International Conference on Learning Representations Β· 2207 papers
Tequila: Trapping-free Ternary Quantization for Large Language Models
Hong Huang (City University Of Hong Kong), Dapeng Wu (City University Of Hong Kong)
CodeComputational EfficiencyTransformerText
π― What it does: This paper proposes a ternary quantization method called Tequila for efficiently deploying large language models on edge devices, addressing the dead zone capture problem caused by traditional ternary quantization.
TerraFM: A Scalable Foundation Model for Unified Multisensor Earth Observation
Muhammad Sohail Danish (Mohamed bin Zayed University of Artificial Intelligence), Salman Khan (Mohamed bin Zayed University of Artificial Intelligence)
Test-Time Adaptation for LLM Agents via Environment Interaction
Arthur Chen (University of Waterloo), Caiming Xiong (Salesforce AI Research)
CodeDomain AdaptationTransformerLarge Language ModelAgentic AITextBenchmark
π― What it does: Propose two test-time adaptation methods: syntactic alignment (SA) and dynamic grounding (DG), to enhance the generalization ability of LLM agents in new environments.
π― What it does: Proposed a test-time domain generalization framework for image super-resolution, MC-TTDG, which utilizes multiple codebooks to achieve pixel-level feature transfer, addressing the low-resolution issues of traditional style transfer in low-level visual tasks.
π― What it does: Propose a general method for iterative error correction (IEC) during the testing phase of deployed efficient diffusion models, which can significantly improve generation quality;
Test-Time Matching: Unlocking Compositional Reasoning in Multimodal Models
Yinglun Zhu (University of California, Riverside), Fuzhi Tang (University of California, Riverside)
CodeRepresentation LearningMultimodality
π― What it does: Proposed a new evaluation metric called GroupMatch and improved the existing GroupScore, then designed a self-supervised iterative training method called Test-Time Matching (TTM) to enhance the performance of multi-modal models in compositional reasoning tasks.
CodeClassificationAnomaly DetectionGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringPoint Cloud
π― What it does: Proposes the PGLLM framework, leveraging 3D point cloud graph to construct neighborhood prompts and refining confidence through contextual guidance via LLM, to achieve classification, OOD detection, and description tasks on 3D point clouds;
TEST-TIME SCALING IN DIFFUSION LLMS VIA HIDDEN SEMI-AUTOREGRESSIVE EXPERTS
Jihoon Lee (Yonsei University), Amrit Singh Bedi (Carnegie Mellon University)
CodeComputational EfficiencyLarge Language ModelMixture of ExpertsDiffusion modelTextBenchmark
π― What it does: For inference-time sequence generation in diffusion-based large language models (dLLMs), this paper proposes a training-free test-time scaling method called HEX, which significantly improves inference accuracy by integrating voting across semi-autoregressive decoding paths with different block sizes.
Test-time Verification via Optimal Transport: Coverage, ROC, & Sub-optimality
Arpan Mukherjee (Imperial College London), Deniz Gunduz (Imperial College London)
CodeOptimizationComputational EfficiencyText
π― What it does: This paper views test-time verification as a sampling problem, using the optimal transport framework to uniformly analyze the interactions among generator coverage, validator ROC, and suboptimality of sampling algorithms.
π― What it does: Propose a graph Transformer architecture named TetraGT, which models bond angles and torsion angles directly as structured tokens, leveraging triangular and tetrahedral geometric constraints to achieve high-quality molecular geometry prediction, and applies it to molecular property prediction.
CodeObject DetectionGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextGraphBenchmark
π― What it does: This paper proposes a system called TEXT2ARCH for generating scientific architecture diagrams based on natural language descriptions. The core idea is to first convert text into intermediate DOT graph code, and then render the architecture diagram using a DOT compiler;
Text2Grad: Reinforcement Learning from Natural Language Feedback
Hanyang Wang (University of Chicago), Dongmei Zhang (Microsoft)
CodeReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
π― What it does: Proposes TEXT2GRAD, a framework that converts natural language feedback into span-level gradients and directly applies them to reinforcement learning.
