ICLR 2026 Papers — Page 50
International Conference on Learning Representations · 5356 papers
Transformers Learn Latent Mixture Models In-Context via Mirror Descent
Francesco D'Angelo (EPFL), Nicolas Flammarion (EPFL)
OptimizationRepresentation LearningTransformerSequential
🎯 What it does: Investigate how transformers infer latent mixture weights in the context during sequence modeling, explaining the learning process of their attention mechanism.
Transformers Trained via Gradient Descent Can Provably Learn a Class of Teacher Models
Chenyang Zhang (University of Hong Kong), Yuan Cao (University of Hong Kong)
Knowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: The paper studies using a Transformer as the student model, learning a teacher model through gradient descent, which includes convolutional layers and average pooling, graph convolution, sparse token selection, group sparse linear prediction, etc.
Transformers with Endogenous In-Context Learning: Bias Characterization and Mitigation
Haotian Wang (National University of Defense Technology), Zhouchen Lin (National University of Defense Technology)
Explainability and InterpretabilityData-Centric LearningTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper first theoretically analyzes the impact of hidden confounders on Transformer pre-training and In-Context Learning (ICL) prediction, revealing that prediction bias is proportional to confounding intensity; subsequently, it proposes a double debiasing method (DDbias) that requires no gradient updates and only uses a few unconfounded prompt samples, achieving debiasing through dual prompting on pre-trained models.
Transitive RL: Value Learning via Divide and Conquer
Seohong Park (University of California, Berkeley), Sergey Levine (University of California, Berkeley)
Reinforcement LearningSequentialBenchmark
🎯 What it does: Propose Transitive RL (TRL), an offline goal-conditioned reinforcement learning (GCRL) value learning algorithm based on the divide-and-conquer paradigm.
Translate Policy to Language: Flow Matching Generated Rewards for LLM Explanations
Xinyi Yang, Tonghan Wang (Tsinghua University)
Explainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningFlow-based ModelTextBenchmark
🎯 What it does: By generating explanations without decision prompts using an interpretive LLM based on context, and reconstructing human reward distributions from noisy rewards produced by a proxy LLM using continuous normalized flows (CNF), sentence-level, distributed rewards for explanation generation are created within the PPO framework.
Translating Flow to Policy via Hindsight Online Imitation
Yitian Zheng (Institute for Interdisciplinary Information Sciences Tsinghua University), Yang Gao (Institute for Interdisciplinary Information Sciences Tsinghua University)
Robotic IntelligenceTransformerVideo
🎯 What it does: Proposed the HinFlow method, which combines high-level point flow planning with online self-imitation to train low-level executable policies.
TraPO: A Semi-Supervised Reinforcement Learning Framework for Boosting LLM Reasoning
Shenzhi Yang (Zhejiang University), Gang Chen (Zhejiang University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a semi-supervised reinforcement learning framework TRAPO, which utilizes a small number of labeled samples to guide RLVR training on large-scale unlabeled data, and dynamically selects reliable unlabeled samples through trajectory similarity.
Trapped by simplicity: When Transformers fail to learn from noisy features
Evan Peters (University of Waterloo), Achim Kempf (University of Waterloo)
ClassificationRecurrent Neural NetworkTransformerTabular
🎯 What it does: Studied the noise-robust learning performance of Transformer and LSTM for Boolean functions under feature noise environments, and analyzed the simplicity bias of Transformer.
Tree Search for LLM Agent Reinforcement Learning
Yuxiang Ji (Xiamen University), Liaoni Wu (Southern University of Science and Technology)
Large Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Proposed a multi-step interactive reinforcement learning method called Tree-GRPO based on tree search, which uses a tree structure with complete think-action-observe nodes for rollout;
Tree-based Dialogue Reinforced Policy Optimization for Red-Teaming Attacks
Ruohao Guo (Georgia Institute of Technology), Dan Roth (Oracle AI University of Pennsylvania)
Adversarial AttackSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a dialogue reinforcement learning framework called DIALTREE based on tree search, which automatically discovers multi-round red team attack strategies and attacks large language models.
Tree-sliced Sobolev IPM
Viet-Hoang Tran (National University of Singapore), Tan Minh Nguyen (National University of Singapore)
GenerationRepresentation LearningDiffusion modelScore-based ModelImageText
🎯 What it does: Propose Tree-Sliced Sobolev Integral Probability Metric (TS-Sobolev) and its spherical variant STS-Sobolev as new methods to compute tree-cutting distances for any p≥1;
TreeGrad-Ranker: Feature Ranking via $O(L)$-Time Gradients for Decision Trees
Weida Li, Bryan Kian Hsiang Low (National University of Singapore)
Explainability and InterpretabilityTabular
🎯 What it does: This paper proposes a gradient ascent-based feature ranking method called TreeGrad-Ranker, which directly addresses the joint optimization problem corresponding to insertion/deletion metrics.
TreeGRPO: Tree-Advantage GRPO for Online RL Post-Training of Diffusion Models
Zheng Ding (University of California San Diego), Weirui Ye (Massachusetts Institute of Technology)
GenerationReinforcement LearningDiffusion modelImage
🎯 What it does: Propose TreeGRPO, an RL post-training framework that transforms the denoising process of diffusion models into a tree search, leveraging tree structures to achieve sample reuse, fine-grained step-wise reward allocation, and multiple policy updates within each forward pass.
