International Conference on Learning Representations · 2207 papers
Speculative Actions: A Lossless Framework for Faster AI Agents
Naimeng Ye (Columbia University), Tianyi Peng (Columbia University)
CodeComputational EfficiencyLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Proposed a 'Speculative Actions' framework that significantly reduces overall latency while maintaining lossless performance by using a fast model to predict future API calls in the environment interaction loop of AI agents and executing these predictions in parallel.
CodeRecognitionExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelWorld ModelAudio
🎯 What it does: Propose a speech world model (SWM) based on a causal graph, decomposing speech understanding into four modules: context activation (WMA), theory of mind (ToM), pragmatic actions (SA), and goals (Prag), and achieving explicit reasoning through causal relationships.
Speech-to-LaTeX: New Models and Datasets for Converting Spoken Equations and Sentences
Dmitrii Korzh (AXXX), Ivan Oseledets (AXXX)
CodeRecognitionData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmarkAudio
🎯 What it does: Developed and released a large-scale bilingual (English, Russian) spoken math dataset (S2L), and proposed multiple speech-to-LaTeX methods covering two tasks: standalone equations and embedded sentences;
SpeechJudge: Towards Human-Level Judgment for Speech Naturalness
Xueyao Zhang (Chinese University of Hong Kong Shenzhen), Zhizheng Wu (Chinese University of Hong Kong Shenzhen)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningBenchmarkChain-of-ThoughtAudio
🎯 What it does: Developed the SpeechJudge series, which includes a large-scale naturalness preference dataset, evaluation benchmarks, and generation-based reward models.
CodeTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Developed the SPELL framework, a multi-agent self-play reinforcement learning system that enhances long-context reasoning capabilities through a three-role cycle of question-answering, answering, and verification using a single LLM without human-labeled data.
🎯 What it does: Proposed a reinforcement learning algorithm called Sandwiched Policy Gradient (SPG) for post-training Masked Diffusion Language Models (MDLM);
🎯 What it does: Proposed a spiking-based digital brain model called Spike-DB, which encodes fMRI time series into spike sequences using IIR filter spiking neurons, and achieves high-precision modeling of brain activity and downstream tasks through self-supervised Anchor-Target prediction;
🎯 What it does: Proposed and implemented SpikeGen, a dual-modal visual processing framework based on latent generative models, capable of fusing RGB and spiking (event) data to achieve defogging, frame reconstruction, and high-frequency perspective synthesis.
🎯 What it does: Proposed SpikeStereoNet, which can directly estimate stereo depth from raw pulse streams of two spike cameras, and created a large-scale synthetic and real-world pulse stream depth dataset.
SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning
Bo Liu (National University Of Singapore), Natasha Jaques
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: In the paper, the authors propose the SPIRAL framework, which trains large language models (LLMs) through self-play in two-player zero-sum language games, thereby enhancing reasoning capabilities without requiring human-annotated rewards.
SplitLoRA: Balancing Stability and Plasticity in Continual Learning Through Gradient Space Splitting
Haomiao Qiu (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
CodeMeta LearningTransformerImageBenchmark
🎯 What it does: Propose the SplitLoRA method, which balances stability and plasticity by leveraging low-rank adapters and gradient subspace partitioning to achieve continuous learning.
🎯 What it does: Created SportR, a cross-sport multimodal question-answering benchmark containing 4,789 images and 2,052 videos, along with 6,841 human-annotated chain-of-thought (CoT) examples serving as training and evaluation standards.
Spotlight on Token Perception for Multimodal Reinforcement Learning
Siyuan Huang (Shanghai AI Laboratory), Yu Cheng (Nanjing University)
CodeReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Propose the VPPO algorithm by integrating visual perception into multi-modal reinforcement learning, achieving hierarchical optimization of vision-dependent tokens to enhance the reasoning performance of large vision-language models.
🎯 What it does: Propose a multi-objective optimization framework SPREAD based on diffusion models, which can generate and refine approximate Pareto front points in the decision space.
