards Knowledge‑and‑Data‑Driven Organic Reaction Prediction: RAG‑Enhanced and Reasoning‑Powered Hybrid System with LLMs
Qingyu Wang (Institute of Automation, Chinese Academy of Sciences), Bo XU
Explainability and InterpretabilityDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Reaction-Thinker system, integrating retrieval-augmented generation (RAG) with chain-of-thought (CoT) large language models to achieve interpretable and high-accuracy predictions for organic reaction forecasting;
Towards Lossless Memory-efficient Training of Spiking Neural Networks via Gradient Checkpointing and Spike Compression
Yifan Huang (Peking University), Yonghong Tian (Peng Cheng Laboratory)
CompressionComputational EfficiencySpiking Neural NetworkImageVideoSequential
🎯 What it does: Proposed an automated memory optimization pipeline that integrates hierarchical gradient checkpointing, lossless synaptic compression, and adaptive spatiotemporal segmentation strategies, achieving up to 8× compression of memory for SNN training;
Towards Multimodal Data-Driven Scientific Discovery Powered by LLM Agents
Fan Liu (Hong Kong University of Science and Technology), Hao Liu (Hong Kong University of Science and Technology)
Drug DiscoveryLarge Language ModelAgentic AIImageTextMultimodalityTabularTime SeriesBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Built MoSciBench — the first benchmark for multi-modal data-driven scientific discovery, covering 6 disciplines, 7 data modalities, and a total of 88 end-to-end tasks.
Towards Multimodal Time Series Anomaly Detection with Semantic Alignment and Condensed Interaction
Shiyan Hu (East China Normal University), Chenjuan Guo (East China Normal University)
Anomaly DetectionTransformerLarge Language ModelContrastive LearningTextMultimodalityTime Series
🎯 What it does: Proposes the MindTS model, integrating time series and text information for multimodal anomaly detection;
Towards One-step Causal Video Generation via Adversarial Self-Distillation
Yongqi Yang (Wuhan University), Yu Wu (Wuhan University)
GenerationKnowledge DistillationDiffusion modelGenerative Adversarial NetworkVideo
🎯 What it does: Propose a one/two-step causal video generation framework by distilling multi-step diffusion models, and introduce adversarial self-distillation and first-frame enhancement strategies to achieve high-quality real-time video generation.
Towards Persistent Noise-Tolerant Active Learning of Regular Languages with Class Query
Lekai Chen (University of Colorado Boulder), Alvaro Velasquez (University of Colorado Boulder)
Large Language ModelPrompt EngineeringText
🎯 What it does: Propose an algorithm called CAPAL for learning deterministic finite automata (DFA) under persistent noise, treating large language models (LLMs) as probabilistic teachers.
Towards Personalized Deep Research: Benchmarks and Evaluations
Yuan Liang (Zhejiang University), Wangchunshu Zhou (Zhejiang University)
Large Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper designs the first benchmark for personalized deep research, PDR-Bench, and proposes a three-dimensional evaluation framework PQR (Personalization, Content Quality, Fact Reliability) to systematically evaluate the personalization capabilities of deep research agents (DRA).
Towards Physically Executable 3D Gaussian for Embodied Navigation
Bingchen Miao (Zhejiang University), Juncheng Li (Zhejiang University)
Robotic IntelligenceLarge Language ModelVision-Language-Action ModelGaussian SplattingTextMultimodalityPoint CloudMeshBenchmark
🎯 What it does: Propose SAGE-3D, transforming 3D Gaussian Splatting into an executable, semantic, and physically aligned environment that supports continuous vision-language navigation
Towards Privacy-Guaranteed Label Unlearning in Vertical Federated Learning: Few-Shot Forgetting Without Disclosure
Hanlin Gu (WeBank AI Lab), Chee Seng Chan (Universiti Malaya)
Federated LearningSafty and PrivacyImageTextBiomedical Data
🎯 What it does: To address the label forgetting problem in vertical federated learning (VFL), a few-shot label forgetting framework is proposed: first, manifold mixup is applied to the representation layer to synthesize embeddings, then gradient ascent is performed on active and passive models to achieve label forgetting, followed by a recovery phase to maintain the performance of retained samples.