Triangle Multiplication is All You Need for Biomolecular Structure Representations
Jeffrey Ouyang-Zhang (Genesis Research), Maruan Al-Shedivat (Genesis Research)
Protein Structure PredictionTransformerDiffusion modelBiomedical Data
🎯 What it does: Proposed a simplified biomolecular structure prediction backbone network called Pairmixer, removing Triangle Attention and retaining only Triangle Multiplication and FFN to reduce computational costs;
TRIBE: TRImodal Brain Encoder for whole-brain fMRI response prediction
Stéphane d'Ascoli (Meta AI), Jean-Remi King
Representation LearningTransformerLarge Language ModelMixture of ExpertsVideoTextMultimodalityMagnetic Resonance ImagingAudio
🎯 What it does: TRIBE is a multimodal deep encoding model that leverages pre-trained features from text, audio, and video, combining Transformers and multi-subject conditional layers to predict whole-brain fMRI responses when watching videos.
TriC-Motion: Tri-Domain Causal Modeling Grounded Text-to-Motion Generation
Yiyang Cao (Huazhong University of Science and Technology), Jingdong Chen (Ant Group)
GenerationGraph Neural NetworkTransformerDiffusion modelScore-based ModelVideoTextMultimodality
🎯 What it does: Proposed and implemented a diffusion model framework called TriC-Motion, which utilizes joint modeling across spatial, temporal, and frequency domains, and incorporates causal interventions during training to eliminate noise-related non-motion information, thereby generating high-quality motion sequences highly aligned with text.
Tricks or Traps? A Deep Dive into RL for LLM Reasoning
Zihe Liu (Alibaba Group), Bo Zheng (Alibaba Group)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Systematically reproduce and fine-grained evaluate commonly used reinforcement learning (RL) techniques in LLM inference tasks under a unified open-source framework, ultimately proposing LitePPO, which requires only two core techniques (advantage normalization and token-level loss aggregation), significantly improving the inference accuracy of baseline models.
TRIDENT: Cross-Domain Trajectory Spatio-Temporal Representation via Distance-Preserving Triplet Learning
Guan Yi Jhang, Jiun-Long Huang (National Yang Ming Chiao Tung University)
RetrievalDomain AdaptationRepresentation LearningGraph Neural NetworkContrastive LearningTime SeriesSequential
🎯 What it does: Propose the TRIDENT framework, which uses parameter-free self-supervised triplet learning and distance-preserving multi-kernel triplet loss to unify the encoding and retrieval of continuous and discrete trajectories.
TRIM: Hybrid Inference via Targeted Stepwise Routing in Multi-Step Reasoning Tasks
Vansh Kapoor (Carnegie Mellon University), Aviral Kumar (Amazon)
OptimizationComputational EfficiencyReinforcement LearningMixture of ExpertsText
🎯 What it does: Propose the TRIM framework, which upgrades to larger models only at critical steps in multi-step reasoning tasks, achieving on-demand acceleration of inference;
TrimR: Verifier-based Training-Free Thinking Trimming for Efficient Test-Time Scaling
Weizhe Lin (Huawei), Mingxuan Yuan (Huawei)
Computational EfficiencyTransformerPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose TrimR, a training-agnostic, lightweight-validator-based reasoning trimming framework that dynamically identifies and prematurely terminates redundant Chain-of-Thought (CoT) reasoning processes in large inference models (LRMs), significantly reducing inference time and token consumption;
Trinity: An Evolved LLM Coordinator
Jinglue Xu (Sakana AI), Yujin Tang (Sakana AI)
OptimizationTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose TRINITY, a lightweight coordinator that uses a 0.6B-scale SLM and a 10,000-parameter head to assign three roles—thinking, working, and verifying—to various large language models during multi-round interactions, achieving task collaboration.
Trion: FFT-based Dynamic Subspace Selection for Low-Rank Adaptive Optimization of LLMs
Ionut-Vlad Modoranu (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)
OptimizationComputational EfficiencyTransformerText
🎯 What it does: Propose a dynamic column selection method based on Discrete Cosine Transform (DCT), replacing traditional SVD/QR low-rank projections, and construct two low-rank adaptive optimizers: Trion and DCT-AdamW;
Triple-BERT: Do We Really Need MARL for Order Dispatch on Ride-Sharing Platforms?
Zijian Zhao (Hong Kong University of Science and Technology), Sen Li (Hong Kong University of Science and Technology)
OptimizationTransformerReinforcement LearningTabular
🎯 What it does: Proposed and implemented Triple-BERT, a centralized framework based on single-agent reinforcement learning for large-scale ride-hailing order dispatch.
TripleSumm: Adaptive Triple-Modality Fusion for Video Summarization
Sumin Kim (Seoul National University), Joonseok Lee (Seoul National University)
TransformerVision Language ModelVideoTextMultimodalityAudio
🎯 What it does: Propose the TripleSumm architecture, which can dynamically weight and fuse visual, textual, and audio multimodal information at the frame level for video summarization.
TriQDef: Disrupting Semantic and Gradient Alignment to Prevent Adversarial Patch Transferability in Quantized Neural Networks
Amira Guesmi (New York University Abu Dhabi), Muhammad Shafique (New York University Abu Dhabi)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: To address the cross-bitwidth transferability of adversarial patches in quantized neural networks (QNNs), the TriQDef defense framework is proposed, significantly reducing the success rate of patch attacks across different quantization levels.
