ICLR 2026 Papers — Page 48
International Conference on Learning Representations · 5356 papers
THE SELF-RE-WATERMARKING TRAP: FROM EXPLOIT TO RESILIENCE
Vithurabiman Senthuran (Deakin University), Uthayasanker Thayasivam (University of Moratuwa)
Safty and PrivacyAdversarial AttackConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This paper proposes a robust watermarking framework that defends against malicious re-embedding attacks by the same encoder, utilizing Lipschitz constraints and self-watermark adversarial training;
The Serial Scaling Hypothesis
Yuxi Liu (University Of California Berkeley), Yutong Bai (University Of California Berkeley)
Computational EfficiencyTransformerDiffusion modelTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes and proves the 'Serial Scaling Hypothesis (SSH)', linking inherently sequential problems in machine learning (such as cellular automata, physics simulations, reinforcement learning, and mathematical question answering) with the TC class in complexity theory, and demonstrates that traditional parallel architectures (such as Transformers, SSMs, and diffusion models with TC0 backgrounds) cannot solve these inherently sequential tasks;
The Shape of Adversarial Influence: Characterizing LLM Latent Spaces with Persistent Homology
Aideen Fay (Microsoft Security Response Center), Anthea Monod (Imperial College London)
Representation LearningAdversarial AttackTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper conducts topological analysis of the internal representation space of large language models (LLMs) using persistent homology, investigating the geometric impacts of two distinct attack methods (indirect prompt injection and backdoor fine-tuning) on hidden layer representations, and proposes a topology-based attack fingerprint through topological compression.
The Softmax Bottleneck Does Not Limit the Probabilities of the Most Likely Tokens
Ronen Basri (Weizmann Institute of Science), David Jacobs
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper investigates the softmax bottleneck caused by the transformer output projection matrix (OPM), presents theoretical bounds, and experimentally verifies that both randomly initialized and trained OPMs can accurately specify a large number of high-probability tokens.
The Spacetime of Diffusion Models: An Information Geometry Perspective
Rafal Karczewski, Vikas K Garg
GenerationDrug DiscoveryDiffusion modelImageBiomedical DataStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose to view the noise space of diffusion models as a spatiotemporal statistical manifold and construct analytical geodesics using Fisher-Rao geometry;
The State of Reinforcement Finetuning for Transformer-based Agents
Shengchao Hu (Shanghai Jiao Tong University), Dacheng Tao (Nanyang Technological University)
TransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringSequential
🎯 What it does: Investigated reinforcement learning fine-tuning (RFT) methods for Transformer-based agents in meta-RL environments, and proposed Q‑guided Policy Optimization (QP), a lightweight improvement;
The Tool Decathlon: Benchmarking Language Agents for Diverse, Realistic, and Long-Horizon Task Execution
Junlong Li (Hong Kong University of Science and Technology), Junxian He (Hong Kong University of Science and Technology)
Agentic AIPrompt EngineeringTextSequentialBenchmark
🎯 What it does: Proposed and implemented the Tool Decathlon (TOOLATHLON) benchmark to evaluate the tool calling and task execution capabilities of language models in real-world, long-sequence, multi-application environments;
The Tutor-Pupil Augmentation: Enhancing Learning and Interpretability via Input Corrections
Darya Biparva (University of Minnesota), Donatello Materassi (University of Minnesota)
Explainability and InterpretabilityAuto EncoderImage
🎯 What it does: Propose the Tutor-Pupil structure, which improves accuracy while preserving interpretability by applying minimal input correction to a fixed main model (Pupil) through an auxiliary model (Tutor);
The Unseen Bias: How Norm Discrepancy in Pre-Norm MLLMs Leads to Visual Information Loss
Bozhou Li (Peking University), Wentao Zhang (Peking University)
Representation LearningTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: To address the L2 norm imbalance between visual and textual features caused by the Pre-Norm architecture in multimodal large language models (MLLMs), this work proposes inserting a specially initialized LayerNorm layer after the visual projector, combined with a Global Weight Compensation (GWC) mechanism, to achieve norm alignment. This mitigates the 'representation inertia' of visual features and enhances cross-modal information fusion.
The Unseen Frontier: Pushing the Limits of LLM Sparsity with Surrogate-Free ADMM
Kwanhee Lee (POSTECH), Namhoon Lee (POSTECH)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Developed an ADMM-based proxy-free sparsification framework called ELSA, which can increase the sparsity of LLMs to over 90% while maintaining language modeling performance.
The Value of Information in Human-AI Decision-making
Ziyang Guo (Northwestern University), Jessica Hullman (Northwestern University)
Explainability and InterpretabilityImageVideoTextTabularBiomedical Data
🎯 What it does: Propose a framework based on Bayesian decision theory to quantify information value in human-machine collaborative decision-making, and introduce global and instance-level complementary information value (ACIV, ILIV) metrics; design a new explanation method called ILIV-SHAP based on ILIV; validate the framework and explanation method through online human-machine collaborative experiments, demonstrating improved decision performance; and demonstrate the framework's application in model selection and feature importance analysis in tasks such as chest X-ray diagnosis and deepfake detection.
THEMIS: Towards Holistic Evaluation of MLLMs for Scientific Paper Fraud Forensics
Tzu-Yen Ma (Beijing University of Posts and Telecommunications), Haihong E (Beijing University of Posts and Telecommunications)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes a multimodal benchmark named THEMIS for evaluating the expert-level capabilities of large language models in visual fraud reasoning within academic papers.
