ICML 2026 Papers — Page 57
International Conference on Machine Learning · 6554 papers
SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks?
Xiang Deng (Scale AI), Brad Kenstler (Scale AI)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper introduces the SWE-Bench Pro benchmark, providing nearly two thousand industrial-scale long-term software tasks aimed at evaluating LLM-driven automated coding agents.
SWE-Compass: Towards Unified Evaluation of Agentic Coding Abilities for Large Language Models
Jingxuan Xu (Kuaishou Technology), Jiaheng Liu (Nanjing University)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the SWE-Compass benchmark, generating 2000 executable multi-task, multi-language, and multi-scenario code examples using real GitHub PRs to uniformly evaluate the agent coding capabilities of LLMs.
SWE-fficiency: Can Language Models Optimize Real-World Repositories on Real Workloads?
Jeffrey Jian Ma (Harvard University), Parthasarathy Ranganathan (Google)
OptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes the SWE-FFICIENCY benchmark to evaluate the ability of language models to optimize the performance of open-source software repositories under real-world workloads.
SWE-IF: Aligning Code Evaluation with Human Preference
Ming Zhong (Google DeepMind), Jiao Sun (Google DeepMind)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose VERICODE, a set of 30 verifiable code instruction categories, and the SWE-IF evaluation platform, which conducts dual assessments of LLMs in terms of functionality and instruction following.
SWE-MiniSandbox: Container-Free Reinforcement Learning for Building Software Engineering Agents
Danlong Yuan (Peking University), Dongyan Zhao (Peking University)
AI Code AssistantReinforcement LearningAgentic AIText
🎯 What it does: Proposed and implemented SWE-MiniSandbox, a lightweight sandbox without containers, based on Linux namespace and chroot, used for training software engineering agents in reinforcement learning.
SWE-Perf: Can Language Models Optimize Code Performance on Real-World Repositories?
Xinyi He (Xi'an Jiaotong University), Zejun MA
OptimizationAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Created a benchmark dataset named SWE-Perf to evaluate the ability of large language models to perform code performance optimization in real open-source repositories.
SWE-rebench V2: Language-Agnostic SWE Task Collection at Scale
Ibragim Badertdinov (Nebius), Alexander Golubev (Nebius)
AI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Developed SWE-rebench V2, a language-agnostic automated pipeline capable of collecting executable real-world software engineering tasks from thousands of repositories and 20 programming languages;
Swift-SVD: Theoretical Optimality Meets Practical Efficiency in Low-Rank LLM Compression
Ruoling Qi (Institute of Artificial Intelligence, China Telecom), Qizhen Weng (Dolby Laboratories)
CompressionOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Propose a training-free low-rank LLM compression framework called Swift-SVD, which utilizes activation-aware closed-form solutions and incremental covariance aggregation to achieve single feature decomposition, supporting dynamic hierarchical compression rate allocation.
SWING: Unlocking Implicit Graph Representations for Graph Random Features
Alessandro Manenti (Universit` a della Svizzera italiana), Krzysztof Marcin Choromanski
Computational EfficiencyRepresentation LearningGraph Neural NetworkImagePoint CloudMeshGraph
🎯 What it does: Propose the SWING algorithm, which uses random walking in continuous space instead of traditional random walks on graph nodes, thereby efficiently approximating the computation of graph random feature methods for implicit graphs.
SwitchCraft: A Programmatic Framework for Designing State-Switching Proteins
Bowen Jing (MIT), Bonnie Berger (MIT)
Drug DiscoveryProtein Structure PredictionTransformerSupervised Fine-TuningDiffusion modelContrastive LearningBiomedical DataBenchmark
🎯 What it does: A programmable multi-state protein design framework called SWITCHCRAFT was designed, which uses backpropagation from structural prediction models to achieve constraint optimization for different states, enabling the switching of multiple functional states on the same sequence.
Swordsman: Entropy-Driven Adaptive Block Partition for Efficient Diffusion Language Models
Yu Zhang (Tongji University), Longbing Cao (Macquarie University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Proposes Swordsman, an adaptive block partitioning and dynamic thresholding-based, no-training block decoding framework driven by entropy, aimed at improving the generation speed and quality of diffusion language models.
Sycophancy Towards Researchers Drives Performative Misalignment
David D. Baek (Massachusetts Institute of Technology), Shi Feng (George Washington University)
Federated LearningExplainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringAuto EncoderTextBenchmark
🎯 What it does: The study proposes an alternative explanation for alignment camouflage, introducing the concept of 'performative misalignment,' suggesting that the model's 'aligned behavior' during evaluation is a concession to the researchers' expectations rather than a self-protective deception.
Symbal: Detecting Systematic Misalignments in Model-Generated Captions
Maya Varma (Stanford University), Curtis Langlotz (Stanford University)
Anomaly DetectionRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodalityBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed a new task — systematic misalignment detection, aiming to identify recurring misaligned text errors in image-text pairs caused by visual features.
Symbiosis-Inspired Knowledge Distillation for Incremental Object Detection
Mingyue Zeng (Xidian University), Xinbo Gao (Xidian University)
Object DetectionKnowledge DistillationTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This paper proposes the Symbiosis-Inspired Knowledge Distillation (SIKD) framework for incremental object detection, leveraging spatial and semantic symbiosis information formed by object coexistence and occlusion to retain old knowledge and learn new categories.
