International Conference on Learning Representations Β· 2207 papers
Endowing GPT-4 with a Humanoid Body: Building the Bridge Between Off-the-Shelf VLMs and the Physical World
Yingzhao Jian (Zhejiang University), Hehe Fan (Zhejiang University)
CodeRobotic IntelligenceTransformerReinforcement LearningVision Language ModelDiffusion modelAuto EncoderMultimodality
π― What it does: Propose the BiBo framework, leveraging existing Vision-Language Models (e.g., GPT-4o) and diffusion models to control humanoid robots to perform diverse physical interaction tasks.
CodeMeta LearningSupervised Fine-TuningMixture of ExpertsVision Language ModelContrastive LearningMultimodalityBenchmark
π― What it does: Introduces the ConDU framework, which uses model fusion to achieve continuous learning in vision-language models, compatible with full-parameter fine-tuning and parameter-efficient fine-tuning, and supports zero-shot inference.
π― What it does: This paper proposes a difference-aware communication compression method based on locally calibrated data, significantly reducing communication volume and improving model accuracy in federated learning.
Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejection
Minjae Kang (Yonsei University), Jaehyung Kim (Yonsei University)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Proposes a dynamic activation regulation method called DIRECTER to enhance large language models' (LLM) instruction-following capability during inference while avoiding text quality degradation caused by over-regulation.
Enhancing Language Model Reasoning with Structured Multi-Level Modeling
Siheng Xiong (Georgia Institute of Technology), Faramarz Fekri (Georgia Institute of Technology)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: Proposes a Multi-Level Reasoning (MLR) framework that decomposes the reasoning process into high-level planning descriptors and low-level detailed reasoning, forming an alternating plan-execute cycle; simultaneously introduces an iterative Step-DPO training process and employs Twisted Sequential Monte Carlo (TSMC) to construct process-level preferences, addressing the long-sequence credit assignment challenge caused by sparse single-result rewards.
π― What it does: Propose the Chain-of-Decomposition (CoD) framework, decomposing the KBQA task into three steps: retrieval, reconstruction, and reasoning, reducing the burden on LLMs and improving performance.
Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval
Fanpu Cao (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)
CodeConvolutional Neural NetworkTime Series
π― What it does: Propose a lightweight, plug-and-play Global Temporal Retriever (GTR) module to expand the temporal awareness range of multivariate time series prediction models, thereby capturing long-term global periodic patterns.
Enhancing Visual Token Representations for Video Large Language Models via Training-free Spatial-Temporal Pooling and Gridding
Bingjun Luo (Tsinghua University), Xinpeng Ding (Xidian University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelVideo
π― What it does: Proposes ST-GridPool, a training-agnostic visual token enhancement method to improve visual representations in video large language models;
EnsembleSHAP: Faithful and Certifiably Robust Attribution for Random Subspace Method
Yanting Wang (Pennsylvania State University), Jinyuan Jia (Pennsylvania State University)
CodeExplainability and InterpretabilityComputational EfficiencyAdversarial AttackImageText
π― What it does: Propose EnsembleSHAP, an efficient and provably secure feature attribution method specifically designed for random subspace methods.
Entering the Era of Discrete Diffusion Models: A Benchmark for SchrΓΆdinger Bridges and Entropic Optimal Transport
Xavier Aramayo Carrasco (Applied AI Institute), Alexander Korotin (Applied AI Institute)
CodeOptimizationDiffusion modelBenchmarkPhysics Related
π― What it does: This paper proposes a benchmark for the SchrΓΆdinger Bridge and Entropic Optimal Transport (EOT) in discrete space, utilizing analytical SB solutions to construct discrete distribution pairs, and develops and evaluates three new algorithms (DLightSB, DLightSB-M, Ξ±-CSBM) on this benchmark.
Entropy-Based Block Pruning for Efficient Large Language Models
Liangwei Yang (Salesforce Ai Research), Shelby Heinecke (Salesforce Ai Research)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This paper proposes EntroDrop, a block pruning method based on the entropy increase of hidden representations, aimed at reducing the computational and storage requirements of large language models.
π― What it does: Propose a multi-task keyframe manipulation strategy called EquAct, which maps language instructions, point cloud observations, and actions into a shared space through SE(3) equivariant networks, achieving theoretical generalization for 3D pose variations.
π― What it does: Proposed a self-supervised learning framework called Equivariant Splitting (ES), for image recovery using only a single incomplete measurement.
ERGO: Efficient High-Resolution Visual Understanding for Vision-Language Models
Jewon Lee (Nota Inc), Bo-Kyeong Kim (Nota Inc)
CodeComputational EfficiencyTransformerReinforcement LearningVision Language ModelMultimodality
π― What it does: Designed and implemented a two-stage coarse-to-fine high-resolution visual reasoning pipeline, training the model via reinforcement learning to autonomously locate task-related regions for fine-grained reasoning under low-resolution inputs.