TrojanTO: Action-Level Backdoor Attacks Against Trajectory Optimization Models
Yang Dai (National University of Defense Technology), Li Shen (Sun Yat-sen University)
OptimizationAdversarial AttackTransformerSequential
🎯 What it does: This paper proposes a post-training action-level backdoor attack framework called TrojanTO for trajectory optimization (Trajectory Optimization) models, which can implant a powerful trigger on only 0.3% of trajectory samples while maintaining the original task performance;
TROLL: Trust Regions Improve Reinforcement Learning for Large Language Models
Philipp Becker (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a trust region-based gradient update framework called TROLL to replace the traditional PPO-clip;
True Self-Supervised Novel View Synthesis is Transferable
Thomas Mitchel (Adobe), Vincent Sitzmann (MIT)
GenerationData SynthesisTransformerContrastive LearningVideo
🎯 What it does: Proposes a fully geometry-prior-free, purely self-supervised view synthesis framework called XFactor, achieving true transferable view generation
Trust but Verify: Adaptive Conditioning for Reference-Based Diffusion Super-Resolution via Implicit Reference Correlation Modeling
Yuan Wang (Nankai University), Peng-Tao Jiang (vivo BlueImage Lab)
Super ResolutionDiffusion modelGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Designed and implemented Ada-RefSR, a reference-guided super-resolution framework based on single-step diffusion;
Trust The Typical
Debargha Ganguly (Ohio State University, Google Research), Vipin Chaudhary (Ohio State University, Google Research)
Anomaly DetectionSafty and PrivacyLarge Language ModelText
🎯 What it does: Propose the T3 framework, which leverages the 'typicality' principle to convert LLM safety issues into out-of-distribution (OOD) detection, actively identifying atypical prompts and outputs to avoid manually enumerating harmful patterns.
Trust-Region Adaptive Policy Optimization
Mingyu Su (Tsinghua University), Hongning Wang (Tsinghua University)
TransformerSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes the TRAPO framework, which alternates supervised fine-tuning (SFT) and reinforcement learning (RL) within each training instance, guided by expert prefixes to enable model learning and exploration.
TrustGen: A Platform of Dynamic Benchmarking on the Trustworthiness of Generative Foundation Models
Yue Huang, Xiangliang Zhang
GenerationLarge Language ModelVision Language ModelMultimodalityBenchmark
🎯 What it does: Proposed the TRUSTGEN platform for dynamically and scalably evaluating the trustworthiness of multimodal foundation models.
TrustJudge: Inconsistencies of LLM-as-a-Judge and How to Alleviate Them
Yidong Wang (Peking University), Shikun Zhang (Peking University)
Large Language ModelTextBenchmark
🎯 What it does: Systematically analyze inconsistencies in LLM evaluation frameworks and propose the TrustJudge method to address inconsistencies in score comparison and pairwise transitivity.
Truthful or Fabricated? Using Causal Attribution to Mitigate Reward Hacking in Explanations
Pedro Lobato Ferreira (University of Amsterdam), Ivan Titov (University of Edinburgh)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Investigate the reward hacking phenomenon in chain-of-thought (CoT) explanations of reward models, proposing to supplement reward model inputs with interpretive information derived from causal attribution to detect and reduce the probability of generating unfaithful CoT.
Truthfulness Despite Weak Supervision: Evaluating and Training LLMs Using Peer Prediction
Tianyi Alex Qiu (University of California, Berkeley), Cameron Allen (University of California, Berkeley)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposed and verified the use of peer prediction mechanisms for unsupervised evaluation and training of large language models, addressing the deception problem in strong model assessment.
TS-Attn: Temporal-wise Separable Attention for Multi-Event Video Generation
Hongyu Zhang (Nankai University), Daquan Zhou (Nankai University)
GenerationTransformerDiffusion modelVideoTextMultimodality
🎯 What it does: Proposes a training-agnostic temporal separable attention mechanism, TS-Attn, to enhance the accuracy and coherence of multi-event video generation.
TS-DDAE: A Novel Temporal-Spectral Denoising Diffusion AutoEncoder for Wireless Signal Recognition Model Pre-training
Yaoqi Liu (Beijing University of Posts and Telecommunications), Chuan Shi (Beijing University of Posts and Telecommunications)
RecognitionConvolutional Neural NetworkTransformerDiffusion modelAuto EncoderTime SeriesSequential
🎯 What it does: Propose a wireless signal pre-training framework TS-DDAE based on diffusion models, which learns robust features by adding noise and learning denoising simultaneously in both the time domain and spectral domain;
TS$^2$: Training with Sparsemax+, Testing with Softmax for Accurate and Diverse LLM Fine-Tuning
XuZiyang, Yuangang Pan (Centre for Frontier AI Research, Agency for Science, Technology and Research)
Large Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes a method (TS²) for fine-tuning large language models that first uses Sparsemax+ during training and then employs Softmax for inference, significantly improving generation quality and diversity.
TSLM: Tree-Structured Language Modeling for Divergent Thinking
Doyoung Kim (KAIST AI), Minjoon Seo (KAIST AI)
GenerationTransformerSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Propose Tree-Structured Language Modeling (TSLM), which serializes tree structures using special tokens, enabling the language model to generate a complete search tree and perform branch expansion in a single forward pass
TSM-Bench: Detecting LLM-Generated Text in Real-World Wikipedia Editing Practices
Gerrit Quaremba (King's College London), Elena Simperl (Wikimedia Foundation)
ClassificationTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper constructs a multilingual, multi-generator, multi-task benchmark dataset called TSM-BENCH to evaluate the detection performance of machine-generated text in Wikipedia editing scenarios.
