ICLR 2026 Papers — Page 46

Structural Inference: Interpreting Small Language Models with Susceptibilities

Garrett Baker, Daniel Murfet

🎯 What it does: This paper proposes a linear response-based explanation framework called susceptibility, which views neural networks as differentiable systems from the perspective of Bayesian statistical mechanics, and quantifies the response of each network component (e.g., attention heads) to minor changes in data distribution; it efficiently estimates the susceptibility of each component through local SGLD sampling, and then performs PCA on the response matrix via structural inference to identify functional modules within the model (e.g., induction circuits, multi-word heads, parenthesis-matching heads) and explain activation and suppression behaviors.

Structural Prognostic Event Modeling for Multimodal Cancer Survival Analysis

Yilan Zhang (King Abdullah University of Science and Technology), Xin Gao (King Abdullah University of Science and Technology)

🎯 What it does: Proposed a Slot-based structured prognosis event modeling framework, SlotSPE, for multi-modal (tissue slide + genomic) cancer survival analysis.

Structurally Human, Semantically Biased: Detecting LLM-Generated References with Embeddings and GNNs

Melika Mobini (Vrije Universiteit Brussel), Vincent Ginis (Vrije Universiteit Brussel)

🎯 What it does: Compare the citation network structure and semantic features of LLM-generated references with those of real references.

Structure Learning from Time-Series Data with Lag-Agnostic Structural Prior

Taiyu Ban (University of Science and Technology of China), Huanhuan Chen (University of Science and Technology of China)

🎯 What it does: Integrate temporal structure learning with lag-agnostic structural priors to propose a continuous optimization framework for precise lag recovery.

Structure-Aware Graph Hypernetworks for Neural Program Synthesis

Wenhao Li (University of Toronto), Scott Sanner (University of Toronto)

🎯 What it does: This paper synthesizes neural programs in a continuous weight space through meta-learning and hypernetworks, directly generating complete weights of fixed template networks from user intent.

Structured Flow Autoencoders: Learning Structured Probabilistic Representations with Flow Matching

Yidan Xu (University of Michigan), XuanLong Nguyen (University of Michigan)

🎯 What it does: Propose a Structured Flow Autoencoder (SFA) that combines graphical models with conditional continuous normalizing flows to achieve high-fidelity generation and structured latent representation learning.

Structured Reasoning for LLMs: A Unified Framework for Efficiency and Explainability

Yubo Dong (Zhejiang University), Yi Yang (Zhejiang University)

🎯 What it does: Proposes a structured reasoning framework that models the reasoning process as a sparse directed graph, and enhances reasoning efficiency, stability, and interpretability by applying reinforcement learning to LLMs with MaxFlow and LCS rewards.

Study of Training Dynamics for Memory-Constrained Fine-Tuning

Aël Quélennec (Télécom Paris), Enzo Tartaglione (Télécom Paris)

🎯 What it does: In extreme memory-constrained environments, transfer learning is applied to pre-trained models, proposing the TraDy dynamic channel selection scheme to achieve efficient fine-tuning;

STVG-R1: Incentivizing Instance-Level Reasoning and Grounding in Videos via Reinforcement Learning

Xiaowen Zhang (Xidian University), Qing Li (State Key Laboratory of General Artificial Intelligence, BIGAI)

🎯 What it does: By overlaying unique instance ID visual prompts on video frames, dense coordinate regression is transformed into an instance-level identification problem, and end-to-end optimization of spatiotemporal localization is achieved using the reinforcement learning framework STVG-R1.

StyliTruth : Unlocking Stylized yet Truthful LLM Generation via Disentangled Steering

Chenglei Shen (Renmin University of China), Jun Xu (Renmin University of China)

🎯 What it does: This paper achieves stylized generation while maintaining answer truthfulness by performing representation editing on the attention heads of large language models, and proposes a new framework named StyliTruth.

Stylos: Multi-View 3D Stylization with Single-Forward Gaussian Splatting

Hanzhou Liu (Texas A&M University), Peng Jiang (Texas A&M University)

🎯 What it does: Proposed a single-forward 3D Gaussian stylization framework named Stylos, which can instantly generate multi-view stylized scenes from unposed images

SubDyve: Subgraph-Driven Dynamic Propagation for Virtual Screening Enhancement Controlling False Positive

Jungseob Yi (Seoul National University), Sangseon Lee (Inha University)

🎯 What it does: Construct a subgraph fingerprint network under low-label conditions, and enhance early identification in virtual screening through dynamic seed refinement calibrated by local false discovery rate.

