ICLR 2026 Papers — Page 47
International Conference on Learning Representations · 5356 papers
Teaching VLMs to Admit Uncertainty in OCR from Lossy Visual Inputs
Shuhao Guan (University College Dublin), Derek Greene (University College Dublin)
RecognitionData SynthesisLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: This study proposes an uncertainty-aware OCR method, training visual-language models (VLMs) to insert <C></C> tags before and after potentially erroneous text segments when recognizing distorted documents, explicitly indicating the model's uncertainty.
TEDM: Time Series Forecasting with Elucidated Diffusion Models
Edgardo Solano Carrillo, Julia Niebling (German Aerospace Center)
Diffusion modelTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: In multivariate time series forecasting tasks, the TEDM framework is proposed, based on diffusion models and unifying the time axis with diffusion time, achieving efficient and accurate probabilistic prediction through data-driven noise and scale scheduling.
Tell me Habibi, is it Real or Fake?
Kartik Kuckreja (MBZUAI), Abhinav Dhall (Monash University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringDiffusion modelVideoMultimodalityBenchmarkAudio
🎯 What it does: Generated and released the first large-scale Arabic-English code-switching audio-visual deepfake dataset ArEnAV, covering 387k videos and 765 hours;
Temperature as a Meta-Policy: Adaptive Temperature in LLM Reinforcement Learning
Haoran Dang (Tsinghua University), Yan Lu (Microsoft Research Asia)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose a hierarchical reinforcement learning framework called TAMPO, which treats the sampling temperature as a learnable meta-policy, achieving online adaptive balance between exploration and exploitation through the temperature likelihood information of trajectories, thereby improving the reinforcement learning fine-tuning of large language models (LLMs).
TEMPFLOW-GRPO: WHEN TIMING MATTERS FOR GRPO IN FLOW MODELS
Xiaoxuan He (ZheJiang University), Bo Zhang (ZheJiang University)
GenerationReinforcement LearningFlow-based ModelImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose TempFlow-GRPO, a temporal-aware reinforcement learning framework for flow matching models, to improve the quality of text-to-image generation and alignment with human preferences.
Temporal Concept Dynamics in Diffusion Models via Prompt-Conditioned Interventions
Ada Görgün (Max Planck Institute for Informatics), Jonas Fischer (Max Planck Institute for Informatics)
GenerationPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Proposed the Prompt-Conditioned Intervention (PCI) framework, which analyzes the dynamic evolution of concepts over time by switching text prompts at different stages of the diffusion process, and introduces the Concept Insertion Success (CIS) metric to quantify the probability of successful concept insertion;
Temporal Generalization: A Reality Check
Divyam Madaan (New York University), Kyunghyun Cho (New York University)
Supervised Fine-TuningTextTime Series
🎯 What it does: Studied achieving temporal generalization using interpolation or extrapolation of past model parameters when future data is unavailable.
Temporal Geometry of Deep Networks: Hyperbolic Representations of Training Dynamics for Intrinsic Explainability
Ambarish Moharil (Eindhoven University of Technology)
Explainability and InterpretabilityRepresentation LearningGraph Neural NetworkImageTime Series
🎯 What it does: Construct and embed the time parameter graph of MLP training trajectories, utilizing the Poincaré ball to implement hypergeometric graph attention and self-evolving kernels, performing parameter's positive and negative weight regression and connectivity prediction;
Temporal Graph Thumbnail: Robust Representation Learning with Global Evolutionary Skeleton
Weining Shi (Xiamen University), Zhihong Zhang (Xiamen University)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: Proposed Temporal Graph Thumbnail (TGT), enhancing the robustness of representation learning by constructing a global evolutionary skeleton for temporal graphs.
Temporal Representations for Exploration: Learning Complex Exploratory Behavior without Extrinsic Rewards
Faisal Mohamed (Mila-Quebec AI Institute), Glen Berseth (Mila-Quebec AI Institute)
Representation LearningReinforcement LearningContrastive Learning
🎯 What it does: This paper proposes a self-encouraging exploration method (C-TeC) that utilizes representations obtained through time contrastive learning. These representations serve as reward signals, encouraging agents to visit states where future states are unpredictable, thereby learning complex behaviors without external rewards.
Temporal Slowness in Central Vision Drives Semantic Object Learning
Timothy Schaumlöffel (Goethe University Frankfurt), Jochen Triesch (Goethe University Frankfurt)
ClassificationRecognitionRepresentation LearningTransformerContrastive LearningVideo
🎯 What it does: Process 5 months of head-mounted video (Ego4D) with human-like visual experience, using state-of-the-art gaze prediction models to predict fixation points, then cropping the central visual region based on these predictions. Subsequently, train a temporal slow self-supervised learning model (MoCoV3) on these image sequences to learn slow-varying and semantically informative visual representations.
Temporal Sparse Autoencoders: Leveraging the Sequential Nature of Language for Interpretability
Usha Bhalla (Harvard University), Flavio Calmon (Harvard University)
Explainability and InterpretabilityAuto EncoderContrastive LearningText
🎯 What it does: Propose Temporal Sparse Autoencoders (T-SAE), improving the interpretability of sparse autoencoders in language models by introducing temporal contrastive loss on high-level features.