TSPulse: Tiny Pre-Trained Models with Disentangled Representations for Rapid Time-Series Analysis
Vijay Ekambaram (IBM), Jayant Kalagnanam (IBM)
ClassificationAnomaly DetectionComputational EfficiencyRepresentation LearningTransformerSupervised Fine-TuningTime Series
🎯 What it does: Proposed a lightweight time series pre-trained model TSPulse with 1M parameters, capable of achieving zero-shot and fine-tuning usage across multiple diagnostic tasks (anomaly detection, classification, interpolation, similarity search).
TTOM: Test-Time Optimization and Memorization for Compositional Video Generation
Leigang Qu (National University Of Singapore), Tat-Seng Chua (National University Of Singapore)
GenerationTransformerLarge Language ModelDiffusion modelVideoText
🎯 What it does: Designed and implemented an unsupervised test-time optimization and memory framework called TTOM, which uses spatiotemporal layouts generated by LLMs to guide video generation during inference, while maintaining an insertable, readable, updatable, and deletable parametric memory to support streaming generation.
TTSDS2: Resources and Benchmark for Evaluating Human-Quality Text to Speech Systems
Christoph Minixhofer (University of Edinburgh), Peter Bell (University of Edinburgh)
GenerationTransformerTextBenchmarkAudio
🎯 What it does: Proposed and made public TTSDS2, a multidimensional evaluation metric based on distribution similarity, and built a sustainable, updatable multilingual TTS benchmark pipeline.
TTT3R: 3D Reconstruction as Test-Time Training
Xingyu Chen (Zhejiang University), Anpei Chen (Westlake University)
GenerationPose EstimationDepth EstimationAutonomous DrivingRecurrent Neural NetworkTransformerPoint CloudSequential
🎯 What it does: Propose TTT3R, which improves the sequence length generalization of 3D reconstruction models through test-time training perspective.
Tucker-FNO: Tensor Tucker-Fourier Neural Operator and its Universal Approximation Theory
Guancheng Zhou (Xi'an Jiaotong University), Deyu Meng (Xi'an Jiaotong University)
RestorationImageVideoPhysics Related
🎯 What it does: Proposed Tucker-FNO, which decomposes high-dimensional Fourier Neural Operator into multiple one-dimensional FNOs via Tucker tensor decomposition, significantly reducing FFT computational complexity while preserving expressive power, and validated on PDE solving and high-dimensional visual signal recovery tasks.
Tug-of-War No More: Harmonizing Accuracy and Robustness in Vision-Language Models via Stability-Aware Task Vector Merging
Junhao Dong (Nanyang Technological University), Yew-Soon Ong (Nanyang Technological University)
ClassificationRecognitionAdversarial AttackVision Language ModelImageTextMultimodality
🎯 What it does: Propose a training-free parameter space task vector merging method called PISTOLE, which uses gradient information and parameter space paths to balance the accuracy and adversarial robustness of vision-language models.
TUMIX: Multi-Agent Test-Time Scaling with Tool-Use Mixture
Yongchao Chen (MIT), Jinsung Yoon (Google Cloud AI Research)
OptimizationComputational EfficiencyTransformerLarge Language ModelAgentic AIMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the TUMIX framework, which utilizes multi-agent parallel execution and iterative sharing to improve answers, achieving test-time extension through interaction with tools (Code Interpreter and Search) and LLM.
TumorChain: Interleaved Multimodal Chain-of-Thought Reasoning for Traceable Clinical Tumor Analysis
Sijing Li (Zhejiang University), Ling Zhang (DAMO Academy, Alibaba Group)
ClassificationSegmentationAnomaly DetectionConvolutional Neural NetworkTransformerPrompt EngineeringVision Language ModelMultimodalityBiomedical DataComputed TomographyBenchmarkChain-of-Thought
🎯 What it does: Constructed the TumorChain framework and the TumorCoT-1.5M dataset, achieving multi-modal chain reasoning from CT images to pathological predictions.
Tuning the burn-in phase in training recurrent neural networks improves their performance
Julian D. Schiller (Leibniz University Hannover), Matthias A. Müller (Leibniz University Hannover)
OptimizationHyperparameter SearchRecurrent Neural NetworkTime Series
🎯 What it does: This paper studies and quantifies the impact of the burn-in phase (the first m steps ignored after network initialization) on training performance when training recurrent neural networks using truncated backpropagation through time (TBPTT), and provides theoretical upper bounds on the performance loss (regret).
Turbo-DDCM: Fast and Flexible Zero-Shot Diffusion-Based Image Compression
Amit Vaisman (Technion Israel Institute of Technology), Tomer Michaeli (Technion Israel Institute of Technology)
CompressionDiffusion modelImage
🎯 What it does: Propose Turbo-DDCM, a zero-shot diffusion image compression method that uses a reusable Gaussian codebook for multi-atom selection.
TurboBoA: Faster and Exact Attention-aware Quantization without Backpropagation
Junhan Kim (Samsung Research), Yongkweon Jeon (Samsung Research)
Computational EfficiencyTransformerText
🎯 What it does: Proposes TurboBoA, a gradient-free post-training quantization algorithm that can quickly quantize large language models to low bit widths while maintaining accuracy.
TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate
Amir Zandieh (Google Research), Vahab Mirrokni (Google Research)
CompressionOptimizationComputational EfficiencyTextTabular
🎯 What it does: Proposes an online vector quantization method called TURBOQUANT, which achieves near-optimal distortion rates at any bit width.
Turning Internal Gap into Self-Improvement: Promoting the Generation-Understanding Unification in MLLMs
Yujin Han (University of Hong Kong), Difan Zou (University of Hong Kong)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 What it does: Systematic evaluation of the internal gap between generation and understanding in unified multi-modal large language models (MLLMs), and the proposal of a self-improvement framework based on internal gaps. The framework uses a stronger understanding branch to guide a weakened generation branch, constructs high-quality image data through post-training (SFT/DPO); during the self-improvement process, co-improvement phenomena of generation and understanding are observed, and further combined with curriculum learning to dynamically expand training samples.
TusoAI: Agentic Optimization for Scientific Methods
Alistair Turcan (Carnegie Mellon University), Martin Jinye Zhang (Carnegie Mellon University)
OptimizationNeural Architecture SearchTransformerLarge Language ModelAgentic AIBiomedical DataBenchmark
🎯 What it does: Developed TusoAI, an agent-based AI system that leverages large language models and structured domain knowledge trees to automate the development and optimization of scientific computing methods.
Tversky Neural Networks: Psychologically Plausible Deep Learning with Differentiable Tversky Similarity
Moussa Koulako Bala Doumbouya (Stanford University), Christopher D Manning (Stanford University)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper proposes a differentiable Tversky similarity and its projection layer, replacing traditional linear projection layers in image recognition and language modeling tasks, demonstrating higher accuracy and fewer parameters.
TVTSyn: Content-Synchronous Time-Varying Timbre for Streaming Voice Conversion and Anonymization
Waris Quamer (Texas A&M University), Ricardo Gutierrez-Osuna (Texas A&M University)
GenerationSafty and PrivacyTransformerAudio
🎯 What it does: This paper proposes a real-time voice conversion and speaker anonymization system called TVTSyn, which utilizes time-synchronized variable timbre (TVT) and global timbre memory to match the time-varying characteristics of content, achieving low latency (<80 ms GPU) and natural synthesis;
TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows
Zhenglin Cheng (Westlake University), Tao Lin (Westlake University)
GenerationData SynthesisSupervised Fine-TuningFlow-based ModelImageTextMultimodality
🎯 What it does: Proposed the TWINFLOW framework, achieving one-step generation for large generative models through self-adversarial dual-track training;
TwinVLA: Data-Efficient Bimanual Manipulation with Twin Single-Arm Vision-Language-Action Models
Hokyun Im (Yonsei University), Youngwoon Lee (Yonsei University)
Computational EfficiencyRobotic IntelligenceMixture of ExpertsVision-Language-Action ModelImageText
🎯 What it does: By replicating a pre-trained single-arm VLA model and achieving bimanual coordination through joint attention, TwinVLA is constructed to enable dual-arm manipulation.
Two (narrow) heads are better than (an arbitrarily wide) one
Amanuel Tesfaye (Northeastern University), Ravi Sundaram (Northeastern University)
OptimizationGraph Neural NetworkTransformerGraph
🎯 What it does: Investigates the representational capabilities of single-head and dual-head Transformers on the Endpoint Selection Problem (ESP), proving that single-head models are infeasible on cyclic graphs, while dual-head models can fully resolve the problem.
Two failure modes of deep transformers and how to avoid them: a unified theory of signal propagation at initialisation
Alessio Giorlandino (International School of Advanced Studies), Sebastian Goldt (International School of Advanced Studies)
TransformerTextPhysics Related
🎯 What it does: Built a comprehensive theory of signal propagation in deep Transformers, unifying the explanation and prevention of rank collapse and entropy collapse
Two-Layer Convolutional Autoencoders Trained on Normal Data Provably Detect Unseen Anomalies
Yanbo Chen (Wuhan University), Weiwei Liu (Wuhan University)
Anomaly DetectionConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: The paper studies the learning dynamics of a 2-layer convolutional autoencoder when trained using only normal data, and theoretically explains its effectiveness in anomaly detection.
Two-Way Is Better Than One: Bidirectional Alignment with Cycle Consistency for Exemplar-Free Class-Incremental Learning
Hongye Xu (Rochester Institute of Technology), Bartosz Krawczyk (Rochester Institute of Technology)
ClassificationKnowledge DistillationRepresentation LearningImageBenchmark
🎯 What it does: Propose a bidirectional cyclic consistency projection (BiCyc) framework to compensate for feature drift in class-incremental learning without sample memory.
Type-Compliant Adaptation Cascades
Chu-Cheng Lin (Google), Eugene Ie (Google)
OptimizationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the Type-Compatible Adaptable Cascading (TACs) framework, viewing large language model workflows as unnormalized probabilistic programs with type constraints, and achieving end-to-end training through gradient optimization.
TyphoonMLA: A Mixed Naive-Absorb MLA Kernel For Shared Prefix
Ahmet Caner Yüzügüler (Huawei), Lukas Cavigelli (Huawei)
Computational EfficiencyText
🎯 What it does: Proposed TyphoonMLA, a hybrid MLA kernel combining naive and absorb implementations, leveraging shared prefixes to enhance attention computation efficiency.