Sublinear Spectral Clustering Oracle with Little Memory

Ranran Shen (University of Science and Technology of China), Zengfeng Huang (Fudan University)

🎯 What it does: Investigated the technology for constructing sublinear spectral clustering oracles under sparse memory, providing a scheme to perform graph clustering with memory usage far below the traditional √n;

Sublinear Time Quantum Algorithm for Attention Approximation

Zhao Song (Simons Institute for Theory of Computing, UC Berkeley), Lichen Zhang (MIT Computer Science and Artificial Intelligence Laboratory)

🎯 What it does: Proposes a quantum data structure to approximate the Transformer's attention matrix in sublinear time under a row query model, along with corresponding preprocessing and query algorithms.

Submodular Function Minimization with Dueling Oracle

Kaien Sho (University of Tokyo), Shinji Ito (University of Tokyo)

🎯 What it does: This paper proposes minimizing submodular functions under the condition of using only a dueling oracle, and provides algorithms along with error upper and lower bounds for linear and sigmoid transfer functions.

Subquadratic Algorithms and Hardness for Attention with Any Temperature

Shreya Gupta (University of Washington), Christopher Ye (University of California San Diego)

🎯 What it does: This paper addresses the Attention mechanism in Transformers, proposing a sub-quadratic approximation algorithm for computing Attention when the head dimension is a constant, and provides the corresponding lower bound, elucidating the complexity of Attention computation under different head dimensions and temperature ranges;

Subspace Kernel Learning on Tensor Sequences

Lei Wang (Griffith University), Piotr Koniusz (University of New South Wales)

🎯 What it does: This paper proposes a kernel learning framework named UKTL for subspace kernel comparison on multi-dimensional tensors, achieving end-to-end training through uncertainty weighting, dynamic Nyström linearization, and multi-modal fusion.

SUIT: Knowledge Editing with Subspace-Aware Key-Value Mappings

Haewon Park (Seoul National University), Yohan Jo (Seoul National University)

🎯 What it does: Proposes the SUIT method, which edits knowledge in language models through subspace-constrained key vectors and residual vectors.

Summaries as Centroids for Interpretable and Scalable Text Clustering

Jairo Diaz-Rodriguez (York University)

🎯 What it does: Proposes two variants of k-means, k-NLPmeans and k-LLMmeans, which periodically replace numerical centroids with text summaries to enhance clustering interpretability and effectiveness, and extend to a mini-batch version for real-time streaming text clustering.

SumRA: Parameter Efficient Fine-tuning with Singular Value Decomposition and Summed Orthogonal Basis

Chin Yuen Kwok (Nanyang Technological University), Eng Siong Chng (Nanyang Technological University)

🎯 What it does: Propose the SumRA method, which uses singular value decomposition (SVD) to initialize the low-rank matrix A with multi-vector summation in LoRA adaptation, and freezes A to reduce trainable parameters.

SupCLAP: Controlling Optimization Trajectory Drift in Audio-Text Contrastive Learning with Support Vector Regularization

Jiehui Luo (Peking University), Dongchao Yang (Peking University)

🎯 What it does: Proposed the SupCLAP framework, which employs support vector regularization (SVR) to address the optimization trajectory drift problem in audio-text contrastive learning.

SuperF: Neural Implicit Fields for Multi-Image Super-Resolution

Sander Riisøen Jyhne (University of Agder), Nico Lang (University of Copenhagen)

🎯 What it does: Proposed a multi-image super-resolution method called SuperF based on test-time optimization, achieving high-resolution reconstruction by sharing an implicit neural field (INR) and jointly optimizing sub-pixel alignment.

Superficial Safety Alignment Hypothesis

Jianwei Li (North Carolina State University), Jung-Eun Kim (North Carolina State University)

🎯 What it does: This paper proposes the 'Surface Safety Alignment Hypothesis (SSAH)', achieving safety alignment for large language models (LLMs) by identifying and freezing neuron-level safety-critical units (SCUs), and explores using redundant units as alignment budget.

SuperMAN: Interpretable and Expressive Networks over Temporally Sparse Heterogeneous Data

Andrea Zerio (Aalborg University), Aleksejs Sazonovs (Aalborg University)

🎯 What it does: Proposed the SUPERMAN framework to directly learn heterogeneous and sparse time series data by modeling them as implicit graph collections, providing interpretability at the node, graph, and subset levels.

Supervised Fine-Tuning or Contrastive Learning? Towards Better Multimodal LLM Reranking

Xin Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)

🎯 What it does: Compare the effectiveness of Contrastive Learning (CL) and Supervised Fine-Tuning (SFT) in re-ranking for large multimodal language models (LLMs), propose a unified loss splitting framework (weight + direction), build a general-purpose multimodal re-ranking model GMR, and release the MRB evaluation benchmark covering 40 datasets.

Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning

Yihe Deng (UCLA), Chen-Yu Lee (Google Cloud AI Research)

🎯 What it does: Proposed and implemented a Supervised Reinforcement Learning (SRL) framework that decomposes complex problems into a series of executable actions. The model generates internal reasoning before executing actions, enabling small-scale LLMs to learn on challenging tasks.

Supporting High-Stakes Decision Making Through Interactive Preference Elicitation in the Latent Space

Michael Eichelbeck (Technical University of Munich), Matthias Althoff (Technical University of Munich)

🎯 What it does: Propose an interactive preference mining framework that performs preference Bayesian optimization in the latent space of an autoencoder, integrating personalized priors generated by large language models (LLMs)

Supporting Multimodal Intermediate Fusion with Informatic Constraint and Distribution Coherence

Yi Li (Institute of Software Chinese Academy of Sciences), Jiangmeng Li (Institute of Software Chinese Academy of Sciences)

🎯 What it does: Propose a multi-modal learning framework called IID based on intermediate fusion (IF), which combines information-constrained linear objective mapping and distribution coordination with restricted isometric projection, theoretically proving that IF outperforms LF under specific constraints.

SurfSplat: Conquering Feedforward 2D Gaussian Splatting with Surface Continuity Priors

Bing He (Shanghai Jiao Tong University), Wenjun Zhang (Shanghai Jiao Tong University)

🎯 What it does: Construct a feed-forward 3D scene reconstruction model called SurfSplat using 2D Gaussian splatting, and enhance geometric coherence through surface continuity prior and forced alpha blending

SURGE: Surprise-Guided Token Reduction for Efficient Video Understanding with VLMs

Chong Tang (University of Southampton), Jagmohan Chauhan (University College London)

🎯 What it does: Propose SURGE, a training-agnostic and backend-agnostic surprise mask based on token prediction error, dynamically trimming redundant visual tokens before video VLM inference, significantly reducing computational cost.

SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis

Shahriar Noroozizadeh (Carnegie Mellon University), George H. Chen (Carnegie Mellon University)

🎯 What it does: Constructed the SURVHTE-BENCH benchmark to unify the evaluation of heterogeneous treatment effect estimation methods in right-censored survival data.

SUSD: Structured Unsupervised Skill Discovery through State Factorization

Seyed Mohammad Hadi Hosseini (Sharif University of Technology), Mahdieh Soleymani Baghshah (Sharif University of Technology)

🎯 What it does: This paper proposes a structured unsupervised skill discovery framework called SUSD, which decomposes environmental states into multiple factors and assigns independent skill variables to each factor, enabling the learning of rich and decomposable skills in reward-free environments.

SVD Provably Denoises Nearest Neighbor Data

Ravindran Kannan (UC Berkeley), David Woodruff (Carnegie Mellon University)

🎯 What it does: Under a high-dimensional noise model, we propose an algorithm that only uses SVD dimensionality reduction to recover the nearest neighbor of the original data from noisy data, provided the noise level satisfies σ=O(1/k^{1/4}).

Swap-guided Preference Learning for Personalized Reinforcement Learning from Human Feedback

Gihoon Kim (Yonsei University), Euntai Kim (Korea Institute of Science and Technology)

🎯 What it does: Proposed a framework named Swap-guided Preference Learning (SPL) to learn personalized reward functions from human preference data and address the posterior collapse problem in VPL.

SWE-RM: Execution-free Feedback for Software Engineering Agents

KaShun SHUM, Junxian He (Hong Kong University of Science and Technology)

🎯 What it does: Developed an executive-free reward model SWE-RM that performs well in both Test Time Scaling (TTS) and Reinforcement Learning (RL) scenarios.

SWERank: Software Issue Localization with Code Ranking

Revanth Gangi Reddy (University of Illinois at Urbana-Champaign), Shafiq Joty (Salesforce Research)

🎯 What it does: Propose the SWERANK framework, treating software problem localization as a specialized retrieval task, combining a retriever and re-ranker for efficient localization, and constructing the SWELOC dataset for training.

SwiftTS: A Swift Selection Framework for Time Series Pre-trained Models via Multi-task Meta-Learning

Tengxue Zhang (East China Normal University), Bin Yang (East China Normal University)

🎯 What it does: Proposes SwiftTS, a fast selection framework for time-series pre-trained models based on dual encoders and multi-task meta-learning, designed to efficiently select the most suitable model from a model library for a given time-series prediction task.