Towards Quantifying Long-Range Interactions in Graph Machine Learning: a Large Graph Dataset and a Measurement
Huidong Liang (University of Oxford), Xiaowen Dong (University of Oxford)
Graph Neural NetworkTransformerGraphBenchmark
🎯 What it does: Proposes City-Networks, a large-scale urban road network dataset, aimed at evaluating the long-range dependency capabilities of graph neural networks, and introduces a long-range measurement method based on the Jacobian matrix.
Towards Quantization-Aware Training for Ultra-Low-Bit Reasoning LLMs
Yasuyuki Okoshi (Institute of Science), Masato Motomura (Institute of Science)
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a two-stage quantization-aware training (QAT) workflow tailored for post-training inference large language models, aiming to maintain inference capabilities after ultra-low-bit (<4 bits) quantization;
Towards Reliable Benchmarking: A Contamination Free, Controllable Evaluation Framework for Multi-step LLM Function Calling
Seiji Maekawa (Megagon Labs), Estevam Hruschka (Megagon Labs)
Data SynthesisLarge Language ModelTextGraphBenchmark
🎯 What it does: Propose the FuncBenchGen framework for automatically generating clean, controllable difficulty multi-step function call evaluation tasks;
Towards Reliable Detection of Empty Space: Conditional Marked Point Processes for Object Detection
Tobias Riedlinger (Technical University of Berlin), Hanno Gottschalk (Technical University of Berlin)
Object DetectionAutonomous DrivingConvolutional Neural NetworkImage
🎯 What it does: Designed a target detection model based on conditional marked Poisson point processes, which can provide confidence estimates for the probability of any image region being empty (drivable), and employs maximum likelihood loss during training.
Towards Robust Real-World Multivariate Time Series Forecasting: A Unified Framework for Dependency, Asynchrony, and Missingness
Jinkwan Jang (Seoul National University), Taesup Kim (Seoul National University)
TransformerTime SeriesBenchmark
🎯 What it does: Propose ChannelTokenFormer, which uniformly handles channel asynchronous sampling, block-level missing data, and cross-channel dependencies in multivariate time series, achieving robust real-time prediction.
Towards Safe and Optimal Online Bidding: A Modular Look-ahead Lyapunov Framework
Hengquan Guo (ShanghaiTech University), Xin Liu (ShanghaiTech University)
OptimizationTabularFinance Related
🎯 What it does: Proposed an online bidding framework called L2FOB based on Look-ahead Lyapunov, aiming to achieve safe and approximately optimal bidding decisions under budget and ROI constraints.
Towards Safe Reasoning in Large Reasoning Models via Corrective Intervention
Yichi Zhang (Tsinghua University), Jun Zhu (Tsinghua University)
Safty and PrivacyLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: To address safety issues in the reasoning process of large reasoning models (LRMs), this paper achieves safe reasoning alignment by correcting unsafe reasoning steps and conducting preference learning;
Towards Sampling Data Structures for Tensor Products in Turnstile Streams
Zhao Song (University of California Berkeley), Samson Zhou (Texas Aandm University)
Computational Efficiency
🎯 What it does: Investigate and propose an attention sampler in streaming environments, significantly reducing the computational complexity of attention matrices through sparse sampling
Towards Scalable Oversight via Partitioned Human Supervision
Ren Yin (University of Tokyo), Masashi Sugiyama (RIKEN)
ClassificationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningAgentic AITextBenchmarkFinance Related
🎯 What it does: Proposed a scalable supervised framework—Assessment and Training with Partitioned Human Supervision—which utilizes 'not this class' complementary labels provided by professional experts to estimate the accuracy of AI systems and use them as training signals.
Towards Self-Evolving Agent Benchmarks : Validatable Agent Trajectory via Test-Time Exploration
Dadi Guo (Shanghai Artificial Intelligence Laboratory), Jing Shao (Hong Kong University of Science and Technology)
Large Language ModelAgentic AITextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the TRACE framework to achieve self-evolution of benchmark tasks: through evolutionary proposals, exploration execution, and multi-level trajectory verification, incrementing task difficulty while preserving verifiable execution trajectories.