Temporal superposition and feature geometry of RNNs under memory demands
Pratyaksh Sharma (Imperial College London), Pedro A. M. Mediano
Representation LearningRecurrent Neural NetworkTime Series
🎯 What it does: Studied the feature representation geometry of recurrent neural networks under temporal (memory) pressure, proposed and analyzed the concept of temporal superposition, revealing how the hidden space is compressed and forgotten.
Temporally Detailed Hypergraph Neural ODE for Disease Progression Modeling
Tingsong Xiao (University of Florida), Zhe Jiang (University of Florida)
Graph Neural NetworkTabularTime SeriesBiomedical DataElectronic Health RecordsOrdinary Differential Equation
🎯 What it does: Proposed a learnable time-refined hypergraph neural ordinary differential equation model (TD-HNODE) based on clinically known progression paths for continuous-time disease progression prediction.
TEN-DM: Topology-Enhanced Diffusion Model for Spatio-Temporal Event Prediction
Yuxin Liu (University of California), Yuzhou Chen (University of California)
Graph Neural NetworkTransformerDiffusion modelGraphTime Series
🎯 What it does: Proposed the Topology-Enhanced Diffusion Model (TEN-DM) for predicting spatiotemporal point process events.
Tensor learning with orthogonal, Lorentz, and symplectic symmetries
Wilson G. Gregory (Johns Hopkins University), Soledad Villar (Johns Hopkins University)
Representation LearningBiomedical DataPhysics Related
🎯 What it does: Designed a generalizable tensor-to-tensor equivariant mapping expressibility and constructed the corresponding machine learning model;
Tequila: Trapping-free Ternary Quantization for Large Language Models
Hong Huang (City University Of Hong Kong), Dapeng Wu (City University Of Hong Kong)
Computational EfficiencyTransformerText
🎯 What it does: This paper proposes a ternary quantization method called Tequila for efficiently deploying large language models on edge devices, addressing the dead zone capture problem caused by traditional ternary quantization.
Terminal Velocity Matching
Linqi Zhou, Jiaming Song
GenerationTransformerDiffusion modelFlow-based ModelImage
🎯 What it does: Propose the Terminal Velocity Matching (TVM) method, training single-step/few-step generative models by matching terminal velocity
Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces
Mike A Merrill, Ludwig Schmidt (Stanford University)
AI Code AssistantLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed the Terminal-Bench 2.0 framework and 89 real command-line tasks to evaluate AI agents' ability to complete long-term technical tasks in terminal environments.
TerraFM: A Scalable Foundation Model for Unified Multisensor Earth Observation
Muhammad Sohail Danish (Mohamed bin Zayed University of Artificial Intelligence), Salman Khan (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationSegmentationTransformerContrastive LearningImageMultimodalityBenchmark
🎯 What it does: Proposed and trained a scalable multi-sensor Earth Observation foundation model called TerraFM;
TESSAR: Geometry-Aware Active Regression via Dynamic Voronoi Tessellation
Seong Jin Cho (Korea Institute of Oriental Medicine), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
OptimizationData-Centric LearningTabular
🎯 What it does: Proposed an active regression framework called TESSAR based on Voronoi partitioning, with the core metric being the Voronoi-based Least Disagree Metric (VLDM), and combined VLDM with distance and density scores to form a unified sampling strategy;
Test-Time Accuracy-Cost Control in Neural Simulators via Recurrent-Depth
Harris Abdul Majid (University of Edinburgh), Francesco Tudisco (University of Edinburgh)
OptimizationComputational EfficiencyTransformerDiffusion modelPhysics Related
🎯 What it does: Propose a pluggable neural simulation framework RecurrSim, which allows explicitly controlling the precision and computational cost of the simulation during inference by adjusting the recurrent depth K;
Test-Time Adaptation for LLM Agents via Environment Interaction
Arthur Chen (University of Waterloo), Caiming Xiong (Salesforce AI Research)
Domain AdaptationTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose two test-time adaptation methods: syntactic alignment (SA) and dynamic grounding (DG), to enhance the generalization ability of LLM agents in new environments.
Test-Time Adaptation without Source Data for Out-of-Domain Bioactivity Prediction
Yiming Yang (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
Domain AdaptationDrug DiscoveryGraph Neural NetworkContrastive LearningGraphBiomedical DataBenchmark
🎯 What it does: Propose a source-data-missing test-time adaptation framework TAB for cross-domain generalization in drug activity prediction;
Test-Time Alignment for Large Language Models via Textual Model Predictive Control
Kuang-Da Wang (National Yang Ming Chiao Tung University), Ping-Chun Hsieh (National Yang Ming Chiao Tung University)
OptimizationReinforcement Learning from Human FeedbackTransformerText
🎯 What it does: Propose a test-time alignment framework TMPC, which leverages the idea of model predictive control (MPC) to optimize the generation process of large language models, achieving alignment without parameter updates.