U-MARVEL: Unveiling Key Factors for Universal Multimodal Retrieval via Embedding Learning with MLLMs
Xiaojie Li (Tencent PCG), Xi Chen (Tencent PCG)
RetrievalKnowledge DistillationVision Language ModelContrastive LearningImageVideoTextMultimodalityBenchmark
🎯 What it does: Developed and evaluated a general-purpose retrieval framework called U-MARVEL based on multimodal large language models.
U2-BENCH: Benchmarking Large Vision-Language Models on Ultrasound Understanding
Anjie Le (Dolphin AI), Hongcheng Guo (Dolphin AI)
ClassificationObject DetectionPrompt EngineeringVision Language ModelMultimodalityBiomedical DataUltrasound
🎯 What it does: Designed and released the U2-BENCH benchmark to evaluate the performance of large-scale vision-language models in ultrasound image understanding;
UALM: Unified Audio Language Model for Understanding, Generation and Reasoning
Jinchuan Tian (Carnegie Mellon University), Wei Ping (NVIDIA)
GenerationTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelTextMultimodalityChain-of-ThoughtAudio
🎯 What it does: Developed a unified audio-language model, UALM, integrating audio understanding, text-to-audio generation, and multi-modal reasoning, and further introduced UALM-Reason to achieve cross-modal chain-of-thought reasoning and self-reflection;
UFO-4D: Unposed Feedforward 4D Reconstruction from Two Images
Junhwa Hur (Google), Deqing Sun (Google)
Pose EstimationAutonomous DrivingTransformerNeural Radiance FieldGaussian SplattingOptical FlowImagePoint Cloud
🎯 What it does: Propose UFO-4D, a forward unified 4D reconstruction framework based on two pose-free images, directly outputting dynamic 3D Gaussian Splat and camera poses
UI-Ins: Enhancing GUI Grounding with Multi-Perspective Instruction as Reasoning
Liangyu Chen (Renmin University of China), Steven HOI (Tongyi Lab, Alibaba Group)
TransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextMultimodalityBenchmark
🎯 What it does: Studied instruction diversity and quality in GUI localization tasks, constructed a high-quality, multi-perspective instruction dataset, and proposed the Instruction-as-Reasoning framework, training two-stage models UI-Ins-7B/32B.
UIS-Digger: Towards Comprehensive Research Agent Systems for Real-world Unindexed Information Seeking
Chang Liu (Huawei Technologies Ltd), Lifeng Shang (Huawei Technologies Ltd)
RetrievalLarge Language ModelSupervised Fine-TuningAgentic AITextMultimodalityBenchmark
🎯 What it does: This paper defines the unindexed information retrieval (UIS) task and establishes the UIS-QA benchmark, while constructing a multi-agent framework called UIS-Digger. It achieves automatic answering for UIS tasks by utilizing a dual-mode browser, file reading, and a two-stage fine-tuning strategy (SFT+RFT).
ULD-Net: Enabling Ultra-Low-Degree Fully Polynomial Networks for Homomorphically Encrypted Inference
Xi Xie (University of Connecticut), Caiwen Ding (University of Minnesota - Twin Cities)
Computational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: This work proposes a training method called ULD-Net, which can train fully low-order (multiplication depth ≤3) polynomial networks on ImageNet and Transformer scales, achieving high accuracy while maintaining HE inference efficiency.
ULTRA-360: Unconstrained Dataset for Large-scale Temporal 3D Reconstruction across Altitudes and Omnidirectional Views
Xijun Liu (Johns Hopkins University), Cheng Peng (Johns Hopkins University)
Neural Radiance FieldGaussian SplattingImageVideoBenchmark
🎯 What it does: Propose ULTRA-360—a cross-seasonal, cross-height, 360° panoramic, campus-level large-scale temporal 3D reconstruction dataset containing approximately 37.7k frames—and systematically evaluate feature matching, dense reconstruction, and multi-appearance NVS methods on it, revealing the bottlenecks of existing techniques.
Ultra-Fast Language Generation via Discrete Diffusion Divergence Instruct
Haoyang Zheng (Purdue University), Guang Lin (Purdue University)
GenerationKnowledge DistillationTransformerLarge Language ModelReinforcement LearningDiffusion modelScore-based ModelText
🎯 What it does: This study proposes the DiDi-Instruct framework, which distills a pre-trained diffusion large language model (dLLM) into a student model capable of generating high-quality text with only a few steps, significantly improving generation speed.
UltraGauss: Ultrafast Gaussian Reconstruction of 3D Ultrasound Volumes
Mark C. Eid (University of Oxford), Joao F. Henriques (University of Oxford)
Gaussian SplattingBiomedical DataUltrasound
🎯 What it does: Proposed UltraGauss, an efficient Gaussian splatting framework specifically designed for ultrasound, to rapidly reconstruct 3D volumes from 2D ultrasound scans.
UltraLLaDA: Scaling the Context Length to 128K for Diffusion Large Language Models
Guangxin He (Hong Kong University of Science and Technology), Binhang Yuan (Renmin University of China)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelDiffusion modelText
🎯 What it does: This paper extends the context window of diffusion-based large language models to 128K through lightweight post-training, and proposes diffusion-aware NTK rotation position encoding and cross-document intervention mask strategies;
UltraMemV2: Memory Networks Scaling to 120B Parameters with Superior Long-Context Learning
Zihao Huang (Bytedance), Siyuan Qiao (Bytedance)
Computational EfficiencyTransformerMixture of ExpertsTextBenchmark
🎯 What it does: Design and evaluate UltraMemV2, a novel memory layer architecture aimed at competing with 8-expert MoE models while maintaining low memory access.