Towards Self-Robust LLMs: Intrinsic Prompt Noise Resistance via CoIPO
Xin Yang (Zhejiang University), Wenyuan Jiang (ETH Zurich)
Reinforcement Learning from Human FeedbackTransformerContrastive LearningTextBenchmark
🎯 What it does: To address the issue of performance fluctuations in large language models when prompts contain minor noise, this paper proposes the CoIPO framework, which enhances the model's robustness to prompt noise through contrastive learning and inverse DPO (Inverse DPO) during the post-training phase.
Towards Sequence Modeling Alignment between Tokenizer and Autoregressive Model
Pingyu Wu (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
GenerationTransformerDiffusion modelImage
🎯 What it does: Proposed the AliTok tokenizer, which aligns with autoregressive models by introducing forward dependencies in image token sequences through a causal decoder that constrains the bidirectional encoder; subsequently, standard decoder-only autoregressive models are used to generate images;
Towards Strategic Persuasion with Language Models
Zirui Cheng (University of Illinois Urbana-Champaign), Jiaxuan You (University of Illinois Urbana-Champaign)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This paper proposes a framework based on Bayesian persuasion theory, which reconstructs a multi-agent environment using human persuasion data to evaluate and train large language models (LLMs) in strategic persuasion, and enhances the persuasion effectiveness of small LLMs through reinforcement learning.
Towards Sustainable Investment Policies Informed by Opponent Shaping
Juan Agustin Duque (University of Montreal), Aaron Courville (University of Montreal)
Reinforcement LearningTime SeriesFinance Related
🎯 What it does: Model the interaction between investors and companies under climate risk as a multi-agent reinforcement learning game, InvestESG, and demonstrate it as a temporal social dilemma.
Towards Text-Mask Consistency in Medical Image Segmentation
Jie Gui (Anhui University), Xiuquan Du (Anhui University)
SegmentationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningMultimodalityBiomedical Data
🎯 What it does: Proposed the C2Seg two-stage medical image segmentation framework, aiming to enhance the consistency between image segmentation results and medical text descriptions.
Towards True Speech-to-Speech Models Without Text Guidance
Xingjian Zhao (Fudan University), Xipeng Qiu (Fudan University)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: Proposed a true speech-to-speech large language model capable of directly understanding and generating speech without relying on text as an intermediary;
Towards Understanding Subliminal Learning: When and How Hidden Biases Transfer
Simon Schrodi (University of Freiburg), Thomas Brox (University of Freiburg)
Explainability and InterpretabilityKnowledge DistillationRepresentation LearningTransformerPrompt EngineeringText
🎯 What it does: Studied the unconscious learning phenomenon in language models during distillation, revealing through comparative experiments and mechanism analysis that differential words are the core driving force behind this phenomenon.
Towards Understanding The Calibration Benefits of Sharpness-Aware Minimization
Chengli Tan (Northwestern Polytechnical University), Yong Xu (Xi'an Jiaotong University)
ClassificationOptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: Studied the impact of Sharpness-Aware Minimization (SAM) on the calibration of deep neural networks, and proposed an improved CSAM algorithm
Towards Understanding the Nature of Attention with Low-Rank Sparse Decomposition
Zhengfu He (Shanghai Innovation Institute), Xipeng Qiu (Shanghai Innovation Institute)
Explainability and InterpretabilityTransformerAuto EncoderText
🎯 What it does: Designed and implemented a low-rank sparse attention model called Lorsa to replace the original multi-head self-attention (MHSA), decomposing and interpreting attention units through sparsification and single-dimensional OV circuits.
Towards Understanding the Shape of Representations in Protein Language Models
Kosio Beshkov (University of Oslo), Anders Malthe-Sorenssen
Representation LearningDrug DiscoveryTransformerLarge Language ModelBiomedical Data
🎯 What it does: This paper studies the geometric and structural encoding of protein language models (PLM) in different hierarchical representation spaces through shape analysis (SRV representation and graph filtering), quantifying metrics such as effective dimensions, Fréchet radius, and graph filter moments;
Towards Understanding Valuable Preference Data for Large Language Model Alignment
Zizhuo Zhang (Hong Kong Baptist University), Masashi Sugiyama (University Of Tokyo)
Data-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Propose a truncated influence function (TIF) to evaluate preference data quality, and based on TIF, introduce a lightweight LossDiff-IRM selection method to improve large language model alignment effectiveness.