Test-time Domain Generalization for Image Super-resolution
Zaizuo Tang, Yu-Bin Yang (Nanjing University)
Super ResolutionDomain AdaptationConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: Proposed a test-time domain generalization framework for image super-resolution, MC-TTDG, which utilizes multiple codebooks to achieve pixel-level feature transfer, addressing the low-resolution issues of traditional style transfer in low-level visual tasks.
Test-Time Efficient Pretrained Model Portfolios for Time Series Forecasting
Mert Kayaalp (Dalle Molle Institute for Artificial Intelligence), Bernie Wang (Amazon Web Services)
Computational EfficiencyTransformerSupervised Fine-TuningMixture of ExpertsTime SeriesBenchmark
🎯 What it does: Built a combination of small-scale pre-trained time series prediction models (Chroma), which fine-tunes expert models on different data subsets and performs model selection or ensemble during testing, replacing a single large model.
Test-Time Iterative Error Correction for Efficient Diffusion Models
Yunshan Zhong (Hainan University), Yuxin Zhang (Xiamen University)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: Propose a general method for iterative error correction (IEC) during the testing phase of deployed efficient diffusion models, which can significantly improve generation quality;
Test-Time Matching: Unlocking Compositional Reasoning in Multimodal Models
Yinglun Zhu (University of California, Riverside), Fuzhi Tang (University of California, Riverside)
Representation LearningMultimodality
🎯 What it does: Proposed a new evaluation metric called GroupMatch and improved the existing GroupScore, then designed a self-supervised iterative training method called Test-Time Matching (TTM) to enhance the performance of multi-modal models in compositional reasoning tasks.
Test-Time Mixture of World Models for Embodied Agents in Dynamic Environments
Jinwoo Jang (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
Domain AdaptationKnowledge DistillationRobotic IntelligenceGraph Neural NetworkMixture of ExpertsWorld Model
🎯 What it does: Proposes the TMoW framework, which can dynamically mix world models during testing to adapt to new domains, and achieves continuous adaptation through multi-granularity prototype routing, test-time prototype refinement, and model incremental expansion based on mixing.
Test-Time Optimization of 3D Point Cloud LLM via Manifold-Aware In-Context Guidance and Refinement
Tiankai Chen (Southwest Jiaotong University), Xun Xu (Institute for Infocomm Research, A*STAR)
ClassificationAnomaly DetectionGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringPoint Cloud
🎯 What it does: Proposes the PGLLM framework, leveraging 3D point cloud graph to construct neighborhood prompts and refining confidence through contextual guidance via LLM, to achieve classification, OOD detection, and description tasks on 3D point clouds;
Test-Time Poisoned Sample Detection by Exploiting Shallow Malicious Matching in Backdoored CLIP
Zhengyao Song (Harbin Institute of Technology), Baoyuan Wu (Chinese University of Hong Kong)
Anomaly DetectionAdversarial AttackPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper studies how the CLIP model detects triggered images during the inference phase after being subjected to backdoor attacks, and proposes a detection method called Subspace Detection based on text subspaces;
TEST-TIME SCALING IN DIFFUSION LLMS VIA HIDDEN SEMI-AUTOREGRESSIVE EXPERTS
Jihoon Lee (Yonsei University), Amrit Singh Bedi (Carnegie Mellon University)
Computational EfficiencyLarge Language ModelMixture of ExpertsDiffusion modelTextBenchmark
🎯 What it does: For inference-time sequence generation in diffusion-based large language models (dLLMs), this paper proposes a training-free test-time scaling method called HEX, which significantly improves inference accuracy by integrating voting across semi-autoregressive decoding paths with different block sizes.
Test-Time Scaling with Reflective Generative Model
Zixiao Wang (University of Science and Technology of China), Hongtao Xie (University of Science and Technology of China)
Computational EfficiencyAI Code AssistantLarge Language ModelReinforcement LearningText
🎯 What it does: Designed and implemented a Reflective Generative Model (RGM) that, during testing, generates multiple reasoning paths through the same network while self-evaluating and selecting the optimal path, achieving controllable reasoning expansion for large language models.
Test-Time Training Done Right
Tianyuan Zhang (Massachusetts Institute of Technology), Hao Tan (Adobe Research)
Image TranslationGenerationData SynthesisTransformerImageVideoText
🎯 What it does: Propose the large-block test-time training framework LaCT, which can efficiently learn context on long sequences and has been validated on three tasks (novel view synthesis, language modeling, autoregressive video diffusion).
Test-time Verification via Optimal Transport: Coverage, ROC, & Sub-optimality
Arpan Mukherjee (Imperial College London), Deniz Gunduz (Imperial College London)
OptimizationComputational EfficiencyText
🎯 What it does: This paper views test-time verification as a sampling problem, using the optimal transport framework to uniformly analyze the interactions among generator coverage, validator ROC, and suboptimality of sampling algorithms.