Theoretical Analysis of Contrastive Learning under Imbalanced Data: From Training Dynamics to a Pruning Solution
Haixu Liao (New Jersey Institute of Technology), Shuai Zhang (New Jersey Institute of Technology)
Representation LearningData-Centric LearningTransformerContrastive LearningImage
🎯 What it does: Theoretically analyze the contrastive learning training dynamics of the Transformer-MLP model under imbalanced data, and propose an amplitude clipping method to enhance minority feature learning.
Theoretical Guarantees for Causal Discovery on Large Random Graphs
Mathieu Chevalley (GSK.ai), Patrick Schwab (GSK.ai)
Explainability and InterpretabilityGraph
🎯 What it does: Studies the finite-dimensional convergence properties of the false negative rate (FNR) in causal structure discovery over random graphs (ER, BA) under assumptions of random single-variable intervention and ε-intervention credibility.
Theoretical Modeling of Large Language Model Self-Improvement Training Dynamics Through Solver-Verifier Gap
Yifan Sun (Shanghai University of Finance and Economics), Jiaye Teng (Shanghai University of Finance and Economics)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkStochastic Differential Equation
🎯 What it does: This paper constructs a theoretical model based on the gap between the solver and verifier to characterize the training dynamics of large language models' self-improvement, and verifies its exponential convergence properties.
Theory of Scaling Laws for In-Context Regression: Depth, Width, Context and Time
Blake Bordelon (Harvard University), Cengiz Pehlevan (Harvard University)
OptimizationComputational EfficiencyMeta LearningTransformerTabular
🎯 What it does: This paper establishes and analyzes the in-context learning (ICL) performance of deep linear attention models under three different covariance structures (ISO, FS, RRS), deriving scaling relationships for width, depth, context length, and pre-training time;
Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration?
Pingyue Zhang (Northwestern University), Manling Li (Stanford University)
Representation LearningTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelWorld ModelMultimodality
🎯 What it does: Proposed and evaluated a framework called Theory of Space, investigating the ability of multimodal large models to actively explore, build, revise, and utilize spatial beliefs in partially observable environments.
Theory-Grounded Evaluation of Human-Like Fallacy Patterns in LLM Reasoning
Andrew Keenan Richardson (University of Oxford), Philipp Koralus (University of Oxford)
Explainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper uses the Erotetic Theory of Reasoning (ETR) and its open-source implementation PyETR to automatically generate 383 formal reasoning problems, evaluates reasoning errors in 38 large language models, and investigates whether their error patterns align with common human reasoning fallacies.
There and Back Again: On the relation between Noise and Image Inversions in Diffusion Models
Łukasz Staniszewski (Warsaw University of Technology), Kamil Deja (Warsaw University of Technology)
GenerationDiffusion modelScore-based ModelImage
🎯 What it does: This paper systematically analyzes the DDIM inversion process, revealing that the potential noise generated does not follow a Gaussian distribution, particularly leading to a lack of diversity in latent variables in image smooth regions. It proposes replacing the initial steps with the forward diffusion process to eliminate these correlations, thereby improving the quality of interpolation and editing.
There is No VAE: End-to-End Pixel-Space Generative Modeling via Self-Supervised Pre-Training
Jiachen Lei (AMAP, Alibaba Group), Xiangxiang Chu (AMAP, Alibaba Group)
GenerationTransformerDiffusion modelScore-based ModelContrastive LearningImage
🎯 What it does: Proposes a two-stage self-supervised pre-training and end-to-end fine-tuning framework for training diffusion models and consistency models in pixel space,
There Was Never a Bottleneck in Concept Bottleneck Models
Antonio Almudévar (ViVoLab, I3A University of Zaragoza), Alfonso Ortega (ViVoLab, I3A University of Zaragoza)
Explainability and InterpretabilityRepresentation LearningImageBenchmark
🎯 What it does: Propose Minimal Concept Bottleneck Models (MCBMs), which incorporate information bottleneck (IB) constraints into Concept Bottleneck Models (CBMs), explicitly limiting each latent variable to retain only the minimal sufficient statistics corresponding to its concept, thereby eliminating information leakage issues;
Thicker and Quicker: The Jumbo Token for Fast Plain Vision Transformers
Anthony Fuller (Carleton University), James R Green (Carleton University)
ClassificationSegmentationRetrievalTransformerContrastive LearningImageTextMultimodalityTime Series
🎯 What it does: Propose an architecture that integrates a wide Jumbo token (J times the width) into a standard ViT, maintaining pure attention and non-hierarchical characteristics, significantly enhancing model capacity and performance.
Think in Parallel, Answer as One: Logit Averaging for Open-Ended Reasoning
Haonan Wang (National University Of Singapore), Tianyu Pang (National University Of Singapore)
Computational EfficiencyLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose a training-agnostic, pluggable parallel inference decoding strategy called THINKMERGE, which generates a single answer by averaging logits from multiple reasoning paths during the answer stage;
Think Then Embed: Generative Context Improves Multimodal Embedding
Xuanming Cui (University of Central Florida), Xiangjun Fan (Meta)
ClassificationRetrievalKnowledge DistillationTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the Think-Then-Embed (TTE) framework, which first generates task-instruction-related reasoning (ECR) using a multimodal large language model, and then encodes the original input along with the reasoning results into a generic multimodal embedding.