Symbol-Equivariant Recurrent Reasoning Models
Richard Freinschlag (Johannes Kepler University Linz), Günter Klambauer
Depth EstimationOptimizationComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackRecurrent Neural NetworkTransformerSupervised Fine-TuningAuto EncoderSequentialChain-of-Thought
🎯 What it does: Propose a Symbol-Equivariant Recursive Reasoning Model (SE-RRM), which explicitly utilizes symbol permutation equivariance in structured reasoning tasks;
SyMerge: From Non-Interference to Synergistic Merging via Single-Layer Adaptation
Aecheon Jung (Sungkyunkwan University), Sungeun Hong (Sungkyunkwan University)
ClassificationSegmentationDepth EstimationDomain AdaptationRecommendation SystemKnowledge DistillationTransformerSupervised Fine-TuningMixture of ExpertsContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes SyMerge, a multi-task model merging framework that achieves collaborative fusion through single-layer adaptation, enabling positive cross-task synergy improvements during the merging process;
Symmetries in language statistics shape the geometry of model representations
Dhruva Karkada (University of California Berkeley), Yasaman Bahri (Google DeepMind)
Representation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This study explores the geometric structure of internal representations in language models, finding that linguistic statistics exhibit translational symmetry, which affects the geometric structure of high-dimensional word embeddings.
Symmetries in PAC-Bayesian Learning
Armin Beck (Max Planck Institute for Informatics), Peter Ochs (Saarland University)
ClassificationRecognitionOptimizationRepresentation LearningConvolutional Neural NetworkGraph Neural NetworkImagePoint Cloud
🎯 What it does: This paper derives new generalization upper bounds under the PAC-Bayes framework, extending the theoretical analysis for non-compact symmetric (such as translation, scaling, Lorentz transformation) and non-translation-invariant data distributions, and provides strict generalization advantages of symmetric models in such scenarios;
Symmetry Reveals Layerwise Dynamics: How Transformers Perform In-Context Classification
Patrick Lutz (Boston University), Venkatesh Saligrama (Boston University)
ClassificationTransformerPrompt EngineeringContrastive LearningTabularBenchmarkChain-of-Thought
🎯 What it does: This paper successfully parses the model's inference process into identifiable hierarchical recursive dynamics by enforcing feature-label permutation symmetry at each layer of the Transformer, and reveals the implementation of a 'coupled feature-label' average drift algorithm;
SymSpectra: Symmetric Information Bottleneck Framework for Molecular Structure Recognition under Imbalanced Settings
Xiaohan Qin (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
RecognitionDrug DiscoveryRecurrent Neural NetworkTransformerContrastive LearningMultimodalityBiomedical DataMagnetic Resonance ImagingReview/Survey Paper
🎯 What it does: Designed and implemented a framework named SymSpectra for molecular structure identification using multimodal spectral data (IR, ¹H‑NMR, ¹³C‑NMR, MS/MS), and proposed an improved solution to address the problem of imbalanced functional group distribution.
Synergistic Intra- and Cross-Layer Regularization Losses for MoE Expert Specialization
Rizhen Hu (Peking University), Kun Yuan (Peking University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes two regularization losses that do not require modifying the model architecture, aimed at improving the specialization of experts and routing efficiency in Mixture-of-Experts (MoE) models.
Synergistic Space-Vision Processing for Predicate Inference
Zhenhua Lei (Xi'an Jiaotong University), yu qiu
RecognitionGenerationRepresentation LearningConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Through the dual-stream collaborative network DSNet, geometric and non-geometric relationships in scene graph generation are modeled separately, and bidirectional cross-attention fusion is used to reduce representation entanglement caused by co-occurrence, thereby enhancing the semantic depth of relationship recognition.
SynerMedGen: Synergizing Medical Multimodal Understanding with Generation via Task Alignment
Weiren zhao, Cheng Chen (University of Hong Kong)
GenerationData SynthesisRepresentation LearningTransformerMixture of ExpertsVision Language ModelDiffusion modelScore-based ModelAuto EncoderImageMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyPositron Emission Tomography
🎯 What it does: Proposed the SynerMedGen unified framework, which combines the generation and alignment understanding tasks to achieve collaborative training for medical multimodal understanding and generation.
SynGR: Unleashing the Potential of Cross-Modal Synergy for Generative Recommendation
Wei Chen (Beihang University), Fuzhen Zhuang (Beihang University)
Recommendation SystemTransformerAuto EncoderContrastive LearningTextMultimodality
🎯 What it does: Propose the SynGR framework, which activates cross-modal collaborative information in generative recommendation by explicitly suppressing single-modal dependencies, introducing saliency masking, and employing collaborative contrastive learning, achieving better generative recommendations.
SynLaD: Latent Diffusion for Generating Synthesizable Molecules Conditioned on 3D Pharmacophore Profiles
Miruna Cretu (University of Cambridge), Colin A Grambow
Drug DiscoveryTransformerDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningGraphBiomedical Data
🎯 What it does: SynLaD learns a unified latent space through a dual-head variational autoencoder (3D geometry decoder + synthetic pathway decoder), and generates novel molecules that meet specified 3D pharmacophore features and are synthetically feasible by utilizing a latent diffusion model in this space.
Syntax vs. Semantics: How Transformers Learn Deep Dependencies
JiangRui Zhao, Xiaoting Du (Beijing University of Posts and Telecommunications)
Explainability and InterpretabilityRepresentation LearningAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextSequentialChain-of-Thought
🎯 What it does: This paper studies how Transformer models gradually learn deep semantic dependencies from surface syntax during training, and explains this dynamic process using a gradient competition framework.