Testing Fourier Sparsity via Implicit Sensing
Arijit Ghosh (Indian Statistical Institute), Manmatha Roy (Indian Statistical Institute)
OptimizationComputational Efficiency
🎯 What it does: This paper proposes a non-adaptive property testing algorithm for the Fourier sparsity of Boolean functions, and provides improved upper and lower bounds.
Testing Most Influential Sets
Lucas Darius Konrad, Nikolas Kuschnig (Monash University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTabularBenchmarkFinance Related
🎯 What it does: A rigorous statistical framework is proposed for the most influential data subsets in linear regression, deriving their exact influence formula and providing distributions under extreme value theory, thereby enabling hypothesis testing for over-influence.
TetraGT: Tetrahedral Geometry-Driven Explicit Token Interactions with Graph Transformer for Molecular Representation Learning
Jinjia Feng (Renmin University of China), Zongyang Qiu (BioMap Research)
Drug DiscoveryGraph Neural NetworkTransformerSupervised Fine-TuningGraphBenchmark
🎯 What it does: Propose a graph Transformer architecture named TetraGT, which models bond angles and torsion angles directly as structured tokens, leveraging triangular and tetrahedral geometric constraints to achieve high-quality molecular geometry prediction, and applies it to molecular property prediction.
Text summarization via global structure awareness
Jiaquan Zhang (University of Electronic Science and Technology of China), Yang Yang (University of Electronic Science and Technology of China)
GenerationGraph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Propose a global structure-aware text summarization framework called GloSA-sum based on topological data analysis;
Text-Aware Image Restoration with Diffusion Models
Jaewon Min (KAIST AI), Seungryong Kim (KAIST AI)
RestorationConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelImageText
🎯 What it does: Proposes the Text-Aware Image Restoration (TAIR) task, focusing on text restoration in low-quality images;
Text-to-3D by Stitching a Multi-view Reconstruction Network to a Video Generator
Hyojun Go (ETH Zurich), Konrad Schindler (ETH Zurich)
GenerationReinforcement LearningDiffusion modelVideoTextMesh
🎯 What it does: Achieve end-to-end text-to-3D generation by model stitching (model stitching) between text-to-video latent diffusion models and pre-trained multi-view 3D reconstruction networks, followed by direct reward finetuning (direct reward finetuning);
Text2Arch: A Dataset for Generating Scientific Architecture Diagrams from Natural Language Descriptions
Shivank Garg (IIT Roorkee), Manish Gupta (Microsoft)
Object DetectionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextGraphBenchmark
🎯 What it does: This paper proposes a system called TEXT2ARCH for generating scientific architecture diagrams based on natural language descriptions. The core idea is to first convert text into intermediate DOT graph code, and then render the architecture diagram using a DOT compiler;
Text2Grad: Reinforcement Learning from Natural Language Feedback
Hanyang Wang (University of Chicago), Dongmei Zhang (Microsoft)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes TEXT2GRAD, a framework that converts natural language feedback into span-level gradients and directly applies them to reinforcement learning.
Text2Interact: High-Fidelity and Diverse Text-to-Two-Person Interaction Generation
Qingxuan Wu (University of Pennsylvania), Lingjie Liu (Snap Inc)
GenerationData SynthesisTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelVideoText
🎯 What it does: Achieved text-driven, high-fidelity two-person interactive motion generation by synthesizing single-person motion with LLM-generated text descriptions, combined with word-level attention and adaptive interaction loss.
Textual Bayes: Quantifying Prompt Uncertainty in LLM-Based Systems
Brendan Leigh Ross (Layer 6 AI), Jesse C. Cresswell (Layer 6 AI)
Explainability and InterpretabilityLarge Language ModelText
🎯 What it does: Propose the Textual Bayes framework, treating LLM prompts as Bayesian text parameters, utilizing Bayesian inference to sample prompts, thereby achieving uncertainty quantification of LLM outputs;
Textual Equilibrium Propagation for Deep Compound AI Systems
Minghui Chen (Nanyang Technological University), Xiaoxiao Li (Stanford University)
OptimizationLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A framework called Textual Equilibrium Propagation (TEP) based on local learning is studied for optimizing text gradients in deep composite AI systems.
Texture Vector-Quantization and Reconstruction Aware Prediction for Generative Super-Resolution
Qifan Li (University of Electronic Science and Technology of China), Shuhang Gu (University of Electronic Science and Technology of China)
Super ResolutionAuto EncoderImage
🎯 What it does: What was done
TGM: A Modular and Efficient Library for Machine Learning on Temporal Graphs
Jacob Chmura (Mila - Quebec AI Institute), Reihaneh Rabbany (Mila - Quebec AI Institute)
Graph
🎯 What it does: Proposed and implemented TGM, a unified and efficient open-source framework for continuous-time and discrete-time dynamic graph learning.