Think-While-Generating: On-the-Fly Reasoning for Personalized Long-Form Generation
Chengbing Wang (University of Science and Technology of China), Tat-Seng Chua (National University of Singapore)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a framework named FlyThinker, called 'thinking while generating,' to achieve user-personalized long-text generation.
Thinking as Society: Multi-Social-Agent Self-Distillation for Multimodal Misinformation Detection
Yifei Gao (Tianjin University), Anan Liu (Tianjin University)
Data SynthesisAnomaly DetectionKnowledge DistillationTransformerLarge Language ModelReinforcement LearningMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a multi-social-agent self-distillation framework, which aggregates multi-perspective reasoning from multi-role LLMs to generate social chain-of-thought data, thereby achieving collective reasoning within a single model to detect multi-modal misinformation.
Thinking on the Fly: Test-Time Reasoning Enhancement via Latent Thought Policy Optimization
Wengao Ye (University of Oxford), Lianlei Shan (University of Chinese Academy of Sciences)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: During the testing phase of reasoning LLMs, online reinforcement learning is applied directly in the latent space on 'thought' vectors to enhance reasoning performance
Thinking with Camera: A Unified Multimodal Model for Camera-Centric Understanding and Generation
Kang Liao, Chen Change Loy
GenerationPose EstimationTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes a unified multimodal model called Puffin that can simultaneously perform camera parameter estimation (understanding) and image generation based on camera perspective (generation).
Thinking-Free Policy Initialization Makes Distilled Reasoning Models More Effective and Efficient Reasoners
Xin Xu (Hong Kong University of Science and Technology), Can Yang (Hong Kong University of Science and Technology)
Computational EfficiencyKnowledge DistillationReinforcement LearningTextChain-of-Thought
🎯 What it does: Designed and implemented a training phase called Thinking-Free Policy Initialization (TFPI), which uses 'thinking-free' operations to pre-train the model with short contexts before RL-V training, significantly reducing token consumption during inference and accelerating subsequent long-chain thinking (CoT) reinforcement learning.
ThinkMorph: Emergent Properties in Multimodal Interleaved Chain-of-Thought Reasoning
Jiawei Gu (National University of Singapore), Yu Cheng (Chinese University of Hong Kong)
TransformerSupervised Fine-TuningVision-Language-Action ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose ThinkMorph, a unified model that enables complementary interactive chained thinking between text and images, fine-tuned on approximately 24K high-quality interactive reasoning trajectories, significantly enhancing performance on vision-centric tasks and demonstrating multiple emergent capabilities.
ThinkOmni: Lifting Textual Reasoning to Omni-modal Scenarios via Guidance Decoding
Yiran Guan (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)
Representation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes a training-agnostic framework called ThinkOmni, which enhances multimodal reasoning capabilities by using a large reasoning model (LRM) to guide the decoding of a multimodal large language model (OLLM) during inference.
ThinKV: Thought-Adaptive KV Cache Compression for Efficient Reasoning Models
Akshat Ramachandran (Georgia Institute of Technology), Tushar Krishna (Georgia Institute of Technology)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposes ThinKV, a thought-adaptive KV Cache Compression Framework that significantly reduces GPU memory usage and enhances inference throughput during long output reasoning processes.
Thompson Sampling via Fine-Tuning of LLMs
Nicolas Menet (IBM Research), Abbas Rahimi (IBM Research)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataPhysics Related
🎯 What it does: An scalable Thompson sampling algorithm called TOSFIT is proposed by fine-tuning a large language model to generate a strategy that maximizes the probability of reward, for Bayesian optimization in large unstructured discrete spaces.
THOR: Tool-Integrated Hierarchical Optimization via RL for Mathematical Reasoning
Qikai Chang, Jianqing Gao (iFLYTEK Research)
OptimizationAI Code AssistantLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Propose the THOR framework, which significantly improves the accuracy of large language models in mathematical reasoning and code generation tasks by leveraging tool integration and hierarchical reinforcement learning.
Thought Branches: Interpreting LLM Reasoning Requires Resampling
Uzay Macar (Mats), Neel Nanda
Explainability and InterpretabilityLarge Language ModelTextChain-of-Thought
🎯 What it does: By performing multiple resampling of chain-of-thought (CoT) generated by large language models, the study investigates the causal impact of each step on the final decision and proposes resilience and adversarial importance metrics.
Threading Keyframe with Narratives: MLLMs as Strong Long Video Comprehenders
Bo Fang (City University of Hong Kong), Antoni B. Chan (City University of Hong Kong)
RecognitionOptimizationRepresentation LearningLarge Language ModelVision Language ModelVideoTextBenchmark
🎯 What it does: Proposes an unsupervised long video understanding framework called Nar-KFC, consisting of two stages: Key Frame Capture (KFC) and Narrative Key Frame Interpolation (Nar-KFC);
Three Forward, One Backward: Memory-Efficient Full-Rank Fine-Tuning of Large Models via Extra Forward Passes
Jia Zhang (Jilin University), Bin Gu (Jilin University)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose LMAO, an alternating optimization framework that combines LoRA and MeZO, achieving full-rank updates through a three-forward-one-backward step, enabling memory-efficient fine-tuning of large models.
Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs
Zhiyu Pan (Huazhong University of Science and Technology), Jieping Ye (Alibaba Cloud)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose the VC-STaR framework, which leverages visual contrast VQA to correct hallucinations in vision-language models, and generates a high-quality visual reasoning dataset called VisCoR55K for self-improvement of VLMs.
Thyme: Think Beyond Images
YiFan Zhang, Guorui Zhou (Kwai Keye)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision-Language-Action ModelMultimodality
🎯 What it does: Propose a new multimodal large language model paradigm called Thyme, enabling the model to autonomously generate and execute code to accomplish various image processing and complex computational tasks.
TianQuan-S2S: A Subseasonal-to-Seasonal Global Weather Model via Incorporate Climatology State
Guowen Li (Sun Yat-Sen University), Haohuan Fu (Huawei Technologies Co., Ltd)
Convolutional Neural NetworkTransformerTime SeriesSequentialPhysics Related
🎯 What it does: Proposed and implemented the TianQuan-S2S model for global subseasonal-to-seasonal weather forecasting spanning 15–45 days, integrating initial meteorological states with 38-year average climate data and injecting uncertainty noise into the Transformer architecture.
TIGaussian: Disentangle Gaussians for Spatial-Awared Text-Image-3D Alignment
Jiarun Liu (Alibaba Group), Sheng Yang (Alibaba Group)
ClassificationRetrievalRepresentation LearningTransformerVision Language ModelDiffusion modelContrastive LearningGaussian SplattingImageTextPoint Cloud
🎯 What it does: This paper proposes the TIGAUSSIAN framework, achieving alignment and pretraining across text-image-3D Gaussian (3DGS) three modalities.
Tight Bounds for Schrodinger Potential Estimation in Unpaired Data Translation
Nikita Puchkin (HSE University), Denis Belomestny (Duisburg-Essen University)
Image TranslationImageBiomedical DataStochastic Differential Equation
🎯 What it does: This paper studies estimating the potential energy of a Schrodinger bridge using an Ornstein-Uhlenbeck reference process, given only i.i.d. samples from the initial and terminal distributions, and provides an upper bound on the generalization error of the empirical risk minimizer;
Tighter Performance Theory of FedExProx
Wojciech Anyszka (Princeton University), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationFederated Learning
🎯 What it does: This paper conducts an in-depth theoretical analysis of FedExProx (a distributed proximal algorithm that utilizes extrapolation), proposing a more compact proof of linear convergence rate, and providing iteration and time complexity results for non-strongly convex quadratic optimization and general functions that satisfy the Polyak-Łojasiewicz condition.
TikZilla: Scaling Text-to-TikZ with High-Quality Data and Reinforcement Learning
Christian Greisinger (University of Technology Nuremberg), Steffen Eger (University of Technology Nuremberg)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelText
🎯 What it does: Constructed a larger-scale, higher-quality DaTikZ-V4 dataset and trained a small Qwen model TikZilla to achieve high-quality generation of TikZ code from text
TileLang: Bridge Programmability and Performance in Modern Neural Kernels
Lei Wang (Peking University), Zhi Yang (Peking University)
OptimizationComputational Efficiency
🎯 What it does: Propose TILELANG, a programmable tile-level system that provides explicit memory placement, data movement, and parallel scheduling primitives, and achieves automated tile recommendation and tile inference through a unified fused tile-level dataflow graph (FTG), significantly reducing the complexity of AI kernel development and improving performance.
Time Is a Feature: Exploiting Temporal Dynamics in Diffusion Language Models
Wen Wang (Zhejiang University), Chunhua Shen (Zhejiang University)
Large Language ModelReinforcement LearningDiffusion modelTextBenchmark
🎯 What it does: Studied the temporal dynamics of diffusion large language models, proposing two methods, time self-consistent voting and time consistency reinforcement, to enhance reasoning accuracy.
Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks
Yi Yu (Nanyang Technological University), Xudong Jiang (Nanyang Technological University)
OptimizationAdversarial AttackSpiking Neural NetworkTime Series
🎯 What it does: Proposed an attack method that only retimes the timing of neural spikes, keeping the spike count and amplitude unchanged;
Time Optimal Execution of Action Chunk Policies Beyond Demonstration Speed
Sunwoo Kim (Seoul National University), Joseph J Lim
Robotic IntelligenceVision-Language-Action ModelDiffusion model
🎯 What it does: Developed and verified a method called RACE, which can increase the execution speed of imitation learning policies beyond the demonstration speed while maintaining a high success rate.
Time-Gated Multi-Scale Flow Matching for Time-Series Imputation
Hangtian Wang (National Institute of Informatics), Mahito Sugiyama (National Institute of Informatics)
RestorationTransformerFlow-based ModelTime Series
🎯 What it does: Propose a deterministic time series missing value imputation method based on flow matching—Time-Gated Multi-Scale Flow Matching (TG-MSFM).
Time-To-Inconsistency: A Survival Analysis of Large Language Model Robustness to Adversarial Attacks
Yubo Li (Carnegie Mellon University), Rema Padman (Carnegie Mellon University)
Adversarial AttackLarge Language ModelTextSequentialBenchmark
🎯 What it does: Treat the first occurrence of inconsistency in multi-turn dialogues as a time-to-event problem, conducting a large-scale evaluation on 36,951 rounds of MT-Consistency dialogues using survival analysis methods.