SPRIG: Improving Large Language Model Performance by System Prompt Optimization
Lechen Zhang (University of Illinois Urbana-Champaign), David Jurgens (University of Michigan)
CodeOptimizationLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Developed the SPRIG system prompt optimization framework, which iteratively constructs system prompts using genetic algorithms to enhance the performance of large language models across multiple tasks, languages, and scales.
🎯 What it does: Proposed a SCER method that regularizes feature representations in the embedding space to suppress domain-related pseudo-correlated features, thereby improving the worst-group performance under subpopulation distribution shifts.
🎯 What it does: Propose a sparse partial weakly supervised oriented object detection framework named SPWOOD, which can achieve high-precision detection using only a small number of sparse weak labels and a large amount of unlabeled data
CodeOptimizationDrug DiscoveryTransformerLarge Language ModelReinforcement LearningAgentic AITextTabularBenchmarkPhysics Related
🎯 What it does: Proposes the SR-SCIENTIST framework, enabling large language models (LLMs) to evolve from mere equation generators into autonomous scientists capable of writing code, analyzing data, evaluating results, and self-optimizing through multi-round interactions to accomplish symbolic regression tasks.
SSD-GS: Scattering and Shadow Decomposition for Relightable 3D Gaussian Splatting
Iris Zheng (Victoria University of Wellington), Fang-Lue Zhang (Victoria University of Wellington)
CodeGenerationGaussian SplattingImagePoint CloudPhysics Related
🎯 What it does: Propose SSD-GS on the 3D Gaussian Splatting framework, enabling high-quality reconstruction and relighting under new lighting conditions.
SSDi8: Accurate and Efficient 8-bit Quantization for State Space Duality
Hyunwoo Kim (Chung-Ang University), Dahuin Jung (Chung-Ang University)
CodeComputational EfficiencyText
🎯 What it does: Proposes SSDi8, a post-training INT8 quantization framework for Mamba-2 structured state space dual-mode (SSD), enabling continuous INT8 computation paths with significant latency reduction while maintaining accuracy close to FP16.
SSG: Scaled Spatial Guidance for Multi-Scale Visual Autoregressive Generation
Youngwoo Shin (Korea Advanced Institute of Science and Technology), Junmo Kim (Korea Advanced Institute of Science and Technology)
CodeGenerationImageText
🎯 What it does: Propose a training-agnostic Scaled Spatial Guidance (SSG) that leverages frequency domain priors to highlight semantic residuals at each step of the VAR model, thereby enhancing high-frequency details and overall coherence between coarse and fine-level generation.
CodeComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A framework called ssToken for token-level data selection in LLM supervised fine-tuning is studied, combining self-modulation and semantic-aware signals to filter useless tokens.
ST-HHOL: Spatio-Temporal Hierarchical Hypergraph Online Learning for Crime Prediction
Keqing Du (Xi'an Jiaotong University), Wei Shao (Data61)
CodeExplainability and InterpretabilityGraph Neural NetworkTransformerSupervised Fine-TuningGraphTabularTime Series
🎯 What it does: Proposed an online spatiotemporal learning framework ST-HOL for urban streaming crime prediction, integrating hierarchical hypergraph convolutional networks with partially frozen LLMs, capable of capturing high-order crime patterns and adapting to concept drift.
ST-SimDiff: Balancing Spatiotemporal Similarity and Difference for Efficient Video Understanding with MLLMs
Bingjun Luo (Tsinghua University), Xinpeng Ding (Xidian University)
CodeCompressionComputational EfficiencyGraph Neural NetworkVision Language ModelVideo
🎯 What it does: Propose ST-SimDiff, a no-training visual token compression framework based on spatial-temporal graphs, which preserves key video information through dual selection based on similarity and difference;
Stabilizing Policy Gradients for Sample-Efficient Reinforcement Learning in LLM Reasoning
Luckeciano Carvalho Melo (University of Oxford), Yarin Gal (University of Oxford)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Proposes the Curvature-Aware Policy Optimization (CAPO) method, which improves sample efficiency in reinforcement learning for large language model inference tasks by monitoring and limiting instability caused by gradients and curvature.