TP-Spikformer: Token Pruned Spiking Transformer
Wenjie Wei (University of Electronic Science and Technology of China), Haizhou Li (Shenzhen Loop Area Institute)
Object DetectionObject TrackingSegmentationComputational EfficiencySpiking Neural NetworkTransformerImage
🎯 What it does: This paper proposes TP-Spikformer, a token pruning method specifically designed for spiking Transformers;
TPDiff: Temporal Pyramid Video Diffusion Model
Lingmin Ran (National University of Singapore), Mike Zheng Shou (National University of Singapore)
GenerationDiffusion modelFlow-based ModelVideo
🎯 What it does: Propose the Time Pyramid Video Diffusion framework TPDiff, combining phased diffusion training and noise-data alignment, significantly improving training and inference efficiency while maintaining video quality.
TPRU: Advancing Temporal and Procedural Understanding in Large Multimodal Models
Zhenkun Gao (East China Normal University), Yuan Xie (East China Normal University)
Robotic IntelligenceSupervised Fine-TuningReinforcement LearningVision Language ModelImageVideoTextMultimodalitySequential
🎯 What it does: Proposed and constructed the TPRU dataset, leveraging three temporal tasks (temporal reordering, next-frame prediction, previous-frame review) and negative samples, combined with reinforcement learning fine-tuning, significantly improving the performance of small-scale multimodal language models in temporal and program reasoning.
TRAC: Tensor-Train based Across-layer Compression for Parameter-Efficient Fine-Tuning
Bangguo Ye, Xiaoqun Zhang (Shanghai Jiao Tong University)
CompressionComputational EfficiencyKnowledge DistillationRepresentation LearningSupervised Fine-TuningImageText
🎯 What it does: Propose a new parameter-efficient fine-tuning framework TRAC, which utilizes Tensor-Train decomposition and shares/freezes cores across layers, while incorporating a lightweight controller to achieve fine-tuning with fewer trainable parameters.
Trace Anything: Representing Any Video in 4D via Trajectory Fields
Xinhang Liu (ByteDance Seed), Bingyi Kang (ByteDance Seed)
Object TrackingRepresentation LearningConvolutional Neural NetworkTransformerImageVideo
🎯 What it does: This paper proposes a 4D video representation called Trajectory Fields, using pixels as atomic units, and designs a single-channel neural network named Trace Anything, capable of predicting continuous 3D trajectories for all pixels simultaneously without requiring depth, optical flow, or post-processing;
TRACE: Your Diffusion Model is Secretly an Instance Edge Detector
Sanghyun Jo (OGQ), Kyungsu Kim (Seoul National University)
SegmentationDiffusion modelImage
🎯 What it does: Propose the TRACE framework, which utilizes the self-attention of text-to-image diffusion models during the denoising phase to extract instance edges, achieving instance and panoptic segmentation without manual annotation.
Traceable Black-Box Watermarks For Federated Learning
Jiahao Xu (University of Nevada, Reno), Zikai Zhang (University of Nevada, Reno)
Federated LearningSafty and PrivacyImage
🎯 What it does: Propose a new federated learning framework called TraMark, which injects traceable black-box watermarks into the global model to protect model intellectual property and trace leakage sources.
Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Method
Haochen Wang (New Laboratory of Pattern Recognition, State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences), Zhaoxiang Zhang (New Laboratory of Pattern Recognition, State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences)
Explainability and InterpretabilityTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose TreeBench, a benchmark for evaluating vision-based image reasoning, and TreeVGR, a training framework, emphasizing traceable visual evidence and multi-step reasoning.
TRACED: Transition-aware Regret Approximation with Co-learnability for Environment Design
Geonwoo Cho (Gwangju Institute of Science and Technology), Sundong Kim (Gwangju Institute of Science and Technology)
Data SynthesisOptimizationReinforcement LearningBenchmark
🎯 What it does: Propose an unsupervised environment design method called TRACED, which enhances the agent's zero-shot generalization capability in RL training through adaptive task generation and a replay mechanism.
TRACEDET: HALLUCINATION DETECTION FROM THE DECODING TRACE OF DIFFUSION LARGE LANGUAGE MODELS
Shenxu Chang (University of Oxford), Jindong Gu (University of Oxford)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Propose TraceDet, which detects hallucinatory outputs by leveraging the multi-step denoising process (decoding trajectory) of diffusion-based large language models (D-LLM).