Textual Bayes: Quantifying Prompt Uncertainty in LLM-Based Systems
Brendan Leigh Ross (Layer 6 AI), Jesse C. Cresswell (Layer 6 AI)
CodeExplainability and InterpretabilityLarge Language ModelText
π― What it does: Propose the Textual Bayes framework, treating LLM prompts as Bayesian text parameters, utilizing Bayesian inference to sample prompts, thereby achieving uncertainty quantification of LLM outputs;
π― What it does: Investigate various pathological phenomena caused by global amortized inference in sparse autoencoders (SAE) and propose local amortized SAE (LocA-SSAE) to alleviate these issues
The Alignment Auditor: A Bayesian Framework for Verifying and Refining LLM Objectives
Matthieu Bou (Imperial College London), Sonali Parbhoo (Imperial College London)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Proposed a three-stage alignment auditor framework (Alignment Auditor) that utilizes Bayesian inverse reinforcement learning (IRL) to recover reward posterior from contrastive data between experts and baselines, quantifies reward uncertainty, identifies shortcut paths and out-of-distribution (OOD) inputs through uncertainty diagnosis, and finally validates its practicality at the policy level by fine-tuning in RLHF using the posterior mean as the reward signal.
The Choice of Divergence: A Neglected Key to Mitigating Diversity Collapse in Reinforcement Learning with Verifiable Reward
Long Li (Fudan University), Yuan Qi (Fudan University)
CodeData-Centric LearningLarge Language ModelReinforcement LearningText
π― What it does: Propose the Diversity-Preserving Hybrid RL (DPH-RL) framework, addressing diversity collapse and catastrophic forgetting in RLVR through f-divergence regularization.
CodeExplainability and InterpretabilityRepresentation LearningTransformerContrastive LearningImageTextMultimodalityAudio
π― What it does: Proposed an unsupervised concept extraction method based on clustering differences, which uses contrastive activation differences for clustering and enhances diversity through inverse skewness weighting;
CodeSafty and PrivacyAdversarial AttackPrompt EngineeringDiffusion modelText
π― What it does: This paper systematically studies the security vulnerabilities of diffusion-based large language models (dLLMs), proposing and implementing an automated jailbreak framework called DIJA, which can bypass existing security mechanisms by interpolating masks and text to generate harmful outputs.
The False Promise of Zero-Shot Super-Resolution in Machine-Learned Operators
Mansi Sakarvadia (University of Chicago), Michael W. Mahoney (Lawrence Berkeley National Laboratory)
CodeSuper ResolutionPhysics Related
π― What it does: Evaluate the performance of machine learning operators on zero-shot super-resolution and propose a multi-resolution training scheme that requires only a few high-resolution samples
The Geometry of LLM Quantization: GPTQ as Babai's Nearest Plane Algorithm
Jiale Chen (Institute of Science and Technology Austria), Dan Alistarh (Institute of Science and Technology Austria)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Studied the geometric interpretation of the GPTQ weight quantization method, equating it to the Babai closest vector algorithm without basis scaling, derived error upper bounds, and proposed an improved unclipped quantization scheme and efficient GPU inference kernel.
The Geometry of Reasoning: Flowing Logics in Representation Space
Yufa Zhou (Duke University), Anru Zhang
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
π― What it does: Study the geometric dynamics of LLM reasoning, modeling reasoning as a smooth flow in the embedding space and revealing logical structures through geometric quantities such as velocity and curvature.
The Human Brain as a Dynamic Mixture of Expert Models in Video Understanding
Christina Sartzetaki (University of Amsterdam), Iris Groen (University of Amsterdam)
CodeExplainability and InterpretabilityRepresentation LearningMixture of ExpertsVideoBiomedical Data
π― What it does: This paper investigates the spatiotemporal characteristics of visual processing through a large-scale model-brain alignment evaluation involving over 100 deep visual models and dynamic EEG video recordings.
The Ideation-Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas
Chenglei Si (Stanford University), Diyi Yang (Stanford University)
CodeAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: This paper conducted a randomized controlled trial, recruiting 43 expert researchers to execute research ideas generated by Claude-3.5-Sonnet and ideas proposed by human experts. The generated code and papers were subjected to blind review.
π― What it does: This paper proposes a new evaluation framework to systematically study the quality, diversity, and consistency of synthetic data generated by text-to-image (T2I) models under different prompt complexities, and compares them with real data.