🎯 What it does: Propose the Stable-LoRA weight decay strategy to address the instability of feature learning during LoRA training, and verify its effectiveness theoretically and experimentally.
StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs
Yuhan Song (Peking University), Zhou Xiao
CodeRepresentation LearningAudio
🎯 What it does: Propose a speech tokenizer called StableToken based on a multi-branch voting mechanism to improve tokenization stability in noisy environments
Stacked from One: Multi-Scale Self-Injection for Context Window Extension
Wei Han (Singapore University of Technology and Design), Shuicheng YAN
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the SHAREDLLM framework, which expands the LLM context window to 128K tokens by sharing KV between a lower-level compressor and an upper-level decoder.
Stackelberg Coupling of Online Representation Learning and Reinforcement Learning
Fernando Martinez (Fordham University), Juntao Chen (Fordham University)
CodeReinforcement Learning
🎯 What it does: Proposed and implemented the SCORER framework, transforming representation learning and value function learning in deep Q learning into a Stackelberg leader-follower game, achieving stable co-adaptation through two-time-scale updates.
🎯 What it does: Conduct a systematic theoretical analysis of classifier-free guidance (CFG) in diffusion models under multimodal conditional distributions, revealing that the sampling dynamics consist of three stages: directional offset, mode separation, and concentration, while explaining the fundamental mechanisms behind diversity degradation, followed by the design and validation of a time-varying guidance strength scheduling scheme.
🎯 What it does: Propose STAIRS-Former, a spatial-temporal hierarchical Transformer structure for offline multi-task multi-agent reinforcement learning, capable of adapting to varying numbers of agents and leveraging historical information;
🎯 What it does: This paper proposes the STAR framework, achieving distillation and fine-tuning of function calling capabilities through a teacher-guided ultra-small model.
Yinghui He (Princeton University), Sanjeev Arora (Princeton University)
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Developed an adaptive training framework named STAT based on skill gaps, leveraging the metacognitive capabilities of state-of-the-art large language models to analyze skill deficiencies in student models on validation sets. The framework enhances model performance in mathematical reasoning tasks by reweighting existing training samples or synthesizing problems related to missing skills.
Statistical Guarantees in the Search for Less Discriminatory Algorithms
Chris Hays, Manish Raghavan
CodeClassificationTabular
🎯 What it does: Propose an adaptive stopping algorithm to help enterprises prove that they have sufficiently searched for a less discriminatory algorithm (LDA) during model retraining under U.S. anti-discrimination laws.
🎯 What it does: Propose the STEDiff framework, which includes an efficient attack module STEBA and a real-time detection module STEDF, leveraging spatial and temporal redundancy in diffusion models to achieve low-cost backdoor injection and detection.
🎯 What it does: Propose a gradient-free, model-agnostic concept contrastive corrector named CO3 to enhance semantic alignment and concept coverage in multi-concept text-to-image diffusion models.
Steering Autoregressive Music Generation with Recursive Feature Machines
Daniel Zhao (University of California, San Diego), Zachary Novack (University of California, San Diego)
CodeGenerationData SynthesisExplainability and InterpretabilityTransformerAudio
🎯 What it does: By migrating Recursive Feature Machines (RFM) to the audio domain, injecting interpretable musical concepts into the frozen MUSICGEN-Large model, achieving real-time, fine-grained controllable music generation.
🎯 What it does: Proposed the Disco model, achieving trustworthy content recommendation by decoupling diffusion models and trustworthy subspace projection.
Constanza Fierro (University of Copenhagen), Fabien Roger (Anthropic)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Propose a post-training contrastive weight steering method that controls the behavior of large language models by leveraging weight differences obtained from fine-tuning on positive and negative behavior data;
Mohsen Fayyaz (University of California, Los Angeles), Nanyun Peng (University of California, Los Angeles)
CodeSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: By comparing the differences in expert activation rates within adversarial behavior pairs, experts closely associated with specific behaviors (such as safety and authenticity) are detected. During inference, these experts are activated or suppressed by adjusting the router's logits, achieving weight-unadjusted control over the behavior of large language models (MoE).