The Price of Amortized inference in Sparse Autoencoders
Wenjie Sun (Mohamed bin Zayed University of Artificial Intelligence), Lijie Hu (Mohamed bin Zayed University of Artificial Intelligence)
Computational EfficiencyRepresentation LearningAuto EncoderText
🎯 What it does: Investigate various pathological phenomena caused by global amortized inference in sparse autoencoders (SAE) and propose local amortized SAE (LocA-SSAE) to alleviate these issues
The Achilles’ Heel of LLMs: How Altering a Handful of Neurons Can Cripple Language Abilities
Zixuan Qin (Renmin University of China), Yifan Sun (Renmin University of China)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper systematically identifies extremely sparse and critical neurons in large language models by combining noise perturbation with causal verification, and demonstrates that their deactivation can lead to an instantaneous collapse in model performance.
The Alignment Auditor: A Bayesian Framework for Verifying and Refining LLM Objectives
Matthieu Bou (Imperial College London), Sonali Parbhoo (Imperial College London)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposed a three-stage alignment auditor framework (Alignment Auditor) that utilizes Bayesian inverse reinforcement learning (IRL) to recover reward posterior from contrastive data between experts and baselines, quantifies reward uncertainty, identifies shortcut paths and out-of-distribution (OOD) inputs through uncertainty diagnosis, and finally validates its practicality at the policy level by fine-tuning in RLHF using the posterior mean as the reward signal.
The Alignment Waltz: Jointly Training Agents to Collaborate for Safety
Jingyu Zhang (Meta Superintelligence Labs), Hongyuan Zhan (Meta Superintelligence Labs)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Propose WALTZRL, a multi-agent reinforcement learning framework that jointly trains dialogue agents and feedback agents, enabling them to collaboratively generate responses that are both safe and not overly rejected through multi-round interactions.
The Art of Scaling Reinforcement Learning Compute for LLMs
Fnu Devvrit, Rishabh Agarwal (Meta)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper addresses the computational scalability challenges in reinforcement learning (RL) training for large language models (LLMs), proposing a predictable compute-performance scaling framework. Based on systematic experiments, it designs an scalable RL training scheme called SCALERL, achieving stable performance improvements at scales up to 100,000 GPU hours.
The Choice of Divergence: A Neglected Key to Mitigating Diversity Collapse in Reinforcement Learning with Verifiable Reward
Long Li (Fudan University), Yuan Qi (Fudan University)
Data-Centric LearningLarge Language ModelReinforcement LearningText
🎯 What it does: Propose the Diversity-Preserving Hybrid RL (DPH-RL) framework, addressing diversity collapse and catastrophic forgetting in RLVR through f-divergence regularization.
The CoT Encyclopedia: Analyzing, Predicting, and Controlling how a Reasoning Model will Think
Seongyun Lee (KAIST AI), Minjoon Seo (KAIST AI)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed COT ENCYCLOPEDIA, a framework based on bottom-up clustering, to automatically identify, organize, and explain the chain-of-thought (CoT) strategies of large language models, and guide the model to improve reasoning effectiveness based on the identified strategies;
The Counting Power of Transformers
Marco Sälzer (RPTU Kaiserslautern-Landau), Anthony Widjaja Lin (MPI-SWS)
Representation LearningTransformerText
🎯 What it does: This paper provides a formal framework to study the counting ability of transformers, proving that transformers can express highly non-linear counting attributes beyond existing (semi) linear counting attributes.
The Coverage Principle: How Pre-Training Enables Post-Training
Fan Chen (MIT), Dylan J Foster
Representation LearningData-Centric LearningLarge Language ModelTextGraph
🎯 What it does: Investigated and theorized the 'coverage' metric during the pretraining process of language models, proving it to be a better predictor of the success rate of downstream tasks;
The Curious Case of In-Training Compression of State Space Models
Makram Chahine (Massachusetts Institute of Technology), T. Konstantin Rusch (ELLIS Institute)
Computational EfficiencyKnowledge DistillationImageTextSequential
🎯 What it does: Proposed COMPRESSM, a compression framework that performs balanced truncation on state-space models (SSMs) during training, reducing hidden state dimensions without significant performance loss.
The Deleuzian Representation Hypothesis
Clément Cornet (University of Paris-Saclay), Hervé Le Borgne (University of Paris-Saclay)
Explainability and InterpretabilityRepresentation LearningTransformerContrastive LearningImageTextMultimodalityAudio
🎯 What it does: Proposed an unsupervised concept extraction method based on clustering differences, which uses contrastive activation differences for clustering and enhances diversity through inverse skewness weighting;
The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs
Zichen Wen (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
Safty and PrivacyAdversarial AttackPrompt EngineeringDiffusion modelText
🎯 What it does: This paper systematically studies the security vulnerabilities of diffusion-based large language models (dLLMs), proposing and implementing an automated jailbreak framework called DIJA, which can bypass existing security mechanisms by interpolating masks and text to generate harmful outputs.
The Diffusion Duality, Chapter II: $\Psi$-Samplers and Efficient Curriculum
Justin Deschenaux (EPFL), Subham Sekhar Sahoo (Cornell Tech)
GenerationDiffusion modelImageText
🎯 What it does: Developed the Psi posterior with the corresponding Predictor-Corrector sampler, and proposed an efficient memory-friendly training curriculum specifically for the unified discrete diffusion model (Duo++)
The Effect of Attention Head Count on Transformer Approximation
Penghao Yu (National University of Singapore), Qianxiao Li (National University of Singapore)
RetrievalTransformerImageText
🎯 What it does: This paper investigates the impact of the number of attention heads in Transformers on their approximation capability, proposes a densely generalizable D retrieval task, and provides upper and lower bounds for Transformers on this task.