Synthesizable Molecular Generation via Soft-constrained GFlowNets with Rich Chemical Priors
Hyeonah Kim (Mila - Quebec AI Institute), Alex Hernández-García
Drug DiscoveryGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningFlow-based ModelContrastive LearningTextGraphTabular
🎯 What it does: Proposed the S3-GFN framework, which generates synthetically feasible molecules using a soft-constrained GFlowNet and a pre-trained SMILES language model.
Synthesizing Multimodal Geometry Datasets from Scratch and Enabling Visual Alignment via Plotting Code
Haobo Lin (Jilin University), Binhang Yuan (Hong Kong University of Science and Technology)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Built a pipeline for synthesizing multimodal geometry problems from scratch and generated the GeoCode dataset, which includes graphics, text, reasoning processes, and drawing code.
Systematic Failures in Collective Reasoning under Distributed Information in Multi-Agent LLMs
Yuxuan Li (Carnegie Mellon University), Hirokazu Shirado (Carnegie Mellon University)
Federated LearningExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper constructs a hidden information benchmark called HIDDENBENCH with 65 questions to systematically evaluate the collective reasoning ability of multi-agent large language models under distributed information.
T-Edit: Triple-Branch Diffusion Anchoring for Consistent Editing
Linsong Shan (Huazhong University of Science and Technology), Qilin Yang (Huazhong University of Science and Technology)
Image TranslationRestorationGenerationTransformerPrompt EngineeringDiffusion modelAuto EncoderImageVideo
🎯 What it does: Proposed a training-agnostic three-branch diffusion framework named T-Edit for achieving structure-consistent image and video editing on the MMDiT (multimodal diffusion transformer) model.
T-GINEE: A Tensor-Based Multi-Graph Representation Learning
Maolin Wang (City University of Hong Kong), Xiangyu Zhao (City University of Hong Kong)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: Proposed a tensor-based generalized estimating equations framework, T-GINEE, for representation learning on multi-layer graphs
T-measure: A Topology-Consistent Metric for Binary Segmentation
pengfei zhang, Jian Ji (Xidian University)
SegmentationContrastive LearningImage
🎯 What it does: Proposed a binary segmentation evaluation metric called T-measure that takes into account topological consistency.
T-POP: Test-Time Personalization with Online Preference Feedback
Zikun Qu (Chinese University of Hong Kong), Zhongxiang Dai (Chinese University of Hong Kong)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Achieve real-time personalization for large language models by collecting user pairwise preference feedback online, learning a reward function, and guiding the decoding process of a frozen model during testing;
T$^2$PO: Uncertainty-Guided Exploration Control for Stable Multi-Turn Agentic Reinforcement Learning
Haixin Wang (University of California, Los Angeles), Yizhou Sun (University of California, Los Angeles)
TransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringTextSequentialChain-of-Thought
🎯 What it does: Proposes a Token- and Turn-level uncertainty-guided exploration control framework, T PO, to achieve stable training in multi-turn reinforcement learning;
T2AV-Compass: Towards Unified Evaluation for Text-to-Audio-Video Generation
Zhe Cao (Nanjing University), Jiaheng Liu (Nanjing University)
GenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Constructed the T2AV-Compass benchmark with 500 complex multimodal text prompts and proposed a two-tier evaluation framework for systematically evaluating text-to-audio-visual generation models.
Tabero: Learning Gentle Manipulation with Closed-Loop Force Feedback from Vision, Touch, and Language
Qiwei Wu (Nvidia), Renjing Xu (Hong Kong University of Science and Technology (Guangzhou))
Autonomous DrivingOptimizationRobotic IntelligenceConvolutional Neural NetworkRecurrent Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelVision-Language-Action ModelDiffusion modelFlow-based ModelContrastive LearningImageTextMultimodalityPoint CloudTime SeriesBenchmark
🎯 What it does: Constructed the Tabero benchmark and VTLA model, utilizing high-fidelity tactile simulation and closed-loop force feedback to achieve gentle mechanical operations under language conditions
TabICLv2: A Better, Faster, Scalable, and Open Tabular Foundation Model
Jingang QU, Marine Le Morvan (INRIA Saclay)
ClassificationData SynthesisOptimizationComputational EfficiencyTransformerMixture of ExpertsDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningTabularBenchmark
🎯 What it does: Developed TabICLv2, a new table-based model that integrates an improved pre-training process, scalable attention mechanisms, a synthetic data generator, and various structural innovations;
TabMGP: Martingale Posterior with TabPFN
Kenyon Ng (Monash University), Susan Wei (Monash University)
ClassificationExplainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerMixture of ExpertsContrastive LearningTabular
🎯 What it does: Propose a Markov posterior (TabMGP) based on TabPFN, which directly obtains the posterior distribution of parameters θ from a pre-trained tabular foundation model through forward sampling.
TabPack: Efficient Hyperparameter Ensembles for Tabular Deep Learning
Yury Gorishniy (Yandex), Artem Babenko (Yandex)
Computational EfficiencyHyperparameter SearchData-Centric LearningMixture of ExpertsTabular
🎯 What it does: Developed TabPack, an efficient Tabular DL framework capable of building and online updating multiple model ensembles in parallel during a single training process.
TabSwift: An Efficient Tabular Foundation Model with Row-Wise Attention
Si-Yang Liu (Nanjing University), Han-Jia Ye (Nanjing University)
Computational EfficiencyRepresentation LearningTransformerPrompt EngineeringMixture of ExpertsContrastive LearningTabularBenchmark
🎯 What it does: Proposes TABSWIFT, a lightweight row-wise attention-only Tabular Foundation Model that supports classification and regression, and improves pretraining stability by incorporating learnable register tokens and gated attention, while achieving per-sample adaptive early exiting.