SWINGARENA: Adversarial Programming Arena for Long-context GitHub Issue Solving

Wendong XU, Ngai Wong (University of Hong Kong)

🎯 What it does: Propose SWINGARENA, an adversarial evaluation framework based on real CI pipelines, simulating interactions between submitters and reviewers in program patch and test generation;

SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs

Dachuan Shi (Georgia Tech), Wen Xiao (Microsoft)

🎯 What it does: Proposed a training-agnostic reasoning framework called SWIREASONING, which dynamically switches between explicit chain-of-thought reasoning and implicit continuous-space reasoning to improve the accuracy and token efficiency of large language models across various reasoning tasks.

Symmetric Space Learning for Combinatorial Generalization

Jaehyoung Jeong (Gwangju Institute of Science and Technology), Kangil Kim (Gwangju Institute of Science and Technology)

🎯 What it does: Propose a framework called CartanFM that utilizes symmetric space structure (Symmetric Space) for combinatorial generalization (Combinatorial Generalization), and learns and generalizes Lie algebra structure through Cartan Loss and Geodesic Symmetry Consistency Loss.

Symmetry-Aware Bayesian Optimization via Max Kernels

Anthony Bardou (EPFL), Patrick Thiran (EPFL)

🎯 What it does: Propose a symmetry-aware Bayesian optimization framework based on the maximum similarity (max kernel), obtaining a kernel k(D)+ applicable to Gaussian processes through PSD projection and Nyström extension, and conducting experiments on multiple black-box functions with symmetries such as rotation, reflection, and permutation.

SYNC: Measuring and Advancing Synthesizability in Structure-Based Drug Design

Yunfan Liu (Zhejiang University), Stan Z. Li (Westlake University)

🎯 What it does: Proposed a 3D structure-based SE(3)-invariant synthetizability classifier, SYNC, to evaluate and drive structure-based drug design (SBDD) for generating synthetizable drugs.

Synchronizing Probabilities in Model-Driven Lossless Compression

Aviv Adler (Analog Devices, Inc.), Jennifer Tang (College of Holy Cross)

🎯 What it does: Proposed and implemented the PMATIC (Probability Matching Interval Coding) algorithm, enabling lossless compression based on LLMs to maintain correct decoding even in cases of prediction mismatch, along with theoretical proofs and experimental validations.

SynCoGen: Synthesizable 3D Molecule Generation via Joint Reaction and Coordinate Modeling

Andrei Rekesh (University of Toronto), Cheng-Hao Liu (Mila - Quebec AI Institute)

🎯 What it does: Propose a multimodal generative model named SYNCOGEN, capable of simultaneously generating synthetically feasible 3D molecular structures and their synthesis pathways.

Syncphony: Synchronized Audio-to-Video Generation with Diffusion Transformers

Jibin Song (Yonsei University), Youngjung Uh (Yonsei University)

🎯 What it does: Developed a audio-synchronized video generation framework called Syncphony, which can generate highly synchronized and high-quality dynamic content in 380×640, 24fps, 5-second videos based on audio, text, and image conditions.

SyncTrack: Rhythmic Stability and Synchronization in Multi-Track Music Generation

Hongrui Wang (Hong Kong University of Science and Technology), Yang Wang (University of Hong Kong)

🎯 What it does: Proposed the SyncTrack model for synchronous multi-track waveform music generation, focusing on improving rhythmic stability and cross-track synchronization.

Synergizing Understanding and Generation with Interleaved Analyzing-Drafting Thinking

Shengqiong Wu (National University of Singapore), Tat-Seng Chua (National University of Singapore)

🎯 What it does: Propose an alternating analysis-draft (AD-Loop) loop, enabling a unified vision-language model (UVLM) to achieve true synergy between understanding and generation tasks; through a two-phase training (supervised learning + reinforcement learning) framework, the model can dynamically decide during inference whether to alternate between text-based and vision-based thinking.

Synthesising Counterfactual Explanations via Label-Conditional Gaussian Mixture Variational Autoencoders

Junqi Jiang (Imperial College London), Francesca Toni (Imperial College London)

🎯 What it does: This paper proposes Label-conditional Gaussian Mixture VAE (L-GMVAE) and the path-based counterfactual generation algorithm LAPACE, which generates a series of actionable counterfactual explanations for given inputs while satisfying effectiveness, proximity, interpretability, diversity, and robustness to input and model variations.