THE END OF MANUAL DECODING: TOWARDS TRULY END-TO-END LANGUAGE MODELS
Zhichao Wang (Chinese University of Hong Kong), Yan Wang (Tencent AILab)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the AutoDeco framework, integrating a lightweight prediction head into the Transformer architecture to enable real-time estimation of temperature and top-p, thus achieving adaptive decoding within a single forward pass.
The Expressive Limits of Diagonal SSMs for State-Tracking
Mehran Shakerinava (McGill University), Sarath Chandar (Polytechnique Montréal)
Representation Learning
🎯 What it does: Investigated and demonstrated the correspondence between the expressiveness of input-dependent complex diagonal state space models (DCD) and the length of solvable group subnormal chains in state tracking tasks, and experimentally validated the gap between expressiveness and learnability.
The False Promise of Zero-Shot Super-Resolution in Machine-Learned Operators
Mansi Sakarvadia (University of Chicago), Michael W. Mahoney (Lawrence Berkeley National Laboratory)
Super ResolutionPhysics Related
🎯 What it does: Evaluate the performance of machine learning operators on zero-shot super-resolution and propose a multi-resolution training scheme that requires only a few high-resolution samples
The First Impression Problem: Internal Bias Triggers Overthinking in Reasoning Models
Renfei Dang (Nanjing University), Jiajun Chen (Nanjing University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Investigate the root cause of excessive thinking in large language models during reasoning tasks, propose and quantify internal bias (model's intuitive answers), demonstrate causality through adversarial interventions (removing questions, injecting bias), and use attention analysis to reveal how internal bias is reactivated during reflection, leading to redundant reasoning.
The Forecast After the Forecast: A Post-Processing Shift in Time Series
Daojun Liang (Qilu University of Technology), Shuo Li (Case Western Reserve University)
Data-Centric LearningTime Series
🎯 What it does: Propose δ-Adapter, which enhances accuracy and calibration by making minor input adjustments and output residual corrections on frozen time series prediction models.
The Gaussian-Head OFL Family: One-Shot Federated Learning from Client Global Statistics
Fabio Turazza (University of Modena and Reggio Emilia), Marco Mamei (University of Modena and Reggio Emilia)
ClassificationFederated LearningKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: Proposed a one-round federated learning framework GH-OFL, achieving data-agnostic model inference by transmitting only clients' category counts and first/second-order moments.
The Geometry of LLM Quantization: GPTQ as Babai's Nearest Plane Algorithm
Jiale Chen (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)
OptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Studied the geometric interpretation of the GPTQ weight quantization method, equating it to the Babai closest vector algorithm without basis scaling, derived error upper bounds, and proposed an improved unclipped quantization scheme and efficient GPU inference kernel.
The Geometry of Reasoning: Flowing Logics in Representation Space
Yufa Zhou (Duke University), Anru Zhang
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Study the geometric dynamics of LLM reasoning, modeling reasoning as a smooth flow in the embedding space and revealing logical structures through geometric quantities such as velocity and curvature.
The Hot Mess of AI: How Does Misalignment Scale With Model Intelligence and Task Complexity?
Alexander Hägele (Anthropic Fellows Program), Jascha Sohl-Dickstein (Anthropic)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper analyzes the error types of AI models across different scales, task complexities, and reasoning lengths using bias-variance decomposition, quantifies error-incoherence, and investigates whether more powerful models fail systematically deviating from the target or exhibiting random 'thermal noise' behavior.
The Human Brain as a Dynamic Mixture of Expert Models in Video Understanding
Christina Sartzetaki (University of Amsterdam), Iris Groen (University of Amsterdam)
Explainability and InterpretabilityRepresentation LearningMixture of ExpertsVideoBiomedical Data
🎯 What it does: This paper investigates the spatiotemporal characteristics of visual processing through a large-scale model-brain alignment evaluation involving over 100 deep visual models and dynamic EEG video recordings.
The Ideation-Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas
Chenglei Si (Stanford University), Diyi Yang (Stanford University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper conducted a randomized controlled trial, recruiting 43 expert researchers to execute research ideas generated by Claude-3.5-Sonnet and ideas proposed by human experts. The generated code and papers were subjected to blind review.
The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs
Akshit Sinha (University of Cambridge), Jonas Geiping (Max Planck Institute for Intelligent Systems)
Data SynthesisTransformerReinforcement LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper quantifies the ability of large language models (LLMs) in long-term task execution by providing explicit plans and knowledge, and investigates the impact of model scale, context self-conditioning, and thinking patterns on execution performance.
The Imitation Game: Turing Machine Imitator is Length Generalizable Reasoner
Zhouqi Hua (Fudan University), Kai Chen (Shanghai AI Laboratory)
Large Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposes a Chain-of-Thought (CoT) generation framework called TAIL based on Turing machine simulation, enabling large language models to learn algorithm execution processes scalable to arbitrary lengths during training, thus achieving cross-length reasoning generalization.