TabularBERT: Binning-Based Self-Supervised Learning for Tabular Representation
Beomjin Park (Gyeongsang National University), Hosik Choi (University of Seoul)
Representation LearningData-Centric LearningTransformerSupervised Fine-TuningAuto EncoderContrastive LearningTabularBenchmark
🎯 What it does: Propose TABULARBERT, a Transformer-based framework for tabular data representation learning, which maps continuous variables through binning (tokenization) to discrete tokens, and learns conditional dependencies and numerical order relationships by combining masked language modeling (MLM) self-supervised pre-training with numerical reconstruction tasks;
TABX: A High-Throughput Sandbox Battle Simulator for Multi-Agent Reinforcement Learning
Hayeong Lee (Korea University), Byung-Jun Lee (Korea University)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes TABX — a highly configurable, high-throughput multi-agent combat simulation environment based on JAX, designed for systematic evaluation of MARL algorithms;
Tackling Fake Forgetting through Uncertainty Quantification
Yingdan Shi (Illinois Institute of Technology), Ren Wang (Illinois Institute of Technology)
ClassificationFederated LearningExplainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: This paper addresses the 'false forgetting' problem in machine learning by proposing an evaluation metric CR based on conformal prediction and an improved CPU framework, which enhances the real forgetting quality of the model.
Tackling Length Inflation Without Trade-offs: Group Relative Reward Rescaling for Reinforcement Learning
Zichao Li (Chinese Information Processing Laboratory Institute of Software Chinese Academy of Sciences), XingYu Li
OptimizationData-Centric LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextMultimodality
🎯 What it does: This paper proposes a multi-modal RL method based on Group Relative Reward Re-calibration (GR3) to address the problem of length inflation in RL training.
TACTIC: Task-Aware Sparse Coordination Graphs for Multi-Task Multi-agent Reinforcement Learning
Kexing Peng (Nanjing University of Information Science & Technology), Jianye HAO
Graph Neural NetworkReinforcement LearningMixture of ExpertsAuto EncoderContrastive LearningGraphTabular
🎯 What it does: Propose a multi-task multi-agent reinforcement learning framework called TACTIC, which can achieve adaptive collaboration in environments with implicitly changing tasks through trajectory abstraction and sparse coordination graphs.
TadA-Bench: A Million-Variant Benchmark for Future-Round Discovery Toward Agentic Protein Engineering
Jin Gao (Shanghai Jiao Tong University), Dequan Wang (Shanghai Jiao Tong University)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerSupervised Fine-TuningPrompt EngineeringContrastive LearningGraphTime SeriesBiomedical DataBenchmark
🎯 What it does: Propose TadA-Bench, a million-variant replayable benchmark based on 31 rounds of TadA directed evolution experiments, used to evaluate models' ability to rank new variants in future experimental rounds.
TAG: Tangential Amplifying Guidance for Hallucination-Resistant Sampling
Hyunmin Cho (Korea University), Kyong Hwan Jin (Korea University)
GenerationData SynthesisPrompt EngineeringDiffusion modelScore-based ModelImageVideoStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a reasoning-time guidance method called TAG, which suppresses hallucinations in diffusion models and improves generation quality by amplifying the tangential component of noisy sample trajectories.
TAGRPO: Boosting GRPO on Image-to-Video Generation with Direct Trajectory Alignment
Jin Wang (University of Hong Kong), Ping Luo (University of Hong Kong)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelContrastive LearningImageVideoText
🎯 What it does: Propose the TAGRPO framework, which performs post-training on image-to-video generation models, leveraging trajectory alignment and a memory pool to enhance rewards.
Tail Annealing for Heavy-Tailed Flow Matching
Jean Pachebat (Institut Polytechnique de Paris)
GenerationData SynthesisFlow-based ModelTabularFinance RelatedOrdinary Differential Equation
🎯 What it does: The study applies a soft logarithmic transformation to heavy-tailed data before training, compressing it into an exponential tail, and then performs flow matching in the log space to achieve stable generation of heavy-tailed data.
Tailoring Strictly Proper Scoring Rules for Downstream Tasks: An Application to Causal Inference
Roman Plaud (Institut Polytechnique de Paris), Matthieu Labeau (Institut Polytechnique de Paris)
Anomaly DetectionOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement LearningScore-based ModelContrastive LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: Designed and derived a strict positive scoring rule tailored for downstream tasks, and applied it to the inverse probability weighting (IPW) average treatment effect (ATE) estimation in causal inference, proposing a complete framework ranging from downstream error curvature matching to closed-form loss and its corresponding canonical probability mapping.
Tailoring the Training: Difficulty-Aware Learning Strategy Allocation for Large Language Models
Xiaoling Zhou (Peking University), Shikun Zhang (Peking University)
OptimizationComputational EfficiencyRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the DALSA framework, which dynamically allocates SFT, RL, or discard strategies based on sample difficulty to enhance LLM inference capabilities.
Taking the GP Out of the Loop
Mehul Bafna (Yeshiva University), David Sweet (Yeshiva University)
OptimizationComputational EfficiencyHyperparameter SearchTabularTime Series
🎯 What it does: Proposed a lightweight K-nearest neighbors surrogate called Epistemic Nearest Neighbors (ENN), as an alternative to the traditional high-cost Gaussian Process (GP), and integrated it into the TuRBO optimization framework to form TuRBO-ENN; meanwhile, provided a fast acquisition scheme without fitting, based on non-dominated sorting for noise-free scenarios.