Synthesizing High-Quality Visual Question Answering from Medical Documents with Generator-Verifier LMMs

Xiaoke Huang (UC Santa Cruz), Yuyin Zhou (Amazon Research)

🎯 What it does: This study synthesizes high-quality medical multimodal question-answering (VQA) data from public biomedical literature through a generator-verifier framework.

Synthetic Bootstrapped Pretraining

Zitong Yang (Apple), Ruoming Pang (Apple)

🎯 What it does: Proposed a new language model pretraining method called Synthetic Bootstrapping Pretraining (SBP), which synthesizes new corpora by learning relationships between documents to improve model performance.

Synthetic History: Evaluating Visual Representations of the Past in Diffusion Models

Maria-Teresa De Rosa Palmini (University of Zurich), Eva Cetinic (University of Zurich)

🎯 What it does: We created the HistVis benchmark, using 30k synthetic images to evaluate the performance of text-to-image diffusion models in historical contexts, including style preferences, temporal consistency, and demographic distribution.

SynthWorlds: Controlled Parallel Worlds for Disentangling Reasoning and Knowledge in Language Models

Ken Gu (University of Washington), Tim Althoff (Google Research)

🎯 What it does: Constructed a controlled parallel corpus and tasks to separate the impact of language models' reasoning ability from parameterized knowledge.

Sysformer: Safeguarding Frozen Large Language Models with Adaptive System Prompts

Kartik Sharma (Georgia Institute of Technology), Srijan Kumar (Georgia Institute of Technology)

🎯 What it does: Without updating LLM parameters, a learnable Transformer module dynamically adjusts system prompts to enhance the security of frozen LLMs.

SysMoBench: Evaluating AI on Formally Specifying Complex Real-World Systems

Qian Cheng (Nanjing University), Tianyin Xu (University of Illinois Urbana-Champaign)

🎯 What it does: This paper proposes the SYSMOBENCH benchmark to evaluate the capability of generative AI in formally modeling large-scale, complex systems (such as distributed protocols and concurrent components);

t-SNE Exaggerates Clusters, Provably

Noah Bergam (Columbia University), Nakul Verma (Columbia University)

🎯 What it does: This paper demonstrates the limitations and failure modes of t-SNE in clustering significance and outlier detection through theoretical proofs and experiments;

T-TAMER: Provably Taming Trade-offs in ML Serving

Yuanyuan Chloe Yang (University of Washington), Haifeng Xu (University of Chicago)

🎯 What it does: This paper proposes a framework named T-Tamer for uniformly handling the trade-off between accuracy and latency/resource usage in multi-model inference (e.g., early stopping and cascaded models), and provides a dynamic indexing strategy that can be solved on various DAG structures (linear, tree, closure).

T1: One-to-One Channel-Head Binding for Multivariate Time-Series Imputation

Dongik Park (Seoul National University), Hyung-Sin Kim (Seoul National University)

🎯 What it does: Propose a CNN-Transformer hybrid architecture T1 for missing value imputation in multivariate time series;

T1: Tool-integrated Verification for Test-time Compute Scaling in Small Language Models

Minki Kang (KAIST), Jaewoong Cho (KRAFTON)

🎯 What it does: Proposes a two-phase framework called Tool-integrated Verification (T1), which first filters generated candidate answers using external tools and then verifies them with a small language model to enhance the inference performance of small models under computational scaling (best-of-N) during testing.

Tab-MIA: A Benchmark Dataset for Membership Inference Attacks on Tabular Data in LLMs

Eyal German (Ben-Gurion University of the Negev), Yuval Elovici (Ben-Gurion University of the Negev)

🎯 What it does: Proposes the Tab-MIA benchmark to assess the vulnerability of large language models (LLMs) to membership inference (MIA) attacks after training on tabular data.

TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding

Xiaobo Xing (University of Queensland), Hongzhi Yin (University of Queensland)

🎯 What it does: Proposed a dynamic self-adaptive multimodal routing framework called TableDART, which selects the optimal inference path instance-by-instance in table understanding tasks using single-modal models (text and image) and a lightweight gated network, achieving cross-modal fusion through an LLM agent.

TableMaster: A Recipe to Advance Table Understanding with Language Models

Lang Cao (University of Illinois Urbana-Champaign), Hanbing Liu (Tsinghua University)

🎯 What it does: Propose the TableMaster framework, which unifies the integration of table structure extraction, table semantic enrichment, and adaptive reasoning strategies to enhance large language models' understanding and reasoning capabilities with tables.

TABLET: A Large-Scale Dataset for Robust Visual Table Understanding

Iñigo Alonso (University of Edinburgh), Mirella Lapata (University of Basque Country UPV/EHU)

🎯 What it does: Developed a large-scale visual table understanding dataset called TABLET, containing over 4 million samples, 21 tasks, and retaining the original visual presentation.