Step-Aware Residual-Guided Diffusion for EEG Spatial Super-Resolution
Hongjun Liu (University of Science and Technology Beijing), Chao Yao (University of Science and Technology Beijing)
CodeSuper ResolutionDiffusion modelAuto EncoderBiomedical Data
🎯 What it does: Propose a dynamic residual-guided diffusion model, SRGDiff, which generates high-density EEG signals from low-density EEG signals to achieve spatial super-resolution.
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: Proposed and implemented the StepORLM framework, which enables joint training of policy models and generative process reward models (GenPRM) through a self-evolutionary loop, achieving dual supervision of both process and results for operations research (OR) problems, significantly enhancing the reasoning and modeling capabilities of large language models (LLMs) in the OR domain.
🎯 What it does: Re-examine Evidential Deep Learning (EDL) from an optimization perspective, revealing its implicit maximum margin characteristic, proposing the optimization consistency principle, and designing a new uncertainty measure MPU that is consistent with the UCE loss. Subsequently, validate its effectiveness in out-of-distribution (OOD) and misclassification detection tasks.
Stop Tracking Me! Proactive Defense Against Attribute Inference Attack in LLMs
Dong Yan (University of Chinese Academy of Sciences), Tieniu Tan (University of Chinese Academy of Sciences)
CodeSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Research and propose a unified defense framework TRACE-RPS, combining fine-grained anonymization (TRACE) with optimization-based rejection induction (RPS), for actively defending against attribute inference attacks in large language models.
Stop Unnecessary Reflection: Training LRMs for Efficient Reasoning with Adaptive Reflection and Length Coordinated Penalty
Zewei Yu (Zhejiang University), Haobo Wang (Zhejiang University)
CodeComputational EfficiencyLarge Language ModelReinforcement LearningTextBenchmarkChain-of-Thought
🎯 What it does: A reinforcement learning framework with adaptive reflection and length synergistic penalties is used to train large reasoning models to reduce excessive reflection and verbose outputs during chain-of-thought generation, thereby improving reasoning efficiency and accuracy.
🎯 What it does: Propose a new training-free fast sampling method called STORK, which combines robust Runge-Kutta formulas with Taylor expansion to achieve high-quality sampling of noise prediction models and flow matching models with only a small number of NFEs.
StoryAlign: Evaluating and Training Reward Models for Story Generation
Haotian Xia (Tsinghua University), Juanzi Li (Tsinghua University)
CodeGenerationReinforcement Learning from Human FeedbackLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: This paper proposes the STORYRMB benchmark and trains the STORYREWARD reward model to enhance the quality of LLMs in story generation and improve alignment with human preferences.
Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs
Alexander Panfilov (ELLIS Institute Tübingen & MPI for Intelligent Systems), Jonas Geiping (ELLIS Institute Tübingen & MPI for Intelligent Systems)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper investigates the strategic dishonesty exhibited by cutting-edge large language models when faced with harmful requests, and verifies its universality and impact through multiple-choice questions (MCQ) experiments, output monitoring, and internal activation detection;
Strategic Planning and Rationalizing on Trees Make LLMs Better Debaters
Danqing Wang (Carnegie Mellon University), Lei Li (Carnegie Mellon University)
CodeReinforcement Learning from Human FeedbackLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed the TreeDebater system for competitive debate scenarios, utilizing a Rehearsal Tree to simulate potential opponent attacks and defenses, a Debate Flow Tree to track debate status in real-time, and combining speech duration control with simulated audience feedback to generate time-constrained, more persuasive speeches.
🎯 What it does: Propose a new streaming drag-based interactive video editing task REVEL, and develop the DragStream method to achieve drag editing and animation at any position and time during video generation without requiring additional training.
StreamingVLM: Real-Time Understanding for Infinite Video Streams
Ruyi Xu (MIT), Song Han (MIT)
CodeComputational EfficiencyTransformerSupervised Fine-TuningVision Language ModelVideoTextBenchmark
🎯 What it does: Proposed a unified training and inference framework called StreamingVLM, which can maintain low-latency and stable visual language understanding in real-time infinite video streams.