Talk, Judge, Cooperate: Gossip-Driven Indirect Reciprocity in Self-Interested LLM Agents
Shuhui Zhu (University of Waterloo), Pascal Poupart (University of Waterloo)
Autonomous DrivingOptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyAdversarial AttackDrug DiscoveryAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AITextBenchmarkChain-of-Thought
🎯 What it does: We constructed the ALIGN framework, enabling self-interested LLM agents to achieve indirect reciprocity through public, hierarchical verbal comments, and validated its effectiveness in various games and market simulations.
Taming Aleatoric Impulse in Off-Policy Reinforcement Learning
Zhouyang Yu (Tsinghua University), Shengbo Eben Li (Tsinghua University)
Reinforcement LearningTabularTime Series
🎯 What it does: This paper first identifies and addresses the "Aleatoric Impulse" phenomenon caused by discrete randomness, proposing DSAC-AID, which suppresses early estimation bias by implementing adaptive lazy evaluation and optimistic exploration within distributed soft actor-critic.
Taming I2V models for Image HOI Editing: A Cognitive Benchmark and Agentic Self-Correcting Framework
Jiayi Gao (Peking University), Yang Liu (Peking University)
Image TranslationRestorationObject DetectionObject TrackingSegmentationGenerationData SynthesisPose EstimationDepth EstimationOptimizationTransformerAgentic AIPrompt EngineeringVision Language ModelDiffusion modelScore-based ModelOptical FlowImageVideoTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Studied dynamic human-object interaction (HOI) editing based on image-to-video models, proposed the HOI-Edit benchmark and HOI-Eval evaluation method, and introduced the SCPE self-correction framework.
Taming Stochastic Gradient Descent: Almost Sure Convergence and Saddle-Point Avoidance under $(L_{0},L_{1})$-Smoothness
Vassilis Apidopoulos (Athens University of Economics & Business), Panayotis Mertikopoulos (University of Grenoble Alpes)
OptimizationStochastic Differential Equation
🎯 What it does: In this paper, two improved stochastic gradient descent (SGD) algorithms—Tamed SGD (TSGD) and Dissipatively Tamed SGD (DTSGD)—are proposed and analyzed. It is proven that they almost surely converge and can avoid strict saddle points under the condition that the non-convex objective function is (L, L₀, L₁)-smooth, even if it may have exponentially growing gradients.
Taming the Loss Landscape of PINNs with Noisy Feynman–Kac Supervision: Operator Preconditioning and Non-Asymptotic Error Bounds
Nathanael Tepakbong (City University of Hong Kong), Xiang ZHOU
OptimizationSupervised Fine-TuningTabularTime SeriesSequentialPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose an improved method called FK-PINN, which incorporates a small number of Feynman-Kac Monte-Carlo supervised points into the PINN loss to precondition the loss landscape, thereby enhancing training stability and convergence speed.
Taming the Recent-Data Bias: Towards Robust Time Series Forecasting with Global Context
Longlong Xu (Tsinghua University), Dan Pei (Tsinghua University)
Anomaly DetectionTransformerSupervised Fine-TuningAuto EncoderContrastive LearningTime Series
🎯 What it does: This paper systematically studies the普遍存在 'recent data bias' in time series forecasting and proposes a new method called TameR. By using random sampling and basis function projection (BARS), learnable periodic extraction (LPE), and two-stage training, TameR significantly improves robustness to recent data perturbations while maintaining good prediction accuracy.
TAMPO: Task- and Model-Aware Automatic Prompt Optimization for Auto-Routing in LLM-based Systems
Yuejun Jiao (East China Normal University), Mingsong Chen (East China Normal University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsText
🎯 What it does: Propose TAMPO, a task- and model-aware automatic prompt optimization framework for LLM automatic routing, which can automatically generate prompts that meet multi-objective requirements for different tasks and models in a single generation.
TapSampling: Inference-Time Sampling with a Task-Progress-Understanding Verifier for Robotic Manipulation
Sizhe Zhao (Harbin Institute of Technology), Xiangyang Ji (Tsinghua University)
Robotic IntelligenceTransformerReinforcement LearningPrompt EngineeringVision-Language-Action ModelAuto EncoderMultimodalityTime SeriesSequential
🎯 What it does: Proposes a sampling framework called TapSampling during inference, which uses Action-VAE to generate diverse action candidates and selects the optimal action through a task progress understanding validator, thereby improving the performance of general robot policies without fine-tuning the strategy.
TarGATE: Target-Aware Data Selection via Token-Attenuation Gates
Xiandi Luo (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningText
🎯 What it does: Propose the TarGATE framework, which inserts a lightweight gating mechanism before the FFN in the Transformer, using token-level information retention ratio (IRR) to evaluate and filter high-quality training data that matches the target task.
Target-Agnostic Calibration under Distribution Shift with Frequency-Aware Gradient Rectification
Yilin Zhang (Xidian University), Wei Zhao (Xidian University)
ClassificationDomain AdaptationConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Propose a target-agnostic training framework FGR, achieving adaptive calibration under distribution shift through frequency-domain low-pass filtering and gradient correction.
Target-Aware Bandit Allocation for Scalable Surrogate Optimization in Chemical Space
Mohammad Haddadnia (Dana-Farber Cancer Institute), Haribabu Arthanari (Dana-Farber Cancer Institute)
OptimizationDrug DiscoveryTransformerReinforcement LearningContrastive LearningGraphTabular
🎯 What it does: This paper proposes the BOBA framework, which partitions a huge chemical library by chemical structure and treats each partition as an arm of a multi-armed bandit. It employs uncertainty-based UCB1 for inference and resource allocation, and performs local Bayesian optimization within the selected subspace, achieving efficient candidate screening without full-library inference.