TabStruct: Measuring Structural Fidelity of Tabular Data

Xiangjian Jiang (University of Cambridge), Mateja Jamnik (University of Cambridge)

🎯 What it does: Introduce the TabStruct benchmark to systematically evaluate the performance of table generators in terms of structural fidelity and traditional dimensions

Tackling Heavy-Tailed Q-Value Bias in Offline-to-Online Reinforcement Learning with Laplace-Robust Modeling

Ruibo Guo (University Of Science And Technology Of China), Bin Li (China Mobile)

🎯 What it does: To address the Q-value estimation bias in offline-to-online reinforcement learning, we propose an algorithm called LAROO, which robustly estimates Q-values through Laplace noise modeling and conservative ensemble, combined with offline pre-training and online fine-tuning.

Tackling the XAI Disagreement Problem with Adaptive Feature Grouping

Gabriel Laberge (Thales cortAIx Lab), Ola Ahmad (Thales cortAIx Lab)

🎯 What it does: Proposes the AGREED algorithm, which reduces disagreements among XAI explanation methods through adaptive feature grouping.

Tackling Time-Series Forecasting Generalization via Mitigating Concept Drift

Zhiyuan Zhao (Georgia Institute of Technology), B. Aditya Prakash (Georgia Institute of Technology)

🎯 What it does: Propose a model-agnostic ShifTS framework that extracts invariant patterns from historical and future exogenous features through a soft attention mask (SAM), addressing concept drift and time drift in time series prediction.

TaCo: A Benchmark for Lossless and Lossy Codecs of Heterogeneous Tactile Data

Zhengxue Cheng (Shanghai Jiao Tong University), Li Song (Shanghai Jiao Tong University)

🎯 What it does: Constructed the TaCo benchmark for systematic evaluation of various lossless and lossy tactile data compression algorithms.

Tactic: Adaptive Sparse Attention with Clustering and Distribution Fitting for Long-Context LLMs

Kan Zhu (University of Washington), Baris Kasikci (University of Washington)

🎯 What it does: Propose an adaptive sparse attention mechanism called Tactic, which dynamically selects key KV pairs based on cumulative attention scores for long-context LLM inference;

Take Note: Your Molecular Dataset Is Probably Aligned

Peter Lippmann (Heidelberg University), Fred A. Hamprecht (Heidelberg University)

🎯 What it does: This study systematically reveals the pose bias present in popular molecular datasets (QM9, QMugs, OMol25), proposes quantitative evaluation and visualization methods, and verifies through classifier and property regression experiments that models can exploit this bias to achieve unreasonably high performance.

Talk, Evaluate, Diagnose: User-aware Agent Evaluation with Automated Error Analysis

Penny Chong (SAP), Daniel Dahlmeier (SAP)

🎯 What it does: Proposes the TED (Talk, Evaluate, Diagnose) framework for systematically assessing the dialogue performance, progress, and error analysis of LLM agents in multi-task scenarios.

Talking Points: Describing and Localizing Pixels

Matan Rusanovsky (Tel Aviv University), Shai Avidan (Tel Aviv University)

Taming Curvature: Architecture Warm-up for Stable Transformer Training

Sameera Ramasinghe (Pluralis Research), Alexander Long (Pluralis Research)

🎯 What it does: The paper proposes two techniques, online curvature estimation and architectural warm-up, to improve the stability of large-scale Transformer training.

Taming Hierarchical Image Coding Optimization: A Spectral Regularization Perspective

Wuyang Cong (Nanjing University), Zhan Ma (Nanjing University)

🎯 What it does: Propose a spectral regularization method for hierarchical image compression, addressing cross-scale energy diffusion and aliasing issues through internal-scale spectral truncation and cross-scale similarity constraints, thereby improving training efficiency and compression performance.

Taming Imperfect Process Verifiers: A Sampling Perspective on Backtracking

Dhruv Rohatgi (MIT), Dylan J Foster

🎯 What it does: Proposes a new process-guided test-time sampling algorithm called VGB, which utilizes theoretically grounded backtracking techniques to enhance robustness against validator errors.

Taming Momentum: Rethinking Optimizer States Through Low-Rank Approximation

Zhengbo Wang (University of Science and Technology of China), Tieniu Tan (NLPR & MAIS, Institute of Automation, Chinese Academy of Sciences)

🎯 What it does: Propose LoRA-Pre, a low-rank optimizer that significantly reduces memory usage by compressing the momentum states of optimizers such as Adam and Muon, treating EMA momentum as online linear regression.