Tracing and Reversing Edits in LLMs
Paul Youssef (Marburg University), Jörg Schlötterer (Marburg University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a task that uses post-editing weights to track and reverse knowledge edits, along with the corresponding solution.
Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning
Martina G. Vilas (Goethe University Frankfurt), Vidhisha Balachandran (Microsoft Research)
Explainability and InterpretabilityComputational EfficiencyLarge Language ModelTextTime SeriesChain-of-Thought
🎯 What it does: This paper investigates using the time trajectory of hidden states (Latent-Trajectory signals) from large language models to predict the quality of reasoning paths, and employs these signals for multi-sample reasoning and early termination strategies to enhance reasoning efficiency and accuracy.
Tracking Equivalent Mechanistic Interpretations Across Neural Networks
Alan Sun (Carnegie Mellon University), Mariya Toneva (Max Planck Institute for Software Systems)
Explainability and InterpretabilityTransformerTextSequential
🎯 What it does: This paper introduces the concept of 'interpretive equivalence' and designs an algorithm called Congruity, which uses the similarity of linear representations realized by models to determine whether two models implement the same high-level algorithm without requiring explicit explanations.
Tractability via Low Dimensionality: The Parameterized Complexity of Training Quantized Neural Networks
Robert Ganian (TU Wien), Manuel Sorge (TU Wien)
Computational Efficiency
🎯 What it does: This paper conducts a systematic complexity analysis of the training problem for fully quantized ReLU neural networks (weights and biases take integer domain values), providing various lower and upper bounds results;
Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading
Zifan Song (Tongji University), Cairong Zhao (Tongji University)
OptimizationTransformerLarge Language ModelAgentic AITime SeriesFinance Related
🎯 What it does: Built a multi-agent system TiMi, enabling the transition of quantitative trading strategies from macro-level analysis to micro-level customization, and achieving minute-level execution through offline strategy optimization
Train on Validation (ToV): Fast data selection with applications to fine-tuning
Ayush Jain (Google Research), Eren Sasoglu (Encord)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a data selection method called ToV based on 'first fine-tuning on a validation set, then evaluating on a training pool,' which quickly assesses the impact of samples on the target distribution by only computing the forward loss difference;
Train Once, Answer All: Many Pretraining Experiments for the Cost of One
Sebastian Bordt (University of Tbingen), Martin Pawelczyk (University of Vienna)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This study proposes and implements a method that simultaneously performs multiple independent experiments during a single pre-training process, successfully completing 10 experiments (including knowledge acquisition, mathematical reasoning, data watermarking, etc.) in the same training session, and verifies its ability to replicate previous experimental results.
Train-before-Test Harmonizes Language Model Rankings
Guanhua Zhang (Max Planck Institute for Intelligent Systems), Moritz Hardt (Max Planck Institute for Intelligent Systems)
Large Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed and implemented the 'train-before-test' method, which uniformly fine-tunes all 61 LLMs using the training set of each benchmark first, then evaluates on the test set to obtain a ranking of model potential.
Trained on Tokens, Calibrated on Concepts: The Emergence of Semantic Calibration in LLMs
Preetum Nakkiran (Apple), Sinead Williamson (Apple)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Study the semantic calibration phenomenon in large language models (LLMs) during long text generation, proving and verifying that sample-based semantic calibration can naturally emerge.
Training Deep Normalization-Free Spiking Neural Networks with Lateral Inhibition
Peiyu Liu (Peking University), Zhaofei Yu (Peking University)
ClassificationSpiking Neural NetworkImage
🎯 What it does: Proposed a completely unnormalized deep spiking neural network (SNN) training framework that leverages the excitatory-inhibitory (E-I) separation and lateral inhibition mechanisms in the cortex to replace traditional normalization techniques.
Training Dynamics Impact Post-Training Quantization Robustness
Albert Catalan-Tatjer (ELLIS Institute Tübingen), Jonas Geiping (ELLIS Institute Tübingen)
OptimizationComputational EfficiencyHyperparameter SearchTransformerLarge Language ModelTextBenchmark
🎯 What it does: Systematically analyze and experimentally verify the impact of training dynamics (learning rate, weight decay, weight averaging, etc.) on the robustness of post-training quantization (PTQ) for large-scale LLMs, and propose to improve low-bit quantization performance by adjusting these hyperparameters.