The Intricate Dance of Prompt Complexity, Quality, Diversity and Consistency in T2I Models
Xiaofeng Zhang (Mila Quebec AI Institute), Adriana Romero-Soriano (Meta)
GenerationData SynthesisPrompt EngineeringDiffusion modelImageText
🎯 What it does: This paper proposes a new evaluation framework to systematically study the quality, diversity, and consistency of synthetic data generated by text-to-image (T2I) models under different prompt complexities, and compares them with real data.
The Lattice Geometry of Neural Network Quantization: A Short Equivalence Proof of GPTQ and Babai's Algorithm
Johann Birnick (University of California San Diego)
CompressionOptimizationComputational Efficiency
🎯 What it does: This paper proves that the GPTQ algorithm for post-training quantization of neural networks is essentially equivalent to the classical Babai's nearest plane algorithm, and explains their similarity from a geometric perspective; meanwhile, it points out that quantization quality can be further improved through lattice basis transformation;
The Lattice Representation Hypothesis of Large Language Models
Bo Xiong (Stanford University)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelText
🎯 What it does: Proposes the Lattice Representation Hypothesis, elucidating the unified framework between the linear attribute directions in the embedding space of large language models and Formal Concept Analysis, and verifies this framework can recover the complete concept lattice through WordNet sub-hierarchy;
The Less You Depend, The More You Learn: Synthesizing Novel Views from Sparse, Unposed Images without Any 3D Knowledge
Haoru Wang (Peking University), Baoquan Chen (Peking University)
Data SynthesisPose EstimationTransformerNeural Radiance FieldAuto EncoderImage
🎯 What it does: This paper investigates the relationship between the dependence on explicit 3D knowledge and model scalability in Neural View Synthesis (NVS), and proposes a Transformer framework named UP-LVSM that is entirely based on 2D images without requiring explicit scene structure or camera pose annotations.
The Lie of the Average: How Class Incremental Learning Evaluation Deceives You?
Guannan Lai (Nanjing University), Han-Jia Ye (Nanjing University)
ClassificationVision Language ModelContrastive LearningImageBenchmark
🎯 What it does: Proposed a distribution and generalization evaluation protocol called EDGE based on extreme sequences, improving the evaluation method for Class Incremental Learning (CIL).
The Limits of Inference Scaling Through Resampling
Benedikt Stroebl (Princeton University), Arvind Narayanan (Princeton University)
AI Code AssistantTextBenchmark
🎯 What it does: Explored the limits of improving model performance through resampling during inference when incomplete verifiers (e.g., unit tests) are present, proving that weak models cannot match the single-sample accuracy of strong models even with infinite sampling;
The Logical Expressiveness of Topological Neural Networks
Amirreza Akbari (Aalto University), Vikas K Garg
Representation LearningGraph Neural Network
🎯 What it does: Proposed a high-order Weisfeiler–Leman (k-CCWL) algorithm for topological neural networks (TNN), constructed the corresponding topological counting logic (TC^k) and topological k+2-coloring game, and proved that the three have equivalent expressive power in distinguishing non-isomorphic unified combinatorial complexes (ACC).
The Markovian Thinker: Architecture-Agnostic Linear Scaling of Reasoning
Milad Aghajohari (Mila), Siva Reddy (Mila)
Computational EfficiencyReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposed and implemented the Delethink algorithm, transforming large language models into 'Markovian Thinkers' to achieve linear computation and constant memory usage for inference and reinforcement learning.
The Matthew Effect of AI Programming Assistants: A Hidden Bias in Software Evolution
Fei Gu (City University of Hong Kong), Hongzong LI
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: At the ICLR 2026 conference, the authors conducted large-scale experiments to evaluate the impact of AI programming assistants on mainstream and niche programming languages and frameworks.
The Mind's Transformer: Computational Neuroanatomy of LLM-Brain Alignment
Cheng-Yeh Chen (Georgia Institute of Technology), Raghupathy Sivakumar (Georgia Institute of Technology)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Systematically decompose the 13 intermediate states of the Transformer block, map them to brain fMRI data, and propose the MindTransformer framework to enhance brain-model alignment
The Natural Geometry of Code: Hyperbolic Representation Learning for Program Reasoning
Weilin Zhou (Nanjing Tech University)
Representation LearningAI Code AssistantGraph Neural NetworkTextGraph
🎯 What it does: Proposes a framework called HypeCodeNet for learning code representations in hyperbolic geometry spaces with negative curvature.
The Open Proof Corpus: A Large-Scale Study of LLM-Generated Mathematical Proofs
Jasper Dekoninck (ETH Zurich), Martin Vechev (ETH Zurich)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Created the Open Proof Corpus (OPC), a dataset containing over 5,000 mathematical proofs generated by LLMs and evaluated by humans, and used it to systematically study the performance of LLMs in proof generation and evaluation.