Strict Subgoal Execution: Reliable Long-Horizon Planning in Hierarchical Reinforcement Learning
Seungyul Han (Ulsan National Institute of Science and Technology), Jeongmo Kim (Ulsan National Institute of Science and Technology)
CodeReinforcement LearningTabular
🎯 What it does: Proposed the Strict Subgoal Execution (SSE) framework, which utilizes Frontier Experience Replay (FER) to enforce the completion of subgoals in a single step within graph-structured hierarchical reinforcement learning, significantly improving the learning efficiency of long-horizon tasks;
String Seed of Thought: Prompting LLMs for Distribution-Faithful and Diverse Generation
Kou Misaki (Sakana AI), Takuya Akiba (Sakana AI)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied a novel prompting method called String Seed of Thought (SSoT), enabling LLMs to select answers according to a given probability distribution and significantly enhance output diversity.
Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs
Yujie Zhao (University of California, San Diego), Jishen Zhao (University of California, San Diego)
CodeAI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Propose the AT-GRPO algorithm and training system, combining multi-agent systems (MAS) with reinforcement learning (RL) to enhance the collaborative and reasoning capabilities of large language models (LLMs) in tasks such as code, mathematics, planning, and games.
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)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerMixture of ExpertsMultimodalityBiomedical Data
🎯 What it does: Proposed a Slot-based structured prognosis event modeling framework, SlotSPE, for multi-modal (tissue slide + genomic) cancer survival analysis.
Chenglei Shen (Renmin University of China), Jun Xu (Renmin University of China)
CodeGenerationRepresentation LearningTransformerLarge Language ModelText
🎯 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)
CodeGenerationTransformerGaussian SplattingImage
🎯 What it does: Proposed a single-forward 3D Gaussian stylization framework named Stylos, which can instantly generate multi-view stylized scenes from unposed images
SuperMAN: Interpretable and Expressive Networks over Temporally Sparse Heterogeneous Data
Andrea Zerio (Aalborg University), Aleksejs Sazonovs (Aalborg University)
CodeExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkGraphTime SeriesBiomedical Data
🎯 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.
🎯 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.
SurvHTE-Bench: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis
Shahriar Noroozizadeh (Carnegie Mellon University), George H. Chen (Carnegie Mellon University)
CodeData SynthesisBiomedical DataBenchmark
🎯 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)
CodeRepresentation LearningReinforcement Learning
🎯 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.
Swap-guided Preference Learning for Personalized Reinforcement Learning from Human Feedback
Gihoon Kim (Yonsei University), Euntai Kim (Korea Institute of Science and Technology)
CodeReinforcement Learning from Human FeedbackReinforcement LearningFlow-based ModelAuto EncoderText
🎯 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.
SWERank: Software Issue Localization with Code Ranking
Revanth Gangi Reddy (University of Illinois at Urbana-Champaign), Shafiq Joty (Salesforce Research)
CodeAI Code AssistantTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
🎯 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)
CodeMeta LearningMixture of ExpertsTime Series
🎯 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.
🎯 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.
🎯 What it does: Propose a multimodal generative model named SYNCOGEN, capable of simultaneously generating synthetically feasible 3D molecular structures and their synthesis pathways.
Synthesizing High-Quality Visual Question Answering from Medical Documents with Generator-Verifier LMMs
Xiaoke Huang (UC Santa Cruz), Yuyin Zhou (Amazon Research)
CodeData SynthesisLarge Language ModelReinforcement LearningVision Language ModelMultimodalityBiomedical Data
🎯 What it does: This study synthesizes high-quality medical multimodal question-answering (VQA) data from public biomedical literature through a generator-verifier framework.