Target-Driven Policy Optimization for Sequential Counterfactual Outcome Control
Xin Wang (University of Science and Technology of China), Huanhuan Chen (University of Science and Technology of China)
OptimizationDrug DiscoveryRecurrent Neural NetworkTransformerReinforcement LearningTabularTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: GIFT models clinical goal achievement tasks as goal-driven Markov decision processes, learning real-time adjustable intervention strategies from offline observational data.
Target-Oriented Pretraining Data Selection via Neuron-Activated Graph
Zijun Wang (UC Santa Cruz), Fengze Liu (ByteDance)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Studies a method for untrained, interpretable screening and ranking of goal-oriented pre-training data through Neuron-Activated Graph (NAG).
Task-and-Model-Aware Fractal-Consistency for Efficient LLM Reasoning
Ziqiu Luo (Tsinghua University), Dan Li (Tsinghua University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Task-and-Model-Aware Fractal-Consistency (TMAFC) framework, which reduces the sampling cost of large language model inference through component-wise self-consistency evaluation and adaptive stopping criteria.
Task-Aware Exploration via a Predictive Bisimulation Metric
Dayang Liang (Xiamen University), Bo An (Nanyang Technological University)
Convolutional Neural NetworkTransformerReinforcement LearningWorld ModelImageTabularBenchmark
🎯 What it does: Propose a task-aware exploration method called TEB based on predictive bisimulation metric, which can simultaneously learn task-related representations and generate global exploration rewards in visual reinforcement learning with sparse rewards;
Task-Aware Mechanism: Hybrid MoE Vision Tower Towards Holistic Video Understanding
Qishen Yin (Peking University), Li Yuan (Peking University)
ClassificationExplainability and InterpretabilityComputational EfficiencyTransformerPrompt EngineeringMixture of ExpertsVision Language ModelContrastive LearningVideoText
🎯 What it does: Proposed a task-aware mechanism called TAM, constructing a Mixture-of-Experts visual tower with hybrid gating, which can dynamically adjust the number of frames and resolution according to the query intent and video duration, and trained a 0.1B text Inductor based on the TA-116K dataset for task classification;
Task-Aware Preference Calibration for Direct Preference Optimization
Mengyang Li (Tianjin Normal University), Pinlong Zhao (Hangzhou Dianzi University)
OptimizationMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes a task-aware preference calibration (TAPC) method to improve the reliability of perplexity adaptation in different task types within direct preference optimization (DPO).
Task-Aware Structured Memory for Dynamic Multi-modal In-Context Learning
Zhirui Chen (University of Chinese Academy of Sciences), Ling Shao (University of Chinese Academy of Sciences)
CompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageVideoTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a training-agnostic Task-Aware Structured Memory (TASM) framework to compress the KV cache of multi-modal large language models and achieve dynamic retrieval.
Task-Awareness Improves LLM Generations and Uncertainty
Tim Tomov (Technical University of Munich), Stephan Günnemann
ClassificationGenerationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes a method that maps the generation results of large language models (LLMs) into a task-specific latent structure space, and directly performs minimum Bayes risk (MBR) decoding on this space, generating task results that are superior to those produced by traditional decoding methods, while quantifying uncertainty through Bayesian risk.
Task-Driven Subspace Decomposition for Knowledge Sharing and Isolation in LoRA-based Continual Learning
Lingfeng He (Xidian University), Xinbo Gao (Xidian University)
ClassificationComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This paper proposes a continuous learning framework called LoDA based on LoRA, which utilizes task-driven projection energy decomposition to divide the update space into a general subspace and a specific subspace. It constructs a dual-branch LoRA module to achieve knowledge sharing and isolation, and after each task, it achieves feature-level joint optimization through closed-form re-calibration.
TaskLoom: Weaving Knowledge Across Tasks in World Models
Qingzhang Zeng (Peking University), Yonghong Tian (Peking University)
TransformerReinforcement LearningMixture of ExpertsWorld ModelImageVideoTabular
🎯 What it does: Propose the TaskLoom framework, achieving grouped two-stage knowledge sharing of world models in online multi-task reinforcement learning.
Taylor-Gaussians-Flow: Towards Non-uniform Motion for Novel View Synthesis from Monocular Video
Zaoming Yan (East China Normal University), Faming Fang (East China Normal University)
GenerationData SynthesisTransformerDiffusion modelScore-based ModelFlow-based ModelRectified FlowNeural Radiance FieldGaussian SplattingOptical FlowImageVideo
🎯 What it does: Proposes Taylor-Gaussians motion modeling based on Taylor series expansion and a differentiable module called TGs-Flow that projects 3D high-order motion into 2D optical flow, achieving precise modeling and supervision of long-term non-uniform motion, thereby improving the quality of dynamic view synthesis from monocular videos over time.
TCAP: Tri-Component Attention Profiling for Unsupervised Backdoor Detection in MLLM Fine-Tuning
Mingzu Liu, Runmin Cong
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose an unsupervised backdoor detection framework called TCAP, which identifies and cleans polluted training samples by leveraging the differences in three-component attention allocation within a multi-modal large language model.
TD-VAD: Breaking Visual Dependence in Video Anomaly Detection with Text-Driven Learning
Shuangqing Zhang (Nanjing University), Fang Zhao (Nanjing University)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningVideoText
🎯 What it does: Propose a video anomaly detection model entirely trained based on textual descriptions, eliminating the need for reliance on target video data.