THE PATH OF LEAST RESISTANCE: GUIDING LLM REASONING TRAJECTORIES WITH PREFIX CONSENSUS
Ishan Jindal (Fujitsu Research India), SACHIN DEV SHARMA
Computational EfficiencyLarge Language ModelText
🎯 What it does: Propose the PoLR method, which samples short prefixes for clustering during inference and only performs full reasoning on the main cluster, thereby reducing the computational cost of Self-Consistency.
The Pensieve Paradigm: Stateful Language Models Mastering Their Own Context
Xiaoyuan Liu (Chinese University of Hong Kong), Yan Wang (Tencent)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Developed a stateful language model called StateLM that can autonomously manage context and achieve self-context engineering through tool calls;
The Polar Express: Optimal Matrix Sign Methods and their Application to the Muon Algorithm
Noah Amsel (New York University), Robert M. Gower (Flatiron Institute)
OptimizationComputational EfficiencyText
🎯 What it does: Propose a new GPU-friendly matrix polarity decomposition algorithm called Polar Express to accelerate the approximation of matrix sign functions in the Muon optimizer.
The Potential of CoT for Reasoning: A Closer Look at Trace Dynamics
Gregor Bachmann, Moin Nabi (Apple)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper introduces a potential metric to quantitatively analyze the chain-of-thought (CoT) process in large language models and investigates the transferability of CoT across different models.
The Potential of Second-Order Optimization for LLMs: A Study with Full Gauss-Newton
Natalie Abreu (Harvard University), Depen Morwani (Harvard University)
OptimizationTransformerText
🎯 What it does: Investigating the performance upper limit of full Gauss-Newton preconditioning in large language model (LLM) training
The Power of Small Initialization in Noisy Low-Tubal-Rank Tensor Recovery
Zhiyu Liu (Shenyang Institute of Automation, Chinese Academy of Sciences), Yao Wang (Xi'an Jiaotong University)
RestorationOptimizationImageVideo
🎯 What it does: The study recovers low-tubal-rank tensors under noisy measurements within the t-product framework using factor gradient descent (FGD), investigates the effects of small initialization in overparameterized scenarios, and provides theoretical error upper bounds.
The Price of Robustness: Stable Classifiers Need Overparameterization
Jonas von Berg (Ludwig-Maximilians-Universitat Munchen), Gitta Kutyniok (Aleph Alpha Research)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes and quantifies the 'class stability' of classifiers and their normalized co-stability, provides generalization upper bounds for both finite and infinite hypothesis classes, derives robustness laws for discrete classification, and demonstrates that over-parameterization is a necessary condition for achieving high stability and good generalization.
The Quest for Efficient Reasoning: A Data-Centric Benchmark to CoT Distillation
Ruichen Zhang (University of North Carolina at Chapel Hill), Tianlong Chen (University of North Carolina at Chapel Hill)
Knowledge DistillationData-Centric LearningTransformerTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the DC-CoT benchmark to systematically evaluate the impact of data augmentation, filtering, and mixing in chain-of-thought (CoT) distillation, and conducted extensive experiments on multiple teacher-student combinations.
The Quest for Generalizable Motion Generation: Data, Model, and Evaluation
Jing Lin (Nanyang Technological University), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationTransformerVision Language ModelDiffusion modelFlow-based ModelVideoTextMultimodalityBenchmark
🎯 What it does: Proposed a large-scale dataset ViMoGen-228K, a cross-modal fusion ViMoGen model and its lightweight version ViMoGen-light, and constructed the MBench fine-grained evaluation benchmark to enhance the generalization ability of text-driven 3D human motion generation.
The Rank and Gradient Lost in Non-stationarity: Sample Weight Decay for Mitigating Plasticity Loss in Reinforcement Learning
Zihao Wu (Tianjin University), Jianye HAO
OptimizationReinforcement Learning
🎯 What it does: Identify plasticity loss in reinforcement learning through theoretical analysis and propose the Sample Weight Decay (SWD) method to alleviate the gradient decay problem.
The Sample Complexity of Online Reinforcement Learning: A Multi-model Perspective
Michael Muehlebach (Max Planck Institute for Intelligent Systems), Michael I. Jordan (Inria Paris)
Reinforcement Learning
🎯 What it does: This paper proposes a class of online non-periodic reinforcement learning algorithms and provides an upper bound on the frequency policy regret in continuous state-action spaces under nonlinear dynamic systems. By conducting a unified analysis across three scenarios: finite model sets, compressible model sets, and parameterized models, the corresponding sample complexity is proven: for finite model sets, the regret upper bound is O(du ln(N/Δ)+du ln(m/Δ)); for parameterized models, the regret upper bound is O(√(d N p u)), thereby achieving near-optimal sample efficiency theoretically.
The Seismic Wavefield Common Task Framework
Alexey Yermakov (University of Washington), J. Nathan Kutz (University of Washington)
Recurrent Neural NetworkTime SeriesBenchmarkPhysics Related
🎯 What it does: Proposed a generic machine learning task framework (CTF) for seismic wavefields, evaluated and benchmarked on three-scale public datasets;