The Lie of the Average: How Class Incremental Learning Evaluation Deceives You?
Guannan Lai (Nanjing University), Han-Jia Ye (Nanjing University)
CodeClassificationVision Language ModelContrastive LearningImageBenchmark
π― What it does: Proposed a distribution and generalization evaluation protocol called EDGE based on extreme sequences, improving the evaluation method for Class Incremental Learning (CIL).
π― What it does: Explored the limits of improving model performance through resampling during inference when incomplete verifiers (e.g., unit tests) are present, proving that weak models cannot match the single-sample accuracy of strong models even with infinite sampling;
The Mind's Transformer: Computational Neuroanatomy of LLM-Brain Alignment
Cheng-Yeh Chen (Georgia Institute of Technology), Raghupathy Sivakumar (Georgia Institute of Technology)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelMultimodalityBiomedical DataMagnetic Resonance Imaging
π― What it does: Systematically decompose the 13 intermediate states of the Transformer block, map them to brain fMRI data, and propose the MindTransformer framework to enhance brain-model alignment
The Polar Express: Optimal Matrix Sign Methods and their Application to the Muon Algorithm
Noah Amsel (New York University), Robert M. Gower (Flatiron Institute)
CodeOptimizationComputational EfficiencyText
π― What it does: Propose a new GPU-friendly matrix polarity decomposition algorithm called Polar Express to accelerate the approximation of matrix sign functions in the Muon optimizer.
The Rank and Gradient Lost in Non-stationarity: Sample Weight Decay for Mitigating Plasticity Loss in Reinforcement Learning
Zihao Wu (Tianjin University), Jianye HAO
CodeOptimizationReinforcement Learning
π― What it does: Identify plasticity loss in reinforcement learning through theoretical analysis and propose the Sample Weight Decay (SWD) method to alleviate the gradient decay problem.
Alexey Yermakov (University of Washington), J. Nathan Kutz (University of Washington)
CodeRecurrent Neural NetworkTime SeriesBenchmarkPhysics Related
π― What it does: Proposed a generic machine learning task framework (CTF) for seismic wavefields, evaluated and benchmarked on three-scale public datasets;
THE SELF-RE-WATERMARKING TRAP: FROM EXPLOIT TO RESILIENCE
Vithurabiman Senthuran (Deakin University), Uthayasanker Thayasivam (University of Moratuwa)
CodeSafty and PrivacyAdversarial AttackConvolutional Neural NetworkAuto EncoderImage
π― What it does: This paper proposes a robust watermarking framework that defends against malicious re-embedding attacks by the same encoder, utilizing Lipschitz constraints and self-watermark adversarial training;
π― What it does: Propose to view the noise space of diffusion models as a spatiotemporal statistical manifold and construct analytical geodesics using Fisher-Rao geometry;
π― What it does: Proposed and implemented the Tool Decathlon (TOOLATHLON) benchmark to evaluate the tool calling and task execution capabilities of language models in real-world, long-sequence, multi-application environments;
The Unseen Frontier: Pushing the Limits of LLM Sparsity with Surrogate-Free ADMM
Kwanhee Lee (POSTECH), Namhoon Lee (POSTECH)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Developed an ADMM-based proxy-free sparsification framework called ELSA, which can increase the sparsity of LLMs to over 90% while maintaining language modeling performance.
CodeExplainability and InterpretabilityImageVideoTextTabularBiomedical Data
π― What it does: Propose a framework based on Bayesian decision theory to quantify information value in human-machine collaborative decision-making, and introduce global and instance-level complementary information value (ACIV, ILIV) metrics; design a new explanation method called ILIV-SHAP based on ILIV; validate the framework and explanation method through online human-machine collaborative experiments, demonstrating improved decision performance; and demonstrate the framework's application in model selection and feature importance analysis in tasks such as chest X-ray diagnosis and deepfake detection.
THEMIS: Towards Holistic Evaluation of MLLMs for Scientific Paper Fraud Forensics
Tzu-Yen Ma (Beijing University of Posts and Telecommunications), Haihong E (Beijing University of Posts and Telecommunications)
CodeAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes a multimodal benchmark named THEMIS for evaluating the expert-level capabilities of large language models in visual fraud reasoning within academic papers.