Time-to-Move: Training-Free Motion-Controlled Video Generation via Dual-Clock Denoising
Assaf Singer (Technion Israel Institute of Technology), Or Litany (Technion Israel Institute of Technology)
GenerationDiffusion modelScore-based ModelVideoBenchmarkStochastic Differential Equation
🎯 What it does: Proposes a no-training, plug-and-play framework that leverages user-provided rough animations and image conditions to control motion and appearance in video generation.
TimeOmni-1: Incentivizing Complex Reasoning with Time Series in Large Language Models
Tong Guan (Griffith University), Shirui Pan (Griffith University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTime Series
🎯 What it does: Proposed the TSR-SUITE dataset and the TIMEOMNI-1 model, focusing on three types of reasoning tasks for time series: perception, extrapolation, and decision-making, filling the gap of lacking high-quality reasoning data and a general reasoning framework.
TimeRecipe: A Time-Series Forecasting Recipe via Benchmarking Module Level Effectiveness
Zhiyuan Zhao (Georgia Institute of Technology), B. Aditya Prakash (Georgia Institute of Technology)
Recurrent Neural NetworkTransformerTime SeriesBenchmark
🎯 What it does: Propose the TIMERECIPE framework to conduct a unified benchmark evaluation at the module level of time series forecasting models, systematically analyzing the effectiveness of components such as preprocessing, embedding, and feedforward modules;
TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning
Junwen Pan (ByteDance), Qi She (ByteDance)
RetrievalReinforcement LearningVideoBenchmark
🎯 What it does: This paper proposes the TimeSearch-R framework, which transforms video temporal search into text-video interactive reasoning and learns optimal search strategies through reinforcement learning.
TimeSeg: An Information-Theoretic Segment-Wise Explainer for Time-Series Predictions
Hwijin Kim (Korea University), Changhee Lee (Korea University)
Explainability and InterpretabilityReinforcement LearningTime SeriesElectrocardiogram
🎯 What it does: Propose TimeSeg, an information-theoretic segmenter that automatically selects a set of continuous, variable-length subsequences as explanations for time series prediction models under strict black-box settings.
TimeSeriesExamAgent: Creating Time Series Reasoning Benchmarks at Scale
Malgorzata Gwiazda (Technical University of Munich), Artur Dubrawski (Carnegie Mellon University)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningAgentic AITime SeriesBiomedical DataBenchmarkFinance Related
🎯 What it does: Proposed two frameworks, TimeSeriesExam and TimeSeriesExamAgent, for constructing controllable synthetic time series evaluation question banks and automatically generating domain-specific evaluation questions based on real data.
TIMESLIVER : SYMBOLIC-LINEAR DECOMPOSITION FOR EXPLAINABLE TIME SERIES CLASSIFICATION
Akash Pandey (Northwestern University), Sinan Keten (Northwestern University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkTime SeriesBiomedical Data
🎯 What it does: Propose TimeSliver, an interpretable deep learning framework that linearly combines original time series with symbolic discretization results;
Tina: Tiny Reasoning Models via LoRA
Shangshang Wang (University of Southern California), Willie Neiswanger (University of Southern California)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Post-training on the 1.5B Tiny model by combining LoRA and reinforcement learning to build the Tina model, aiming to enhance multi-step reasoning capabilities.
TINKER: Diffusion's Gift to 3D--Multi-View Consistent Editing From Sparse Inputs without Per-Scene Optimization
Canyu Zhao (Zhejiang University), Chunhua Shen (Zhejiang University)
GenerationTransformerSupervised Fine-TuningDiffusion modelFlow-based ModelContrastive LearningImage
🎯 What it does: Propose the TINKER framework, which can achieve high-fidelity 3D editing without scene optimization under the condition of having only one or two edited images.
TINY BUT MIGHTY: A SOFTWARE-HARDWARE CO- DESIGN APPROACH FOR EFFICIENT MULTIMODAL IN- FERENCE ON BATTERY-POWERED SMALL DEVICES
Yilong Li (University of Wisconsin Madison), Suman Banerjee (Uber)
Computational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: Proposed NANOMIND, a hardware-software co-designed framework, achieving efficient offline multimodal inference by decomposing large multimodal models into independently operable modules (visual encoding, projection, language decoding, etc.) and dynamically assigning them to the most suitable accelerators (NPU, GPU, CPU) on the SoC based on their respective computational characteristics;
TIPO: Text to Image with Text Pre-sampling for Prompt Optimization
Shih-Ying Yeh (Karolinska Institute), Shang-Hong Lai (National Tsing Hua University)
GenerationTransformerLarge Language ModelPrompt EngineeringImageText
🎯 What it does: Proposed the TIPO framework, which enhances user prompts during the text generation phase by pre-sampling them with a lightweight language model, making them more aligned with the text distribution used in training text-to-image models, thereby improving image quality and semantic consistency.
TIPS: Turn-level Information-Potential Reward Shaping for Search-Augmented LLMs
Yutao Xie (UC San Diego), Xiaolong Wang (UC San Diego)
RetrievalReinforcement LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: For search-augmented large language models (LLMs) in open-domain question answering tasks, reinforcement learning (RL) is employed to achieve more stable training and improve credit assignment for multi-step tool calls.