Taming Polysemanticity in LLMs: Theory-Grounded Feature Recovery via Sparse Autoencoders

Siyu Chen (Yale University), Zhuoran Yang (Yale University)

🎯 What it does: Proposed a feature recovery method using sparse autoencoders (SAE) with theoretical guarantees for LLM interpretability tasks.

Taming Score-Based Denoisers in ADMM: A Convergent Plug-and-Play Framework

Rajesh Shrestha (Oregon State University), Xiao Fu (Oregon State University)

🎯 What it does: Proposed an AC-DC three-stage denoiser embedded in the ADMM-PnP framework to address noise data manifold mismatch and convergence issues.

TAMMs: Change Understanding and Forecasting in Satellite Image Time Series with Temporal-Aware Multimodal Models

Zhongbin Guo (Beijing Institute of Technology), Ertai E (National University of Singapore)

🎯 What it does: Propose a unified framework TAMMs that jointly accomplishes the tasks of change description and future prediction for satellite image time series.

TandemFoilSet: Datasets for Flow Field Prediction of Tandem-Airfoil Through the Reuse of Single Airfoils

Wei Xian Lim (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)

🎯 What it does: Constructed the TandemFoilSet dataset and developed a training and inference framework for multi-body flow field prediction.

TangleScore: Tangle-Guided Purge and Imprint for Unstructured Knowledge Editing

Hao-Xiang Xu (University of Science and Technology of China), Jia-Chen Gu (University of California, Los Angeles)

🎯 What it does: Research and improve knowledge editing methods for large language models, proposing TANGLESCORE to measure the intrinsic editing difficulty and adaptively adjusting the PIPE framework's purge and imprint based on this metric.

TangoFlux: Super Fast and Faithful Text to Audio Generation with Flow Matching and Clap-Ranked Preference Optimization

Chia-Yu Hung (Nanyang Technology University), Soujanya Poria (Nanyang Technology University)

🎯 What it does: Developed a text-to-audio generation model called TANGOFLUX based on rectified flow, supporting high-quality audio generation within 30 seconds, and achieving dynamic alignment through CLAP-Ranked Preference Optimization (CRPO).

TAO-Attack: Toward Advanced Optimization-Based Jailbreak Attacks for Large Language Models

Zhi Xu (Dalian University of Technology), Han Liu (Dalian University of Technology)

🎯 What it does: Proposes an optimized jailbreak attack method called TAO-Attack, capable of generating harmful text that can bypass safety filters in large language models.

TAPTRv3: Spatial and Temporal Context Foster Robust Tracking of Any Point in Long Video

Jinyuan Qu (Tsinghua University South China University of Technology), Lei Zhang (International Digital Economy Academy)

🎯 What it does: Propose TAPTRv3, which improves TAPTR by utilizing Context-aware Cross-Attention and Visibility-aware Long-Temporal Attention to enhance arbitrary point tracking in long videos;

Target-Aware Video Diffusion Models

Taeksoo Kim (Seoul National University), Hanbyul Joo (Seoul National University)

🎯 What it does: Developed a target-aware video diffusion model that generates videos of actors interacting with specified targets using input images, target segmentation masks, and text-based action descriptions.

Task Tokens: A Flexible Approach to Adapting Behavior Foundation Models

Ron Vainshtein (Technion), Chen Tessler (NVIDIA Research)

🎯 What it does: Propose the Task Tokens method, using a lightweight task encoder to achieve task-specific adaptation on behavior foundation models (BFCM) like MaskedMimic, balancing user prompts and reward-driven control;

Task Vectors, Learned Not Extracted: Performance Gains and Mechanistic Insights

Haolin Yang, Naoya Inoue

🎯 What it does: This work proposes directly training the learning task vector (LTV) and achieving in-context learning in large language models by injecting vectors at any layer or position; meanwhile, it systematically analyzes the impact of LTV on model outputs through attention head OV circuits at lower layers and approximate linear rotation-scaling mechanisms at higher layers.

Task-Adaptive Parameter-Efficient Fine-Tuning for Weather Foundation Models

Shilei Cao (Sun Yat-sen University), Haohuan Fu (National Supercomputing Center)

🎯 What it does: Propose WeatherPEFT, a parameter-efficient fine-tuning framework for weather foundation models, integrating task-adaptive dynamic prompting (TADP) and random Fisher information-guided parameter selection (SFAS).