π― What it does: This paper systematically analyzes the DDIM inversion process, revealing that the potential noise generated does not follow a Gaussian distribution, particularly leading to a lack of diversity in latent variables in image smooth regions. It proposes replacing the initial steps with the forward diffusion process to eliminate these correlations, thereby improving the quality of interpolation and editing.
π― What it does: Proposes a two-stage self-supervised pre-training and end-to-end fine-tuning framework for training diffusion models and consistency models in pixel space,
Thicker and Quicker: The Jumbo Token for Fast Plain Vision Transformers
Anthony Fuller (Carleton University), James R Green (Carleton University)
CodeClassificationSegmentationRetrievalTransformerContrastive LearningImageTextMultimodalityTime Series
π― What it does: Propose an architecture that integrates a wide Jumbo token (J times the width) into a standard ViT, maintaining pure attention and non-hierarchical characteristics, significantly enhancing model capacity and performance.
Thinking on the Fly: Test-Time Reasoning Enhancement via Latent Thought Policy Optimization
Wengao Ye (University of Oxford), Lianlei Shan (University of Chinese Academy of Sciences)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: During the testing phase of reasoning LLMs, online reinforcement learning is applied directly in the latent space on 'thought' vectors to enhance reasoning performance
π― What it does: Propose ThinkMorph, a unified model that enables complementary interactive chained thinking between text and images, fine-tuned on approximately 24K high-quality interactive reasoning trajectories, significantly enhancing performance on vision-centric tasks and demonstrating multiple emergent capabilities.
Nicolas Menet (IBM Research), Abbas Rahimi (IBM Research)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical DataPhysics Related
π― What it does: An scalable Thompson sampling algorithm called TOSFIT is proposed by fine-tuning a large language model to generate a strategy that maximizes the probability of reward, for Bayesian optimization in large unstructured discrete spaces.
THOR: Tool-Integrated Hierarchical Optimization via RL for Mathematical Reasoning
Qikai Chang, Jianqing Gao (iFLYTEK Research)
CodeOptimizationAI Code AssistantLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: Propose the THOR framework, which significantly improves the accuracy of large language models in mathematical reasoning and code generation tasks by leveraging tool integration and hierarchical reinforcement learning.
Threading Keyframe with Narratives: MLLMs as Strong Long Video Comprehenders
Bo Fang (City University of Hong Kong), Antoni B. Chan (City University of Hong Kong)
CodeRecognitionOptimizationRepresentation LearningLarge Language ModelVision Language ModelVideoTextBenchmark
π― What it does: Proposes an unsupervised long video understanding framework called Nar-KFC, consisting of two stages: Key Frame Capture (KFC) and Narrative Key Frame Interpolation (Nar-KFC);
Three Forward, One Backward: Memory-Efficient Full-Rank Fine-Tuning of Large Models via Extra Forward Passes
Jia Zhang (Jilin University), Bin Gu (Jilin University)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Propose LMAO, an alternating optimization framework that combines LoRA and MeZO, achieving full-rank updates through a three-forward-one-backward step, enabling memory-efficient fine-tuning of large models.
Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs
Zhiyu Pan (Huazhong University of Science and Technology), Jieping Ye (Alibaba Cloud)
CodeExplainability and InterpretabilityRepresentation LearningData-Centric LearningLarge Language ModelSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Propose the VC-STaR framework, which leverages visual contrast VQA to correct hallucinations in vision-language models, and generates a high-quality visual reasoning dataset called VisCoR55K for self-improvement of VLMs.
TianQuan-S2S: A Subseasonal-to-Seasonal Global Weather Model via Incorporate Climatology State
Guowen Li (Sun Yat-Sen University), Haohuan Fu (Huawei Technologies Co., Ltd)
CodeConvolutional Neural NetworkTransformerTime SeriesSequentialPhysics Related
π― What it does: Proposed and implemented the TianQuan-S2S model for global subseasonal-to-seasonal weather forecasting spanning 15β45 days, integrating initial meteorological states with 38-year average climate data and injecting uncertainty noise into the Transformer architecture.