TD3B: Transition-Directed Discrete Diffusion for Allosteric Binder Generation
Hanqun Cao (Chinese University of Hong Kong), Pranam Chatterjee (University of Pennsylvania)
Drug DiscoveryProtein Structure PredictionTransformerSupervised Fine-TuningDiffusion modelContrastive LearningSequentialBiomedical Data
🎯 What it does: Propose a protein ligand generation framework TD3B based on a discrete diffusion model, which can generate corresponding agonists or antagonists according to the activation or inhibition needs of the target protein;
TDM-R1: Reinforcing Few-Step Diffusion Models with Non-Differentiable Reward
Yihong Luo (Hong Kong University Of Science And Technology), Jing Tang (Hong Kong University Of Science And Technology)
GenerationReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringDiffusion modelScore-based ModelContrastive LearningImageTextMultimodalityRetrieval-Augmented GenerationOrdinary Differential Equation
🎯 What it does: Propose TDM-R1, a reinforcement learning framework based on Trajectory Distribution Matching, which can improve the generation quality of few-step diffusion models by utilizing non-differentiable rewards.
Teaching Models to Teach Themselves: Reasoning at the Edge of Learnability
Shobhita Sundaram (Massachusetts Institute of Technology), Julia Kempe (New York University)
OptimizationData-Centric LearningMeta LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose a self-generated curriculum framework called SOAR, which utilizes meta-reinforcement learning (meta-RL) to enable pre-trained large language models (LLMs) to automatically generate a series of easy-to-learn intermediate problems when facing extremely difficult mathematical problems that cannot be directly solved, thereby breaking through the learning plateau.
Teaching Molecular Dynamics to a Non-Autoregressive Ionic Transport Predictor
Jiyeon Kim (Yonsei University), Won-Yong Shin (Yonsei University)
Drug DiscoveryTransformerSupervised Fine-TuningMixture of ExpertsAuto EncoderContrastive LearningGraphTabularTime SeriesPhysics Related
🎯 What it does: Proposes a non-autoregressive learning framework that utilizes auxiliary modal learning to use atomic trajectory information during training but predicts ionic transport properties during inference based only on static structures.
TEAM: Temporal–Spatial Consistency Guided Expert Activation for MoE Diffusion Language Model Acceleration
Linye Wei (Peking University), Meng Li (Peking University)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelMixture of ExpertsDiffusion modelText
🎯 What it does: Proposed the TEAM framework, which utilizes spatiotemporal consistency to accelerate Mixture-of-Experts (MoE) diffusion large language models (dLLM)
TeamTR: Trust-Region Fine-Tuning for Multi-Agent LLM Coordination
Yi Xie (University of Arizona), Bo Liu (University of Arizona)
OptimizationFederated LearningComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: We study and propose TeamTR, a sequential fine-tuning framework based on the trust region, aimed at improving the coordination and overall reasoning performance of shared context multi-agent LLMs.
TeamWork: Multivariate Time Series Anomaly Detection via Asymmetric Role-aware Channel Modeling
Shiyan Hu (East China Normal University), Chenjuan Guo (East China Normal University)
Anomaly DetectionConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningTime SeriesBenchmarkFinance Related
🎯 What it does: Propose a multivariate time series anomaly detection framework named TeamWork, which classifies variables into dominant and auxiliary categories using the entropy reduction index, and achieves feature fusion through cycle-aware masks and role-aware gating interaction modules.
TEFormer: Structured Bidirectional Temporal Enhancement Modeling in Spiking Transformers
Sicheng Shen (Institute of Autmation, CAS), Yi Zeng (Institute of Autmation, CAS)
ClassificationSpiking Neural NetworkTransformerAuto EncoderContrastive LearningImageVideoTime SeriesSequential
🎯 What it does: Proposes TEFormer — a framework for achieving bidirectional temporal fusion in Spiking Transformers, utilizing forward-time enhanced attention (TEA) and backward gated recurrent MLP (T-MLP) to realize parallel forward fusion and backward feedback.
TelecomTS: A Multi-Modal Observability Dataset for Time Series and Language Analysis
Austin Feng (Yale University), Rex Ying (Yale University)
Data SynthesisAnomaly DetectionRecurrent Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextMultimodalityTime SeriesRetrieval-Augmented Generation
🎯 What it does: Proposed TelecomTS, a large-scale observability time series dataset collected from 5G communication networks, and constructed multi-modal question-answering and root cause analysis tasks.
Telescope: Improving Zero Shot Detection of LLM Generated Content By Measuring Token Repetition Probability
Christopher Nassif (Virginia Tech), Joshua F. Cooper
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a zero-shot detection method called Telescope Perplexity, which measures the repetitiveness of text by utilizing the probability of the same token appearing in the previous context as determined by large language models, thereby distinguishing between content generated by LLMs and human writing.
Temper-Then-Tilt: Principled Unlearning for Generative Models through Tempering and Classifier Guidance
Jacob L. Block (University of Texas at Austin), Sanjay Shakkottai (University of Texas at Austin)
GenerationData SynthesisSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelDiffusion modelScore-based ModelTextBenchmark
🎯 What it does: This paper studies machine unlearning in large-scale generative models, by treating the unlearning task as a distribution ratio estimation. After flattening the base model distribution with temperature scaling, a lightweight classifier is used to tilt the distribution, effectively removing the forgotten samples.