Training Large Language Models To Reason In Parallel With Global Forking Tokens
Sheng Jia (University of Toronto), Shiva Kasiviswanathan
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This paper introduces global branch tokens to train large language models for parallel reasoning, thereby improving reasoning accuracy while maintaining diversity.
Training Large Reasoning Models Efficiently via Progressive Thought Encoding
Zeliang Zhang (University of Rochester), Jianfeng Gao (Microsoft Research)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: In the paper, the authors propose a method called Progressive Thought Encoding for efficiently training large-scale reasoning models under limited cache conditions;
Training-free Counterfactual Explanation for Temporal Graph Model Inference
Mingjian Lu (Case Western Reserve University), Yinghui Wu (Case Western Reserve University)
Explainability and InterpretabilityGraph Neural NetworkGraphTime Series
🎯 What it does: This paper proposes TemGX, a training-free, instance-oriented temporal graph explanation framework that provides interpretable substructures of TGNN outputs through temporal subgraphs and temporal counterfactual analysis, while supporting time-pattern-based queries.
Training-Free Determination of Network Width via Neural Tangent Kernel
Tatsumi Sunada (University of Tokyo), Atsuto Maki (KTH Royal Institute of Technology)
Computational EfficiencyNeural Architecture SearchImageTabular
🎯 What it does: Propose a pre-training width selection method that estimates the required minimum network width using the minimum eigenvalue of the neural tangent kernel (NTK) at initialization, defining the concept of 'cardinality width'.
Training-Free Loosely Speculative Decoding: Accepting Semantically Correct Drafts Beyond Exact Match
Jinze Li (Advanced Micro Devices, Inc.), Emad Barsoum (Advanced Micro Devices, Inc.)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes a training-agnostic loose speculative decoding (FLy) that enhances LLM inference speed by determining the semantic validity of mismatches using entropy gates and deferred windows.
Training-Free Reward-Guided Image Editing via Trajectory Optimal Control
Jinho Chang (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
GenerationDiffusion modelFlow-based ModelImageStochastic Differential Equation
🎯 What it does: This paper proposes a training-free, reward-guided image editing framework that treats the editing process as an optimal control problem for the inverse trajectory of diffusion or flow matching models, achieving trajectory optimization through adjoint state iteration.
Training-Free Text-Guided Color Editing with Multi-Modal Diffusion Transformer
Zixin Yin (Hong Kong University of Science and Technology), Heung-Yeung Shum (Hong Kong University of Science and Technology)
Image HarmonizationGenerationTransformerDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: Developed a training-free text-guided color editing method called ColorCtrl, which utilizes the attention mechanism of a multi-modal diffusion Transformer to achieve precise color adjustment while maintaining consistency in geometry, material, and lighting.
TrainRef: Curating Data with Label Distribution and Minimal Reference for Accurate Prediction and Reliable Confidence
Murong Ma (National University of Singapore), Jin Song Dong (National University of Singapore)
ClassificationOptimizationData-Centric LearningContrastive LearningImage
🎯 What it does: Propose the TrainRef data cleaning framework during training, which uses a minimal number of reference samples to perform distributed labeling of noisy labels, thereby improving the model's prediction accuracy and confidence.
TRAJECT-Bench:A Trajectory-Aware Benchmark for Evaluating Agentic Tool Use
Pengfei He (Michigan State University), Benoit Dumoulin (Hippocratic AI)
Large Language ModelAgentic AITextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Created TRAJECT-Bench, a novel evaluation benchmark for tool usage trajectories, assessing LLMs' ability to invoke, parameterize, and arrange tools in real-world tasks.
Trajectory Generation with Conservative Value Guidance for Offline Reinforcement Learning
Tieru Wang (Beijing University of Posts and Telecommunications), Guoshun Nan (Beijing University of Posts and Telecommunications)
Data SynthesisTransformerReinforcement LearningWorld ModelSequential
🎯 What it does: Proposed a Transformer-based offline reinforcement learning data augmentation framework called TGCVG, which utilizes conservative value-guided strategies and learned dynamics models to generate high-quality trajectories, and mixes them with original data for offline RL training.