TiTok: Transfer Token-level Knowledge via Contrastive Excess to Transplant LoRA
ChanJoo Jung, Jaehyung Kim (Yonsei University)
Data SynthesisKnowledge DistillationRepresentation LearningTransformerContrastive LearningTextBenchmark
🎯 What it does: Proposed the TITOK framework, leveraging the token-level contrastive advantages of the source model LoRA to achieve LoRA knowledge transfer;
TNT: Improving Chunkwise Training for Test-Time Memorization
Zeman Li (University of Southern California), Vahab Mirrokni (Google Research)
Computational EfficiencyRepresentation LearningRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a two-stage training framework TNT, specifically designed for deep memory modules, which first efficiently pre-trains on large block sizes and then fine-tunes on small block sizes, decoupling training efficiency from inference performance.
To Augment or Not to Augment? Diagnosing Distributional Symmetry Breaking
Hannah Lawrence (Massachusetts Institute Of Technology), Robin Walters (Massachusetts Institute Of Technology)
Explainability and InterpretabilityData-Centric LearningImageTextPoint CloudGraph
🎯 What it does: This study investigates the impact of distributed symmetry breaking on equivariant methods and proposes a metric m_{p_X} based on a two-sample classifier to quantify the symmetry bias in datasets.
To Compress or Not? Pushing the Frontier of Lossless GenAI Model Weights Compression with Exponent Concentration
Zeyu Yang (Rice University), Anshumali Shrivastava (Rice University)
CompressionTransformer
🎯 What it does: Analyze and prove the exponential concentration phenomenon in the weights of generative AI models, and based on this, design ECF8, an FP8 format that achieves lossless compression.
To Infinity and Beyond: Tool-Use Unlocks Length Generalization in State Space Models
Eran Malach (Apple), Etai Littwin (Apple)
Recurrent Neural NetworkTransformerSupervised Fine-TuningAgentic AITextChain-of-Thought
🎯 What it does: Investigate the capabilities and limitations of State Space Models (SSM) in long text generation, demonstrating that SSM cannot generate long tables without tools or single-round tool usage, and propose achieving length generalization through interactive tool usage, with theoretical and experimental validation on arithmetic, reasoning, and programming tasks.
To Sink or Not to Sink: Visual Information Pathways in Large Vision-Language Models
Jiayun Luo (University of British Columbia), Leonid Sigal (University of British Columbia)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: Studied visual attention sinks in large vision-language models, analyzed their impact on inference, and proposed a training-free sink-to-the-front reordering method and the DIYSink framework to better leverage sinks for enhancing visual reasoning performance.
To View Transform or Not to View Transform: NeRF-based Pre-training Perspective
Hyeonjun Jeong (KAIST), Dongsuk Kum (KAIST)
Autonomous DrivingNeural Radiance FieldImagePoint Cloud
🎯 What it does: Designed NeRP3D, a continuous point cloud architecture based on NeRF, enabling simultaneous scene reconstruction and multi-task perception without discarding pre-trained models.
Token Alignment Heads: Unveiling Attention's Role in LLM Multilingual Translation
BINBINLIU, Yin Zheng (ByteDance)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Studied and identified 'token alignment heads' in large language models, which are attention heads responsible for mapping source words to target words, and verified their causal impact on translation ability through ablation experiments.
Token Distillation: Attention-Aware Input Embeddings for New Tokens
Konstantin Dobler (Hasso Plattner Institute), Gerard de Melo (Hasso Plattner Institute)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelTextBiomedical DataBenchmark
🎯 What it does: Propose the Token Distillation method for quickly initializing input embeddings of new tokens in pre-trained models.
Token Hidden Reward: Steering Exploration-Exploitation in Group Relative Deep Reinforcement Learning
Wenlong Deng (University of British Columbia), Christos Thrampoulidis (University of British Columbia)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose the Token Hidden Reward (THR) metric, which quantifies the impact of each token on the confidence of the correct answer, and explicitly regulate the balance between exploration and exploitation in Group Relative Policy Optimization (GRPO) through THR-guided reweighting.
Token-Based Audio Inpainting via Discrete Diffusion
Tali Dror (Ben-Gurion University of the Negev), Eliya Nachmani (Ben-Gurion University of the Negev)
RestorationTransformerDiffusion modelAudio
🎯 What it does: Proposed a discrete diffusion-based audio restoration framework called AIDD, which achieves natural reconstruction of long missing audio segments by quantizing audio into discrete tokens and performing the diffusion reverse process in the token space.
Token-Efficient Item Representation via Images for LLM Recommender Systems
Kibum Kim (KAIST), Chanyoung Park (KAIST)
Recommendation SystemLarge Language ModelPrompt EngineeringVision Language ModelImageTextRetrieval-Augmented Generation
🎯 What it does: This work proposes an efficient product representation method based on images (I-LLMRec), which uses a small number of tokens instead of lengthy textual descriptions, enabling large language models (LLMs) to efficiently and comprehensively capture product semantics in recommendation tasks;
Token-Efficient Long-Term Interest Sketching and Internalized Reasoning for LLM-based Recommendation
Zhihao Ding (Hong Kong Polytechnic University), Jieming Shi (Hong Kong Polytechnic University)
Recommendation SystemLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the SIREN framework, which uses large language models to predict user ratings, and addresses issues of long-term historical noise and inference latency by constructing interest sketches and internalizing reasoning.