TIGaussian: Disentangle Gaussians for Spatial-Awared Text-Image-3D Alignment
Jiarun Liu (Alibaba Group), Sheng Yang (Alibaba Group)
CodeClassificationRetrievalRepresentation LearningTransformerVision Language ModelDiffusion modelContrastive LearningGaussian SplattingImageTextPoint Cloud
π― What it does: This paper proposes the TIGAUSSIAN framework, achieving alignment and pretraining across text-image-3D Gaussian (3DGS) three modalities.
π― What it does: This paper studies estimating the potential energy of a Schrodinger bridge using an Ornstein-Uhlenbeck reference process, given only i.i.d. samples from the initial and terminal distributions, and provides an upper bound on the generalization error of the empirical risk minimizer;
TileLang: Bridge Programmability and Performance in Modern Neural Kernels
Lei Wang (Peking University), Zhi Yang (Peking University)
CodeOptimizationComputational Efficiency
π― What it does: Propose TILELANG, a programmable tile-level system that provides explicit memory placement, data movement, and parallel scheduling primitives, and achieves automated tile recommendation and tile inference through a unified fused tile-level dataflow graph (FTG), significantly reducing the complexity of AI kernel development and improving performance.
π― What it does: Proposes a no-training, plug-and-play framework that leverages user-provided rough animations and image conditions to control motion and appearance in video generation.
TimeOmni-1: Incentivizing Complex Reasoning with Time Series in Large Language Models
Tong Guan (Griffith University), Shirui Pan (Griffith University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTime Series
π― What it does: Proposed the TSR-SUITE dataset and the TIMEOMNI-1 model, focusing on three types of reasoning tasks for time series: perception, extrapolation, and decision-making, filling the gap of lacking high-quality reasoning data and a general reasoning framework.
π― What it does: Propose the TIMERECIPE framework to conduct a unified benchmark evaluation at the module level of time series forecasting models, systematically analyzing the effectiveness of components such as preprocessing, embedding, and feedforward modules;
TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning
Junwen Pan (ByteDance), Qi She (ByteDance)
CodeRetrievalReinforcement LearningVideoBenchmark
π― What it does: This paper proposes the TimeSearch-R framework, which transforms video temporal search into text-video interactive reasoning and learns optimal search strategies through reinforcement learning.
TimeSeriesExamAgent: Creating Time Series Reasoning Benchmarks at Scale
Malgorzata Gwiazda (Technical University of Munich), Artur Dubrawski (Carnegie Mellon University)
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningAgentic AITime SeriesBiomedical DataBenchmarkFinance Related
π― What it does: Proposed two frameworks, TimeSeriesExam and TimeSeriesExamAgent, for constructing controllable synthetic time series evaluation question banks and automatically generating domain-specific evaluation questions based on real data.
TIMESLIVER : SYMBOLIC-LINEAR DECOMPOSITION FOR EXPLAINABLE TIME SERIES CLASSIFICATION
Akash Pandey (Northwestern University), Sinan Keten (Northwestern University)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkTime SeriesBiomedical Data
π― What it does: Propose TimeSliver, an interpretable deep learning framework that linearly combines original time series with symbolic discretization results;
Shangshang Wang (University of Southern California), Willie Neiswanger (University of Southern California)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: Post-training on the 1.5B Tiny model by combining LoRA and reinforcement learning to build the Tina model, aiming to enhance multi-step reasoning capabilities.
π― What it does: For search-augmented large language models (LLMs) in open-domain question answering tasks, reinforcement learning (RL) is employed to achieve more stable training and improve credit assignment for multi-step tool calls.
π― What it does: Proposed the TITOK framework, leveraging the token-level contrastive advantages of the source model LoRA to achieve LoRA knowledge transfer;
To Compress or Not? Pushing the Frontier of Lossless GenAI Model Weights Compression with Exponent Concentration
Zeyu Yang (Rice University), Anshumali Shrivastava (Rice University)
CodeCompressionTransformer
π― What it does: Analyze and prove the exponential concentration phenomenon in the weights of generative AI models, and based on this, design ECF8, an FP8 format that achieves lossless compression.