Trajectory-aware Shifted State Space Models for Online Video Super-Resolution
Qiang Zhu (Pengcheng Laboratory), Ronggang Wang (Peking University)
Super ResolutionOptical FlowVideo
🎯 What it does: Proposed a trajectory-aware Shifted SSM model called TS-Mamba, which selects previous frame tokens using long-term trajectories and aggregates them through Hilbert scanning combined with shifted SSM for online video super-resolution.
TrajFlow: Nation-wide Pseudo GPS Trajectory Generation with Flow Matching Models
Peiran Li (Hitotsubashi University), Renhe Jiang (University of Tokyo)
Data SynthesisFlow-based ModelTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: Proposed a flow matching generative framework named TrajFlow for generating multi-scale, multi-transportation mode synthetic GPS trajectories nationwide.
TrajTok: What makes for a good trajectory tokenizer in behavior generation?
Zhiyuan Zhang, Junchi Yan (Shanghai Jiao Tong University)
GenerationAutonomous DrivingRepresentation LearningData-Centric LearningSequential
🎯 What it does: This paper investigates the role of trajectory tokenizers in behavior generation, proposing the TrajTok tokenizer, which combines regularized grids and data-driven filtering/expansion approaches to balance coverage, utilization, symmetry, and robustness, while incorporating spatially aware label smoothing during training;
Transducing Language Models
Vésteinn Snæbjarnarson (ETH Zürich), Tim Vieira
GenerationRepresentation LearningLarge Language ModelTextBiomedical Data
🎯 What it does: Propose a framework based on finite-state transducers (FST) that converts existing language models into different units on demand, directly calculating the probabilities of the transformed units during inference without retraining;
Transductive Visual Programming: Evolving Tool Libraries from Experience for Spatial Reasoning
Shengguang Wu (Stanford University), Serena Yeung-Levy (Stanford University)
Large Language ModelVision Language ModelBenchmark
🎯 What it does: Developed an experience-based visual programming framework called TVP, achieving tool self-evolution through a dual-library closed-loop for 3D spatial reasoning
Transfer Learning in Infinite Width Feature Learning Networks
Clarissa Lauditi (Harvard University), Cengiz Pehlevan (Harvard University)
Domain AdaptationRepresentation LearningSupervised Fine-TuningImageOrdinary Differential Equation
🎯 What it does: Under the absence of width limits, the paper analyzes transfer learning in infinitely wide multi-layer perceptrons using gradient flow, providing a quantitative analysis of how feature learning impacts generalization in pre-training and downstream tasks.
Transferable and Stealthy Adversarial Attacks on Large Vision-Language Models
Zhewen Yao (Peking University), Shiliang Zhang (Peking University)
Adversarial AttackVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: The study investigates transferable and stealthy adversarial attacks against large-scale vision-language models (VLMs) and proposes the Progressive Semantic Infusion (PSI) method.
Transformers are Inherently Succinct
Pascal Bergsträßer (RPTU Kaiserslautern-Landau), Anthony Widjaja Lin (RPTU Kaiserslautern-Landau)
Computational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformer
🎯 What it does: Prove that the transformer (UHAT) is more concise than LTL, RNN, and finite automata in representing languages, and provide the EXPSPACE-completeness conclusion for the verification problem.
Transformers as Measure-Theoretic Associative Memory: A Statistical Perspective and Minimax Optimality
Ryotaro Kawata (University of Tokyo), Taiji Suzuki (University of Tokyo)
OptimizationComputational EfficiencyRepresentation LearningTransformer
🎯 What it does: Studied the association memory (retrieval + prediction) tasks of learned softmax attention Transformers at the probabilistic measure level, and provided a complete proof of statistical learning theory and optimal sampling complexity.
Transformers as Unsupervised Learning Algorithms: A study on Gaussian Mixtures
Zhiheng Chen (Fudan University), Guanhua Fang (Fudan University)
Meta LearningTransformerTabular
🎯 What it does: Propose a meta-learning framework called TGMM based on Transformer, which can simultaneously solve Gaussian Mixture Model (GMM) estimation tasks with different numbers of components during inference;