Temperature Scaling in Discrete Sequence (Language) Models
Hannah Scheufele (Georg-August-Universität Göttingen), Vikas K Garg (Aalto University)
GenerationComputational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelScore-based ModelText
🎯 What it does: This paper proposes a unified framework and implementation method for sequence-level temperature scaling (Sequence-Level Temperature Scaling, SSTS) in discrete sequence models (autoregressive and discrete diffusion models);
Tempora: Characterising the Time-Contingent Utility of Online Test-Time Adaptation
Sudarshan Sreeram (University of Cambridge), Cecilia Mascolo (University of Cambridge)
ClassificationDomain AdaptationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: Introduce the Tempora framework, systematically evaluate the effectiveness of test-time adaptation (TTA) under different time pressures, and define three time-related utility metrics (discrete, continuous, amortized)
Temporal Context Reinstatement Drives Episodic-Like Order Memory in Long-Context Language Models
Mathis Pink (Max Planck Institute for Software Systems), Mariya Toneva (Max Planck Institute for Software Systems)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper uses long-context LLMs to encode full novels and evaluates the model's performance against humans on a binary classification temporal sequence task.
Temporal Difference Calibration in Sequential Tasks: Application to Vision-Language-Action Models
Shelly Francis-Meretzki (Technion Israel Institute of Technology), Aviv Tamar (Technion Israel Institute of Technology)
Robotic IntelligenceReinforcement LearningVision-Language-Action ModelMultimodalitySequential
🎯 What it does: Proposes a calibration method for sequential tasks called TDQC, which uses the temporal difference (TD) loss to learn the probability prediction of task success.
Temporal Difference Learning for Diffusion Models
Qizhen Ying (University of Oxford), Junfeng Wen (Carleton University)
GenerationConvolutional Neural NetworkTransformerReinforcement LearningDiffusion modelScore-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes a training objective for diffusion models based on temporal difference (TD) learning, which enforces cross-time consistency across the entire denoising trajectory, thereby improving the generation quality of few-step sampling.
Temporal Preference Optimization for Unsupervised Retrieval
HyunJin Kim (Sungkyunkwan University), JinYeong Bak (Sungkyunkwan University)
RetrievalOptimizationRepresentation LearningTransformerReinforcement LearningContrastive LearningTextTime Series
🎯 What it does: This paper proposes the TPOUR framework, which utilizes unsupervised contrastive learning combined with time-priority optimization (TRPO) to train a dense retriever, enabling it to consider temporal relevance during retrieval.
Temporal Score Rescaling for Temperature Sampling in Diffusion and Flow Models
Yanbo Xu (Carnegie Mellon University), Shubham Tulsiani (Carnegie Mellon University)
RoboticsGenerationPose EstimationDepth EstimationProtein Structure PredictionDiffusion modelScore-based ModelFlow-based ModelImageTextPoint CloudTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes a training-free and computation-free in time domain score re-calibration (TSR) method, which can real-time adjust the trade-off between sampling diversity and sample likelihood without modifying model weights, by performing time-dependent scaling on the scores of pre-trained diffusion and flow matching models.
Temporal Self-Rewarding Language Models: Decoupling Chosen-Rejected via Past-Future
Yidong Wang (Peking University), Shikun Zhang (Peking University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the Temporal Self-Rewarding framework, which uses past fixed negative samples from the model and future selected positive samples from the model to address the problem of gradient disappearance caused by the gradual convergence of similarity between positive and negative samples in traditional self-rewarding models.
Temporal Straightening for Latent Planning
Ying Wang (New York University), Mengye Ren (New York University)
OptimizationRobotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningAuto EncoderContrastive LearningWorld ModelImageSequential
🎯 What it does: Proposed and implemented the Temporal Straightening technique in autoregressive world models, achieved through curvature regularization, to improve the feasibility of planning in the latent space.
Temporal Weighted Encoding: Towards Maximal-Capacity Spike Coding for ANN–SNN Conversion
Yiwen Gu (Zhejiang University), Kejie Huang (Zhejiang University)
ClassificationImage TranslationObject DetectionOptimizationComputational EfficiencyKnowledge DistillationConvolutional Neural NetworkSpiking Neural NetworkPrompt EngineeringAuto EncoderContrastive LearningImageVideoTabularTime Series
🎯 What it does: Propose Temporal Weighted Encoding (TWE), which achieves high-capacity temporal encoding for ANN-SNN conversion through recursive integration to realize exponentially decaying weight encoding.
Temporal-aware Flow Matching for Video Generation with Temporally Coherent Motion
Zirui Pan (Tsinghua University), Wenwu Zhu (Tsinghua University)
GenerationTransformerSupervised Fine-TuningFlow-based ModelVideoOrdinary Differential Equation
🎯 What it does: Propose a Temporal-aware Flow Matching (TFM) training paradigm that explicitly models inter-frame motion constraints in video generation, thereby enhancing the spatiotemporal coherence of generated videos.
Temporal-Aware Reasoning Optimization for Video Temporal Grounding
Minghang Zheng (Peking University), Yang Liu (Peking University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringVision Language ModelVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented the TaRO framework, which enhances the reasoning ability of multimodal large language models in video temporal localization tasks through constructive reasoning exploration and time-sensitive rewards.
Temporal-Emerged Prompting for Segment Anything in Multiframe Infrared Small Target Detection
Yinghui Xing (Northwestern Polytechnical University), Di Xu (Huawei Technologies Ltd)
Object DetectionObject TrackingSegmentationRecurrent Neural NetworkTransformerPrompt EngineeringContrastive LearningOptical FlowImageVideoBenchmark
🎯 What it does: Propose the TEP-SAM framework, which utilizes gradually emerging target motion information over time. By employing a Discrepancy-Enhanced Temporal Encoder and a Temporal Prompt Generator, temporal features are injected into SAM to achieve interactive-free multi-frame infrared small target detection and segmentation.