π― What it does: Investigate the capabilities and limitations of State Space Models (SSM) in long text generation, demonstrating that SSM cannot generate long tables without tools or single-round tool usage, and propose achieving length generalization through interactive tool usage, with theoretical and experimental validation on arithmetic, reasoning, and programming tasks.
Token-Based Audio Inpainting via Discrete Diffusion
Tali Dror (Ben-Gurion University of the Negev), Eliya Nachmani (Ben-Gurion University of the Negev)
CodeRestorationTransformerDiffusion modelAudio
π― What it does: Proposed a discrete diffusion-based audio restoration framework called AIDD, which achieves natural reconstruction of long missing audio segments by quantizing audio into discrete tokens and performing the diffusion reverse process in the token space.
Token-Efficient Item Representation via Images for LLM Recommender Systems
Kibum Kim (KAIST), Chanyoung Park (KAIST)
CodeRecommendation SystemLarge Language ModelPrompt EngineeringVision Language ModelImageTextRetrieval-Augmented Generation
π― What it does: This work proposes an efficient product representation method based on images (I-LLMRec), which uses a small number of tokens instead of lengthy textual descriptions, enabling large language models (LLMs) to efficiently and comprehensively capture product semantics in recommendation tasks;
Token-Efficient Long-Term Interest Sketching and Internalized Reasoning for LLM-based Recommendation
Zhihao Ding (Hong Kong Polytechnic University), Jieming Shi (Hong Kong Polytechnic University)
CodeRecommendation SystemLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
π― What it does: Propose the SIREN framework, which uses large language models to predict user ratings, and addresses issues of long-term historical noise and inference latency by constructing interest sketches and internalizing reasoning.
Token-Guard: Towards Token-Level Hallucination Control via Self-Checking Decoding
Yifan Zhu (Beijing University of Posts and Telecommunications), Haoran Luo (Nanyang Technological University)
CodeGenerationLarge Language ModelTextBiomedical DataBenchmarkFinance Related
π― What it does: Designed and implemented Token-Guard, a token-level hallucination control decoding framework that performs step-by-step self-inspection, segmented evaluation, and local/global iterative correction during the generation process.
Token-Importance Guided Direct Preference Optimization
Ning Yang (Institute of Automation Chinese Academy of Sciences), Haijun Zhang (University of Science and Technology Beijing)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
π― What it does: Proposes the Token-Importance Guided Direct Preference Optimization (TI-DPO) framework to achieve fine-grained token-level optimization during large language model (LLM) alignment, combining hybrid weighting and triplet loss.
Token-level Data Selection for Safe LLM Fine-tuning
Yanping Li (Hong Kong University of Science and Technology), Jun Zhang (Hong Kong University of Science and Technology)
CodeSafty and PrivacyData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: This work proposes a token-level data filtering framework called TOSS, which assesses the safety risk of each token using the loss difference between a safety degradation model and a utility model, and achieves precise removal of hazardous tokens through global ranking, thereby enhancing both safety and utility during LLM fine-tuning; simultaneously, it introduces an iteratively improved version, TOSS-Pro, to further strengthen the identification capability of the safety degradation model.
TokMem: One-Token Procedural Memory for Large Language Models
Zijun Wu (University of Alberta), Lili Mou (University of Alberta)
CodeTransformerLarge Language ModelText
π― What it does: Propose the TokMem framework, which compresses reusable task programs into a single trainable memory token, maintaining the LLM frozen to achieve procedural memory and composable execution without context length overhead;
TokUR: Token-Level Uncertainty Estimation for Large Language Model Reasoning
Tunyu Zhang (Rutgers University), Hao Wang (Rutgers University)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Proposes a training-agnostic token-level uncertainty estimation framework called TokUR, which generates prediction distributions for each token by applying low-rank random perturbations to attention layer weights, thereby evaluating self-assessment and self-improvement of large language models during multi-step reasoning.
Tools are under-documented: Simple Document Expansion Boosts Tool Retrieval
Xuan Lu (Shanghai Jiao Tong University), Xiaoyu Shen (Ningbo Key Laboratory of Spatial Intelligence and Digital Derivative)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
π― What it does: Propose the TOOL-REX benchmark and build a low-cost LLM document expansion pipeline, enriching tool documentation with structured fields such as functional description, use cases, limitations, and tags; train specialized tool retrieval models, Tool-Embed (dense retriever) and Tool-Rank (LLM reranker), based on the expanded data.