π― What it does: Proposes ERK-Guid, a method that utilizes embedded Runge-Kutta error as a guidance signal to correct local truncation errors in rigid regions during diffusion model sampling, thereby improving sampling quality and efficiency.
ES-dLLM: Efficient Inference for Diffusion Large Language Models by Early-Skipping
Zijian Zhu (Tsinghua University), Kaisheng Ma (Tsinghua University)
CodeComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
π― What it does: Studied computational redundancy in the inference process of diffusion large language models (dLLM), and proposed a training-agnostic early skipping framework called ES-dLLM to reduce computational cost per iteration.
Shujian Gao (Fudan University), Junjun He (Shanghai AI Laboratory)
CodeExplainability and InterpretabilityRepresentation LearningTransformerImage
π― What it does: Proposed a concept bottleneck model (CBM) integrating implicit vector quantization (IVQ) and magnet attention, addressing visual feature collapse and enabling many-to-many visual concept learning.
π― What it does: Investigated estimating the global dimension of neural representations from limited samples and noise, proposing an unbiased participation ratio estimator.
Estimating Semantic Alphabet Size for LLM Uncertainty Quantification
Lucas Hurley McCabe (George Washington University), H Howie Huang (George Washington University)
CodeExplainability and InterpretabilityLarge Language ModelText
π― What it does: To address uncertainty estimation in large language models, this paper proposes an improved method for estimating the size of the semantic alphabet and uses it to correct discrete semantic entropy (DSE), achieving more accurate entropy estimation and error detection with only a small number of samples.
ETGS: Explicit Thermodynamics Gaussian Splatting for Dynamic Thermal Reconstruction
Zhongwen Wang (Nanjing University of Science and Technology), Quansen Sun (Nanjing University of Science and Technology)
CodeGaussian SplattingVideoTime SeriesPhysics Related
π― What it does: Proposed the ETGS method, which achieves dynamic thermal field reconstruction using 3D Gaussian splatting with explicit thermodynamic modeling.
Evaluating Cross-Modal Reasoning Ability and Problem Characteristics with Multimodal Item Response Theory
Shunki Uebayashi, Koh Takeuchi (CyberAgent)
CodeExplainability and InterpretabilityComputational EfficiencyData-Centric LearningLarge Language ModelVision Language ModelMultimodalityBenchmark
π― What it does: This paper proposes Multimodal Multidimensional Item Response Theory (M3IRT) and its simplified version M2IRT to evaluate the cross-modal reasoning capabilities of Multimodal Large Language Models (MLLMs). By decomposing item difficulty and model ability into image, text, and cross-modal dimensions, it achieves automatic construction of high-quality subsets and significantly reduces evaluation costs.
Evaluating Text Creativity across Diverse Domains: a Dataset and Large Language Model Evaluator
Qian Cao (Renmin University of China), Ruihua Song (Beijing Normal University)
CodeLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
π― What it does: Propose a context-based dual comparison framework for evaluating text creativity, and develop a large cross-domain dataset CreataSet and the corresponding LLM evaluator CrEval
EventFlash: Towards Efficient MLLMs for Event-Based Vision
Shaoyu Liu (Xidian University), Xiangyang Ji (Tsinghua University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelMultimodalityTime Series
π― What it does: Proposed EventFlash, an efficient event-based visual multimodal large language model that achieves fast inference and long sequence processing through spatiotemporal token sparsification.
Everything in Its Place: Benchmarking Spatial Intelligence of Text-to-Image Models
Zengbin Wang (AMAP, Alibaba Group), Xiangxiang Chu (AMAP, Alibaba Group)
CodeGenerationData SynthesisLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
π― What it does: Proposed the SpatialGenEval benchmark, which evaluates the spatial intelligence of T2I models with information-dense long prompts and multiple-choice questions across 10 spatial subdomains, and constructed the SpatialT2I dataset for supervised fine-tuning.
Evolution and compression in LLMs: on the emergence of human-aligned categorization
Nathaniel Imel (New York University), Noga Zaslavsky (New York University)
CodeClassificationCompressionLarge Language ModelPrompt EngineeringText
π― What it does: This paper investigates the semantic systems of large language models (LLMs) in color classification tasks. It first evaluates their alignment with human English systems and information bottleneck (IB) efficiency in an English color naming experiment. Subsequently, it simulates cultural evolution through the proposed iterative contextual language learning (IICLL) method, observing how LLMs evolve near-human color category systems without explicit supervision. Preliminary validation is conducted in different semantic domains such as Shepard circles.
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderText
π― What it does: Track and analyze the evolution of features during the pre-training process of large-scale language models, using cross-snapshot sparse autoencoder (crosscoder) to capture the emergence, rotation, and disappearance of features across different training stages, and further associate micro-level evolution with downstream task performance through feature attribution.
Evolving Graph Structured Programs for Circuit Generation with Large Language Models
Yinqi Bai (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeGenerationOptimizationTransformerLarge Language ModelPrompt EngineeringGraphBenchmark
π― What it does: This paper proposes CircuitEvo, an evolutionary circuit program generation framework based on large language models, which iteratively evolves circuit programs and achieves functionally accurate and more compact logic synthesis through structure-aware functional optimization.