Token-Guard: Towards Token-Level Hallucination Control via Self-Checking Decoding
Yifan Zhu (Beijing University of Posts and Telecommunications), Haoran Luo (Nanyang Technological University)
GenerationLarge Language ModelTextBiomedical DataBenchmarkFinance Related
🎯 What it does: Designed and implemented Token-Guard, a token-level hallucination control decoding framework that performs step-by-step self-inspection, segmented evaluation, and local/global iterative correction during the generation process.
Token-Importance Guided Direct Preference Optimization
Ning Yang (Institute of Automation Chinese Academy of Sciences), Haijun Zhang (University of Science and Technology Beijing)
OptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposes the Token-Importance Guided Direct Preference Optimization (TI-DPO) framework to achieve fine-grained token-level optimization during large language model (LLM) alignment, combining hybrid weighting and triplet loss.
Token-level Data Selection for Safe LLM Fine-tuning
Yanping Li (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
Safty and PrivacyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This work proposes a token-level data filtering framework called TOSS, which assesses the safety risk of each token using the loss difference between a safety degradation model and a utility model, and achieves precise removal of hazardous tokens through global ranking, thereby enhancing both safety and utility during LLM fine-tuning; simultaneously, it introduces an iteratively improved version, TOSS-Pro, to further strengthen the identification capability of the safety degradation model.
Tokenisation over Bounded Alphabets is Hard
Violeta Kastreva (Sofia University St Kliment Ohridski), Tiago Pimentel (ETH Zürich)
🎯 What it does: This paper studies the segmentation problem under restricted alphabets (binary or single symbol), proving the NP-completeness and APX-hardness of binary direct/bottom-up segmentation, and that single-symbol direct segmentation is strongly NP-complete.
Tokenization to Transfer: Do Genomic Foundation Models Learn Good Representations?
Kirill Vishniakov (Ruya AI), Shadab Khan (M42)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningBiomedical DataBenchmark
🎯 What it does: Evaluated the performance of seven genomic foundation models on 52 downstream tasks and compared them with randomly initialized models.
Tokenizing Single-Channel EEG with Time-Frequency Motif Learning
Jathurshan Pradeepkumar (University of Illinois Urbana Champaign), Jimeng Sun (University of Illinois Urbana Champaign)
Representation LearningTransformerBiomedical Data
🎯 What it does: Proposed TFM-Tokenizer, a tokenizer that converts single-channel EEG into discrete time-frequency pattern tokens.
TokenSeek: Memory Efficient Fine Tuning via Instance-Aware Token Ditching
Runjia Zeng (Rochester Institute of Technology), Dongfang Liu (Rochester Institute of Technology)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a plugin called TOKENSEEK, which achieves efficient fine-tuning of LLMs through instance-aware token addressing and dropping
TokMem: One-Token Procedural Memory for Large Language Models
Zijun Wu (University of Alberta), Lili Mou (University of Alberta)
TransformerLarge Language ModelText
🎯 What it does: Propose the TokMem framework, which compresses reusable task programs into a single trainable memory token, maintaining the LLM frozen to achieve procedural memory and composable execution without context length overhead;
TokUR: Token-Level Uncertainty Estimation for Large Language Model Reasoning
Tunyu Zhang (Rutgers University), Hao Wang (Rutgers University)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes a training-agnostic token-level uncertainty estimation framework called TokUR, which generates prediction distributions for each token by applying low-rank random perturbations to attention layer weights, thereby evaluating self-assessment and self-improvement of large language models during multi-step reasoning.
ToolACE-MT: Non-Autoregressive Generation for Agentic Multi-Turn Interaction
Xingshan Zeng (Huawei Technologies Co Ltd), Qun Liu (Huawei Technologies Co Ltd)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
🎯 What it does: Propose the ToolACE-MT non-autoregressive iterative generation framework for constructing high-quality multi-turn tool call dialogue data, consisting of three stages: initialization, iterative refinement, and offline validation;
Tools are under-documented: Simple Document Expansion Boosts Tool Retrieval
Xuan Lu (Shanghai Jiao Tong University), Xiaoyu Shen (Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Propose the TOOL-REX benchmark and build a low-cost LLM document expansion pipeline, enriching tool documentation with structured fields such as functional description, use cases, limitations, and tags; train specialized tool retrieval models, Tool-Embed (dense retriever) and Tool-Rank (LLM reranker), based on the expanded data.
ToolTree: Efficient LLM Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning
Shuo Yang (University of Melbourne), Eduard Hovy (University of Melbourne)
OptimizationComputational EfficiencyTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper proposes ToolTree, a tool planning framework based on Monte Carlo Tree Search (MCTS), which helps LLM agents sequentially and efficiently invoke external tools in multi-step tasks;
ToolWeaver: Weaving Collaborative Semantics for Scalable Tool Use in Large Language Models
Bowen Fang (Chinese Academy of Sciences), Liang Wang (Chinese Academy of Sciences)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the ToolWeaver framework, which uses hierarchical tool code representations and collaboration-aware vector quantization to learn tool semantics and collaboration relationships. These codes are injected into LLMs through generation-aligned fine-tuning, achieving end-to-end generation for tool selection and invocation.