ToolWeaver: Weaving Collaborative Semantics for Scalable Tool Use in Large Language Models
Bowen Fang (Chinese Academy of Sciences), Liang Wang (Chinese Academy of Sciences)
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Propose the ToolWeaver framework, which uses hierarchical tool code representations and collaboration-aware vector quantization to learn tool semantics and collaboration relationships. These codes are injected into LLMs through generation-aligned fine-tuning, achieving end-to-end generation for tool selection and invocation.
π― What it does: Developed the TOPOFORMER framework, converting the topological structure of graphs into sequences processable by Transformers, thereby enabling graph-level learning.
π― What it does: Propose the TAQ-GAD framework, utilizing topological anomaly quantization (NBS, PIS) to dynamically screen pseudo-anomalous nodes and enhance the graph structure through a virtual anomaly center, achieving semi-supervised graph anomaly detection.
π― What it does: In generative models, a flow matching framework is used to model signals on structured spaces (such as graphs and simplicial complexes), and Topological Flow Matching (TFM) is proposed by introducing a drift induced by the Laplacian operator in the reference process to capture topological information.
π― What it does: Proposes the TopoRAG framework, which elevates text graphs to cellular complexes, employs topology-based subcomplex retrieval and multi-dimensional message passing, and finally injects the retrieval results into large language models for question answering.
Toward Conservative Planning from Human-AI Preferences in Reinforcement Learning
Huazhong Wang (University of California, Irvine), Wenzhuo Zhou (University of California, Irvine)
CodeReinforcement Learning from Human FeedbackReinforcement LearningBenchmark
π― What it does: This paper proposes a model-driven conservative planning algorithm called MCP for offline reinforcement learning based on human-AI preference signals, learning robust policies under partial data coverage.
Toward Effective Tool-Integrated Reasoning via Self-Evolved Preference Learning
Yifei Chen (Renmin University of China), Zhicheng Dou (Renmin University of China)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes the Tool-Light framework, which significantly improves the tool-integrated reasoning efficiency and accuracy of large language models (LLMs) through entropy-guided sampling and two-stage self-evolution DPO training.
Toward Efficient Exploration by Large Language Model Agents
Dilip Arumugam (Princeton University), Thomas L. Griffiths (Princeton University)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AITextChain-of-Thought
π― What it does: This paper proposes an approach to explicitly implement the posterior sampling reinforcement learning (PSRL) algorithm using large language models (LLMs), aiming to address the exploration efficiency issues of LLM agents in natural language environments;
Toward Faithful Retrieval-Augmented Generation with Sparse Autoencoders
Guangzhi Xiong (University of Virginia), Aidong Zhang (University of Virginia)
CodeRetrievalAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerAuto EncoderTextRetrieval-Augmented Generation
π― What it does: Developed a lightweight retrieval-augmented generation (RAG) hallucination detector called RAGLens based on sparse autoencoders (SAE), utilizing internal activations of large language models (LLMs) to detect the authenticity of generated text and perform interpretable analysis.
Toward Principled Flexible Scaling for Self-Gated Neural Activation
Sudong Cai (Hong Kong Polytechnic University), Bing WANG
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerImageText
π― What it does: Propose the FleS mechanism to address the non-local tension problem in self-gated activation, providing interpretable flexible scaling activation.
Toward Safer Diffusion Language Models: Discovery and Mitigation of Priming Vulnerability
Shojiro Yamabe (Institute of Science Tokyo), Jun Sakuma (Institute of Science Tokyo)
CodeSafty and PrivacyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningDiffusion modelText
π― What it does: Quantitatively and theoretically analyze the 'priming vulnerability' of discrete diffusion language models (MDLM), and propose a novel safety alignment method called Recovery Alignment (RA), enabling the model to recover to safe responses even after encountering harmful confirmation words in intermediate steps.
π― What it does: Proposes a foundational model for crowdsourcing label aggregation called CrowdFM, achieving label inference through pre-trained bilateral graph neural networks without requiring dataset training.