UltraViCo: Breaking Extrapolation Limits in Video Diffusion Transformers
Min Zhao (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationTransformerDiffusion modelVideo
🎯 What it does: Propose the UltraViCo plug-and-play method, which suppresses attention beyond the training window to enhance video extrapolation generation quality and eliminate periodic repetitions.
UME-R1: Exploring Reasoning-Driven Generative Multimodal Embeddings
Zhibin Lan (Xiamen University), Jinsong Su (Xiamen University)
Representation LearningTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringImageVideoTextMultimodalityChain-of-Thought
🎯 What it does: Explore the reasoning-driven generative multi-modal embedding framework UME-R1;
Unbiased Gradient Estimation for Event Binning via Functional Backpropagation
Jinze Chen (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
Autonomous DrivingOptimizationComputational EfficiencySimultaneous Localization and MappingOptical Flow
🎯 What it does: Proposed an unbiased gradient estimation framework based on Functional Backpropagation for gradient computation of discrete packing in event vision.
Unbiased Object Detection Beyond Frequency with Visually Prompted Image Synthesis
Xinhao Cai (Nanjing University of Science and Technology), Wenguan Wang (Zhejiang University)
Object DetectionData SynthesisPrompt EngineeringDiffusion modelImage
🎯 What it does: Proposes a generative unbiased object detection framework that dynamically corrects frequency and diversity biases through representative score-driven layout re-calibration and visual blueprint prompt-based L2I synthesis.
Uncertainty as Feature Gaps: Epistemic Uncertainty Quantification of LLMs in Contextual Question-Answering
Yavuz Faruk Bakman (University of Southern California), Sai Praneeth Karimireddy (University of Southern California)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a theoretically grounded confidence measurement method for context-aware question answering tasks. It estimates empirical uncertainty by leveraging the gap between the model's hidden layer and an ideal model, and decomposes and quantifies uncertainty through three semantic features (context dependency, context understanding, and honesty).
Uncertainty Estimation via Hyperspherical Confidence Mapping
Eunseo Choi (KAIST), Heejin Ahn (KAIST)
ClassificationDepth EstimationAnomaly DetectionOptimizationImageTabular
🎯 What it does: Propose a sampling-free and distribution-assumption-free uncertainty estimation framework called Hyperspherical Confidence Mapping (HCM), which decomposes network outputs into magnitude and unit vectors, using vector deviation from the unit sphere as an uncertainty metric.
Uncertainty Matters in Dynamic Gaussian Splatting for Monocular 4D Reconstruction
Fengzhi Guo (Texas A&M University), Cheng Zhang (Texas A&M University)
GenerationGaussian SplattingVideo
🎯 What it does: Proposes a dynamic Gaussian distribution rendering framework called USPLAT4D, which is uncertainty-aware, for 4D reconstruction from monocular videos;
Uncertainty-Aware 3D Reconstruction for Dynamic Underwater Scenes
Rui Liu (Zhejiang University), Wenguan Wang (Zhejiang University)
GenerationNeural Radiance FieldGaussian SplattingOptical FlowVideo
🎯 What it does: Proposed an uncertainty-aware dynamic scene representation (UDF) that jointly models time-varying underwater geometry and participating media, suppressing low-confidence observations via heteroscedastic uncertainty to achieve high-quality 3D reconstruction and novel view synthesis.
Uncertainty-Aware Diagnostics for Physics-Informed Machine Learning
Mara Daniels (Massachusetts Institute of Technology), Michael W. Mahoney (University of California Berkeley)
Explainability and InterpretabilityPhysics Related
🎯 What it does: Propose Physics-Informed Log Evidence (PILE) as a Gaussian process-based, interpretable, single metric for selecting and diagnosing hyperparameters and kernel functions in physics-informed machine learning (PIML).
Uncertainty-Aware Gaussian Map for Vision-Language Navigation
Jianzhe Gao (Zhejiang University), Wenguan Wang (Zhejiang University)
Autonomous DrivingExplainability and InterpretabilityTransformerVision-Language-Action ModelGaussian SplattingTextMultimodalityPoint CloudBenchmark
🎯 What it does: Built a Semantic Gaussian Map based on semantic Gaussian point clouds, explicitly modeling geometric, semantic, and appearance uncertainties, which are finally fused into a unified 3D Value Map to guide Vision-Language Navigation.
Uncertainty-driven Embedding Convolution
Sungjun Lim (Yonsei University), Kyungwoo Song (Yonsei University)
ClassificationRetrievalConvolutional Neural NetworkContrastive LearningText
🎯 What it does: Propose the Uncertainty-Driven Embedding Convolution (UEC) framework, which first converts pre-trained deterministic word vectors into probabilistic vectors using posterior Laplacian approximation, then adaptively aggregates these probabilistic vectors with uncertainty-driven weights, and employs variance-based similarity estimation for downstream tasks;
Uncover Underlying Correspondence for Robust Multi-view Clustering
Haochen Zhou (Sichuan University), Xi Peng (Sichuan University)
Representation LearningContrastive LearningImage
🎯 What it does: This paper proposes a generative framework called CorreGen for multi-view clustering in the presence of noisy correspondences (category-level and sample-level mismatch), automatically discovering potential cross-view correspondences.