SynthWorlds: Controlled Parallel Worlds for Disentangling Reasoning and Knowledge in Language Models
Ken Gu (University of Washington), Tim Althoff (Google Research)
CodeGenerationData SynthesisRetrievalData-Centric LearningLarge Language ModelTextGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 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)
CodeSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 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)
CodeAI Code AssistantLarge Language ModelAgentic AITextBenchmark
🎯 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);
CodeOptimizationExplainability and InterpretabilityRepresentation LearningTextTabularBiomedical DataFinance Related
🎯 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;
TableDART: Dynamic Adaptive Multi-Modal Routing for Table Understanding
Xiaobo Xing (University of Queensland), Hongzhi Yin (University of Queensland)
CodeTransformerLarge Language ModelAgentic AIMixture of ExpertsMultimodalityTabular
🎯 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.
TABLET: A Large-Scale Dataset for Robust Visual Table Understanding
Iñigo Alonso (University of Edinburgh), Mirella Lapata (University of Basque Country UPV/EHU)
CodeRecognitionData-Centric LearningSupervised Fine-TuningVision Language ModelMultimodalityBenchmark
🎯 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.
🎯 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)
CodeReinforcement LearningBenchmark
🎯 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 Time-Series Forecasting Generalization via Mitigating Concept Drift
Zhiyuan Zhao (Georgia Institute of Technology), B. Aditya Prakash (Georgia Institute of Technology)
CodeDomain AdaptationTransformerTime Series
🎯 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.
Take Note: Your Molecular Dataset Is Probably Aligned
Peter Lippmann (Heidelberg University), Fred A. Hamprecht (Heidelberg University)
CodePose EstimationRepresentation LearningDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
🎯 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)
CodeExplainability and InterpretabilityLarge Language ModelAgentic AITextBenchmark
🎯 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.
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)
CodeOptimizationText
🎯 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.
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)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 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.
Task Vectors, Learned Not Extracted: Performance Gains and Mechanistic Insights
Haolin Yang, Naoya Inoue
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 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-Aware Data Selection via Proxy-Label Enhanced Distribution Matching for LLM Finetuning
Hao Cheng (Hong Kong Baptist University), Bo Han (D5 Data)
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 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.
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 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.
Teaching VLMs to Admit Uncertainty in OCR from Lossy Visual Inputs
Shuhao Guan (University College Dublin), Derek Greene (University College Dublin)
CodeRecognitionData SynthesisLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: This study proposes an uncertainty-aware OCR method, training visual-language models (VLMs) to insert <C></C> tags before and after potentially erroneous text segments when recognizing distorted documents, explicitly indicating the model's uncertainty.
Divyam Madaan (New York University), Kyunghyun Cho (New York University)
CodeSupervised Fine-TuningTextTime Series
🎯 What it does: Studied achieving temporal generalization using interpolation or extrapolation of past model parameters when future data is unavailable.
🎯 What it does: Proposed Temporal Graph Thumbnail (TGT), enhancing the robustness of representation learning by constructing a global evolutionary skeleton for temporal graphs.
🎯 What it does: This paper proposes a self-encouraging exploration method (C-TeC) that utilizes representations obtained through time contrastive learning. These representations serve as reward signals, encouraging agents to visit states where future states are unpredictable, thereby learning complex behaviors without external rewards.
🎯 What it does: Process 5 months of head-mounted video (Ego4D) with human-like visual experience, using state-of-the-art gaze prediction models to predict fixation points, then cropping the central visual region based on these predictions. Subsequently, train a temporal slow self-supervised learning model (MoCoV3) on these image sequences to learn slow-varying and semantically informative visual representations.
CodeExplainability and InterpretabilityAuto EncoderContrastive LearningText
🎯 What it does: Propose Temporal Sparse Autoencoders (T-SAE), improving the interpretability of sparse autoencoders in language models by introducing temporal contrastive loss on high-level features.
Temporal superposition and feature geometry of RNNs under memory demands
Pratyaksh Sharma (Imperial College London), Pedro A. M. Mediano
CodeRepresentation LearningRecurrent Neural NetworkTime Series
🎯 What it does: Studied the feature representation geometry of recurrent neural networks under temporal (memory) pressure, proposed and analyzed the concept of temporal superposition, revealing how the hidden space is compressed and forgotten.