Task-Aware Data Selection via Proxy-Label Enhanced Distribution Matching for LLM Finetuning

Hao Cheng (Hong Kong Baptist University), Bo Han (D5 Data)

🎯 What it does: Propose a task-aware data selection framework based on LLM-generated proxy labels, two-stage filtering, and incremental sampling, enhancing LLM fine-tuning performance through joint distribution matching.

Task-free Adaptive Meta Black-box Optimization

Chao Wang (Xidian University), Shuyuan Yang (Xidian University)

🎯 What it does: Propose a task-agnostic adaptive meta-black-box optimizer ABOM, which can online learn and adaptively update evolutionary operators using only the optimization data of the target task, achieving zero-shot optimization.

Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective

Lei Lei (University of Science and Technology of China), Tong Xu (University of Science and Technology of China)

🎯 What it does: Assess the importance of visual tokens using explainability methods and perform task-related visual token compression during the LLM input phase, significantly reducing computational and memory costs;

TaskCraft: Automated Generation of Agentic Tasks

Dingfeng Shi (OPPO), Wangchunshu Zhou (OPPO)

🎯 What it does: Proposes TASKCRAFT, a complete workflow for automatically generating multi-tool, verifiable, and scalable agentic tasks, progressively increasing task difficulty through depth and width expansion while maintaining quality via incremental verification.

TASTE: Text-Aligned Speech Tokenization and Embedding for Spoken Language Modeling

Liang-Hsuan Tseng (MediaTek Research), Hung-yi Lee (National Taiwan University)

🎯 What it does: Proposed TASTE, a text-aligned speech segmentation and embedding method for constructing a joint text-speech speaking language model;

TaTToo: Tool-Grounded Thinking PRM for Test-Time Scaling in Tabular Reasoning

Jiaru Zou (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)

🎯 What it does: Proposed TATTOO, a tool-assisted process reward model specifically designed for table reasoning, which provides step-based reward supervision during the reasoning process to enhance the reasoning quality and efficiency of large models in table reasoning tasks.

TAVAE: A VAE with Adaptable Priors Explains Contextual Modulation in the Visual Cortex

Balázs Meszéna (HUN-REN Wigner Research Centre for Physics), Gergo Orban (HUN-REN Wigner Research Centre for Physics)

🎯 What it does: This paper constructs a Task-Amortized Variational Autoencoder (TAVAE) to explain the context modulation and multimodal response deviations in the mouse primary visual cortex (V1) during visual Go-NoGo discrimination tasks, while maintaining a pre-trained generative model and variational posterior for natural images and learning task-specific priors.

TCD-Arena: Assessing Robustness of Time Series Causal Discovery Methods Against Assumption Violations

Gideon Stein (Friedrich Schiller University Jena), Joachim Denzler (Friedrich Schiller University Jena)

🎯 What it does: Proposed the TCD-Arena testing tool, systematically evaluating the robustness of over 300,000 time series causal discovery algorithms under 33 progressively intensified hypothesis violations.

TD-JEPA: Latent-predictive Representations for Zero-Shot Reinforcement Learning

Marco Bagatella (ETH Zurich), Andrea Tirinzoni (FAIR at Meta)

🎯 What it does: Propose TD-JEPA, a potential prediction method based on multi-step temporal difference (TD) learning, for learning zero-shot unsupervised reinforcement learning policies directly applicable to arbitrary rewards from offline reward-free data.

TD-MoE: Tensor Decomposition for MoE Models

Yuebin XU, Zeyi Wen

🎯 What it does: Propose a Mixture-of-Experts compression method TD-MoE based on three-dimensional tensor decomposition

Teach to Reason Safely: Policy-Guided Safety Tuning for MLRMs

Jingyu Zhang, Zhaohui Yang (Ant Group)

🎯 What it does: This paper proposes a two-stage safety optimization framework, Policy-Guided Safety Tuning (PST), which significantly enhances the safety of multimodal reasoning models while maintaining their reasoning capabilities through policy-guided supervised fine-tuning and safe reasoning preference optimization.

Teach2Eval: An Interaction-Driven LLMs Evaluation Method via Teaching Effectiveness

Yuhang Zhou (Fudan University), Hongfeng Chai (Fudan University)

🎯 What it does: Propose Teach2Eval—an interactive method for evaluating LLMs by guiding weak students to improve through a teacher model

Teaching Metric Distance to Discrete Autoregressive Language Models

Jiwan Chung (Yonsei University), Youngjae Yu (Seoul National University)

🎯 What it does: Proposed a new distance-aware target, DIST 2 Loss, for discrete autoregressive language models, aiming to replace traditional one-hot targets by leveraging predefined token distances to improve model performance on numerical, spatial coordinate, and similar tasks.