Uncovering Conceptual Blindspots in Generative Image Models Using Sparse Autoencoders
Maty Bohacek, Ekdeep Singh Lubana (Harvard University)
GenerationExplainability and InterpretabilityDiffusion modelAuto EncoderContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a systematic method that utilizes a sparse autoencoder (RA-SAE) to extract interpretable concept embeddings and quantifies the difference in concept distribution between generated and real images, thereby identifying 'concept blind spots' in generative models;
Uncovering Semantic Selectivity of Latent Groups in Higher Visual Cortex with Mutual Information-Guided Diffusion
Yule Wang (Georgia Institute of Technology), Anqi Wu (Georgia Institute of Technology)
Diffusion modelAuto EncoderImage
🎯 What it does: Learn neural latent subspaces through a group-level decoupled variational autoencoder, and visualize semantic features of each subspace using a diffusion model guided by mutual information maximization, revealing semantic selectivity in the higher visual cortex.
Understanding and Improving Continuous LLM Adversarial Training via In-context Learning Theory
Shaopeng Fu (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Studied the effective mechanisms of Continuous Adversarial Training (CAT) in large language models (LLMs) against jailbreak attacks, and proposed an improved method called ER-CAT;
Understanding and Improving Hyperbolic Deep Reinforcement Learning
Timo Klein (University of Vienna), Sebastian Tschiatschek (University of Vienna)
OptimizationReinforcement LearningImage
🎯 What it does: Proposed HYPER++, an architecture that stabilizes hyperbolic space deep reinforcement learning through regularization, feature scaling, and classification value loss;
Understanding and Improving Length Generalization in Hierarchical Sparse Attention Models
Jiaqi Leng (Fudan University), Yucheng Lu (NYU Shanghai)
Computational EfficiencyRepresentation LearningTransformerMixture of ExpertsSequentialBenchmark
🎯 What it does: Propose an architecture based on SWA+HSA that enables a model trained on 4K to directly perform inference on long-sequence tasks with 32M length while maintaining high accuracy;
Understanding and improving Shampoo and SOAP via Kullback-Leibler Minimization
Wu Lin (Vector Institute), Roger Baker Grosse
OptimizationLarge Language ModelText
🎯 What it does: Propose interpreting the structured second-moment estimation of Shampoo and SOAP from the perspective of KL divergence, design two novel optimizers, KL-Shampoo and KL-SOAP, leveraging QR decomposition and exponential moving average (EMA) to achieve an efficient training process without Adam step size grafting.
Understanding and Relaxing the Limitations of Transformers for Linear Algebra
Andres Potapczynski (New York University), Andrew Gordon Wilson (New York University)
OptimizationComputational EfficiencyRecurrent Neural NetworkTransformer
🎯 What it does: Investigate the limitations of Transformers in linear algebra tasks and propose an improved method, RangeFormer, enabling scalable and generalizable matrix operations on large-scale (up to 1000×1000) matrices with diverse distributions.
Understanding Cross-layer Contributions to Mixture-of-Experts Routing in LLMs
Wengang Li (Institute of Science Tokyo), Mohamed Wahib (RIKEN)
Explainability and InterpretabilityTransformerMixture of ExpertsText
🎯 What it does: This paper proposes a recursive decomposition framework to quantify the contributions of different model components (such as tokens, attention layers, MoE outputs, and attention heads) to Mixture-of-Experts (MoE) routing decisions, and performs cross-layer interpretability analysis on four mainstream MoE LLMs.
Understanding Dataset Distillation via Spectral Filtering
Deyu Bo (National University of Singapore), Xinchao Wang (National University of Singapore)
Knowledge DistillationImage
🎯 What it does: This paper proposes a spectral filtering framework called UniDD, aiming to unify different dataset distillation (DD) objectives, and reveals that the essence of DD lies in matching frequency-specific features by extracting specific frequency information through analyzing the eigenvalues of the feature-feature correlation (FFC) matrix.
Understanding In-Context Learning on Structured Manifolds: Bridging Attention to Kernel Methods
Zhaiming Shen (Georgia Institute of Technology), Wenjing Liao (Georgia Institute of Technology)
Meta LearningTransformerText
🎯 What it does: The paper studies context learning regression on structural manifolds with transformers, proposing to connect attention mechanisms with kernel methods to construct transformers for kernel regression and providing an upper bound on generalization error.
Understanding Language Prior of LVLMs by Contrasting Chain-of-Embedding
Lin Long (University of Wisconsin-Madison), Sharon Li (University of Wisconsin-Madison)
Explainability and InterpretabilityRepresentation LearningTransformerVision Language ModelContrastive LearningMultimodalityChain-of-Thought
🎯 What it does: By comparing the embedded sequences of visual and non-visual inputs, revealing and quantifying language priors in large vision-language models.
Understanding Sensitivity of Differential Attention through the Lens of Adversarial Robustness
Tsubasa Takahashi (Turing Inc), Kento Sasaki (Turing Inc)
Explainability and InterpretabilityAdversarial AttackTransformerContrastive LearningImage
🎯 What it does: Investigate the structural vulnerability of Differential Attention under adversarial attacks, and demonstrate through theory and experiments that negative gradient alignment leads to sensitivity amplification.