CodeDrug DiscoveryTransformerDiffusion modelBiomedical Data
π― What it does: This paper proposes an Atom-level Diffusion Transformer (ADiT) based on AlphaFold3, which can serve as a general-purpose foundational model to predict the binding affinity of multiple types of biomolecules.
Towards Anomaly-Aware Pre-Training and Fine-Tuning for Graph Anomaly Detection
Yunhui Liu (Nanjing University), Tieke He (Hong Kong University of Science and Technology (Guangzhou))
CodeAnomaly DetectionGraph Neural NetworkGraphTabularBenchmarkFinance Related
π― What it does: This paper proposes a two-phase framework APF for graph anomaly detection, which first performs anomaly-aware pre-training through unsupervised Rayleigh quotient subgraph sampling and dual spectral filtering, and then significantly improves the anomaly detection effect by adopting node and dimension adaptive fusion and anomaly-aware regularization during the fine-tuning phase.
Towards Better Branching Policies: Leveraging the Sequential Nature of Branch-and-Bound Tree
Ce Zhang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences), Guoliang Fan (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences)
CodeOptimizationContrastive LearningSequential
π― What it does: Propose a sequential branching strategy called Mamba-Branching based on the Mamba network, which utilizes contrastive learning for pre-training embeddings and models the branching paths of B&B trees in an autoregressive manner to achieve more efficient branching decisions.
π― What it does: Propose the Construct-and-Refine (CaR) framework, which jointly trains a neural constructor and refiner to achieve explicit feasibility correction, applicable to complex-constrained vehicle routing problems (VRP).
Towards High Data Efficiency in Reinforcement Learning with Verifiable Reward
Xinyu Tang (Renmin University of China), JUN ZHOU
CodeReinforcement LearningText
π― What it does: This work proposes DEPO, a data efficiency enhancement method for RLVR that combines offline multi-objective sample selection with online explorability filtering.
Towards Improved Sentence Representations using Token Graphs
Krishna Sri Ipsit Mantri (University of Bonn), Moshe Eliasof (University of Cambridge)
CodeComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelText
π― What it does: Propose a lightweight pooling module called GLOT based on token graphs, utilizing a frozen large language model to generate sentence-level representations
Towards Multimodal Data-Driven Scientific Discovery Powered by LLM Agents
Fan Liu (Hong Kong University of Science and Technology), Hao Liu (Hong Kong University of Science and Technology)
CodeDrug DiscoveryLarge Language ModelAgentic AIImageTextMultimodalityTabularTime SeriesBiomedical DataBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Built MoSciBench β the first benchmark for multi-modal data-driven scientific discovery, covering 6 disciplines, 7 data modalities, and a total of 88 end-to-end tasks.
π― What it does: Propose a one/two-step causal video generation framework by distilling multi-step diffusion models, and introduce adversarial self-distillation and first-frame enhancement strategies to achieve high-quality real-time video generation.
CodeLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: This paper designs the first benchmark for personalized deep research, PDR-Bench, and proposes a three-dimensional evaluation framework PQR (Personalization, Content Quality, Fact Reliability) to systematically evaluate the personalization capabilities of deep research agents (DRA).
Towards Privacy-Guaranteed Label Unlearning in Vertical Federated Learning: Few-Shot Forgetting Without Disclosure
Hanlin Gu (WeBank AI Lab), Chee Seng Chan (Universiti Malaya)
CodeFederated LearningSafty and PrivacyImageTextBiomedical Data
π― What it does: To address the label forgetting problem in vertical federated learning (VFL), a few-shot label forgetting framework is proposed: first, manifold mixup is applied to the representation layer to synthesize embeddings, then gradient ascent is performed on active and passive models to achieve label forgetting, followed by a recovery phase to maintain the performance of retained samples.
Towards Quantization-Aware Training for Ultra-Low-Bit Reasoning LLMs
Yasuyuki Okoshi (Institute of Science), Masato Motomura (Institute of Science)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes a two-stage quantization-aware training (QAT) workflow tailored for post-training inference large language models, aiming to maintain inference capabilities after ultra-low-bit (<4 bits) quantization;
CodeData SynthesisLarge Language ModelTextGraphBenchmark
π― What it does: Propose the FuncBenchGen framework for automatically generating clean, controllable difficulty multi-step function call evaluation tasks;