ToonComposer: Streamlining Cartoon Production with Generative Post-Keyframing
Lingen Li (Chinese University of Hong Kong), Ying Shan (Tencent PCG)
GenerationTransformerDiffusion modelRectified FlowImageVideo
🎯 What it does: Propose ToonComposer, unifying the keyframe interpolation and coloring processes in animation production, forming the post-keyframe stage.
TopoFormer: Topology Meets Attention for Graph Learning
Md Joshem Uddin (University of Texas at Dallas), Baris Coskunuzer (University of Texas at Dallas)
Representation LearningDrug DiscoveryTransformerContrastive LearningGraph
🎯 What it does: Developed the TOPOFORMER framework, converting the topological structure of graphs into sequences processable by Transformers, thereby enabling graph-level learning.
Topological Anomaly Quantification for Semi-supervised Graph Anomaly Detection
Ting Guo, Da Wang (Shanxi University)
Anomaly DetectionGraph Neural NetworkAuto EncoderGraph
🎯 What it does: Propose the TAQ-GAD framework, utilizing topological anomaly quantization (NBS, PIS) to dynamically screen pseudo-anomalous nodes and enhance the graph structure through a virtual anomaly center, achieving semi-supervised graph anomaly detection.
Topological Causal Effects
Kwangho Kim (Korea University), Hajin Lee (Korea University)
ImageGraphBiomedical DataComputed Tomography
🎯 What it does: Proposes a causal inference framework based on topological data analysis, defines the topological average treatment effect (TATE), and provides a nonparametric doubly robust estimator along with corresponding inference methods.
Topological Flow Matching
Kacper Wyrwal (University of Oxford), Alexander Tong (AITHYRA)
GenerationData SynthesisFlow-based ModelImageGraphBiomedical DataMagnetic Resonance Imaging
🎯 What it does: In generative models, a flow matching framework is used to model signals on structured spaces (such as graphs and simplicial complexes), and Topological Flow Matching (TFM) is proposed by introducing a drift induced by the Laplacian operator in the reference process to capture topological information.
Topology and geometry of the learning space of ReLU networks: connectivity and singularities
Marco Nurisso (Politecnico di Torino), Francesco Vaccarino (Politecnico di Torino)
OptimizationComputational EfficiencyTabularBiomedical DataOrdinary Differential Equation
🎯 What it does: This paper investigates the gradient flow dynamics of ReLU networks under arbitrary DAG structures, proposes conservation laws generated by rescaling symmetry, and defines and analyzes invariant sets in the parameter space based on this. It provides a complete geometric and topological characterization of their connectivity, bottleneck nodes, and singularities, further inducing singularities through nuclear norm regularization to achieve structured pruning.
Topology Matters in RTL Circuit Representation Learning
Mingyu Zhao (Beijing University of Posts and Telecommunications), Chuan Shi (Beijing University of Posts and Telecommunications)
Representation LearningAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelContrastive LearningTextMultimodalityGraph
🎯 What it does: Proposed and implemented the TopoRTL framework for RTL circuit representation learning based on behavioral and topological information, validated on PPA prediction and natural language code retrieval tasks.
Topology of Reasoning: Retrieved Cell Complex-Augmented Generation for Textual Graph Question Answering
Sen Zhao (Chongqing University of Posts and Telecommunications), Ding Zou (Chongqing University of Posts and Telecommunications)
GenerationRetrievalTransformerSupervised Fine-TuningPrompt EngineeringGraphRetrieval-Augmented Generation
🎯 What it does: Proposes the TopoRAG framework, which elevates text graphs to cellular complexes, employs topology-based subcomplex retrieval and multi-dimensional message passing, and finally injects the retrieval results into large language models for question answering.
Topology-Preserved Auto-regressive Mesh Generation in the Manner of Weaving Silk
Gaochao Song (University of Hong Kong), Shenghua Gao (University of Hong Kong)
GenerationData SynthesisCompressionTransformerMesh
🎯 What it does: Proposes a grid tokenization algorithm based on vertex hierarchical sorting, which can strictly preserve geometric properties such as manifoldness, closure, normal consistency, and partial identifiability during autoregressive generation, while achieving efficient lossless compression;
ToProVAR: Efficient Visual Autoregressive Modeling via Tri-Dimensional Entropy-Aware Semantic Analysis and Sparsity Optimization
Jiayu Chen (Peking University), Xiang Chen (Peking University)
Computational EfficiencyTransformerImage
🎯 What it does: Propose the ToProVAR framework, which optimizes the three-dimensional sparsity of visual autoregressive models through attention entropy;
TOUCH: Text-guided Controllable Generation of Free-Form Hand-Object Interactions
Guangyi Han (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
GenerationData SynthesisRobotic IntelligenceConvolutional Neural NetworkTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelAuto EncoderTextPoint Cloud
🎯 What it does: Proposed the Free-Form Hand-Object Interaction (HOI) generation task, breaking the traditional grasp-centric limitations, capable of generating diverse, controllable, and physically plausible hand-object interaction poses based on fine-grained text intentions.
Toward Complex-Valued Neural Networks for Waveform Generation
Hyung-Seok Oh (Korea University), Seong-Whan Lee (Korea University)
GenerationGenerative Adversarial NetworkAudio
🎯 What it does: Proposed a fully complex network-based iSTFT neural vocoder named ComVo, achieving waveform generation through a GAN framework.