EvoTest: Evolutionary Test-Time Learning for Self-Improving Agentic Systems
Yufei He (National University of Singapore), Bryan Hooi (National University of Singapore)
CodeMeta LearningLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
π― What it does: Proposed the J-TTL benchmark to evaluate AI agents' ability to progressively improve across multiple attempts in the same game, and designed the EvoTest framework to achieve gradient-free self-evolution during testing
EXP-Bench: Can AI Conduct AI Research Experiments?
Patrick Tser Jern Kon, Ang Chen (University Of Michigan)
CodeLarge Language ModelAgentic AITextBenchmark
π― What it does: Propose the EXP-Bench benchmark to evaluate AI agents' ability to complete the full scientific experiment process, from experimental design, implementation, execution, to conclusion.
Expanding Reasoning Potential in Foundation Model by Learning Diverse Chains of Thought Patterns
Xuemiao Zhang (Peking University), Xunliang Cai (Meituan)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningMixture of ExpertsTextBenchmarkChain-of-Thought
π― What it does: By defining the model's reasoning potential (i.e., the reciprocal of the number of attempts required to solve a problem), a core set containing high-value chain reasoning patterns is constructed, and a dual-granularity algorithm (based on reasoning pattern chains and token entropy) is designed to efficiently select training samples that conform to the core set distribution from massive CoT data, thereby significantly improving the performance of large language models on complex mathematical reasoning tasks.
Expert Divergence Learning for MoE-based Language Models
Jiaang Li (Alibaba Group), Bo Zheng (Alibaba Group)
CodeLarge Language ModelMixture of ExpertsText
π― What it does: This paper proposes the Expert Divergence Learning (EDL) pre-training strategy, which introduces a domain label-driven auxiliary loss in MoE models to encourage experts to generate separated routing distributions across different data domains, thereby achieving significant differentiation in expert functions;
Expert Merging in Sparse Mixture of Experts with Nash Bargaining
Dung Viet Nguyen, Tan Minh Nguyen
CodeOptimizationComputational EfficiencyMixture of ExpertsImageText
π― What it does: This paper proposes a Nash equilibrium-based expert merging method called NAMEx, combining complex momentum to achieve efficient convergence.
CodeKnowledge DistillationRepresentation LearningTransformerLarge Language ModelMixture of ExpertsTextMultimodalityBenchmark
π― What it does: Proposed Expert Merging and its improved version Expert Merging++, achieving unsupervised fusion of multi-domain expert models by learning hierarchical (or block-wise) coefficients and aligning hidden states and logits on unlabeled calibration data.
ExpGuard: LLM Content Moderation in Specialized Domains
Minseok Choi (KAIST), Jungmin Son (KakaoBank Corp)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextFinance RelatedChain-of-Thought
π― What it does: Proposed a content safety model for LLMs in professional fields such as finance, healthcare, and law called EXPGUARD, and constructed a large-scale professional safety dataset named EXPGUARDMIX.
Explain in Your Own Words: Improving Reasoning via Token-Selective Dual Knowledge Distillation
Minsang Kim (Korea University), Seung Jun Baek (Korea University)
CodeKnowledge DistillationTextChain-of-Thought
π― What it does: This paper proposes a bidirectional knowledge distillation framework called TSD-KD based on token selection, aiming to enhance the Chain-of-Thought generation capability of small models in complex reasoning tasks.
Junfeng Liao (University of Technology Sydney), Zhen Fang (RIKEN Center for Advanced Intelligence Project)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Propose a reasoning-based unlearning target and a method named TRU that combines a supervised reasoning loss with a gradient ascent (GA) loss, aiming to specifically remove undesirable knowledge from large language models while preserving other model capabilities.
Explaining Grokking and Information Bottleneck through Neural Collapse Emergence
Keitaro Sakamoto (University of Tokyo), Issei Sato (University of Tokyo)
CodeOptimizationExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkTransformerImage
π― What it does: By analyzing the 'neural collapse' phenomenon in the internal representations of neural networks, this paper provides a unified explanation for two late-stage training phenomena in deep learning: grokking (sudden improvement in generalization performance after training) and the information bottleneck (IB) compression phase.
Exploiting Low-Dimensional Manifold of Features for Few-Shot Whole Slide Image Classification
Conghao Xiong (Chinese University of Hong Kong), Irwin King (Chinese University of Hong Kong)
CodeClassificationRepresentation LearningMeta LearningImageBiomedical Data
π― What it does: To address few-shot pathological whole-slide image classification, the authors propose a geometry-aware Manifold Residual (MR) block, which fixes random projections to preserve the low-dimensional manifold structure of pre-trained features and learns task-specific adaptations through low-rank residual learning, significantly enhancing model generalization.
Exploratory Diffusion Model for Unsupervised Reinforcement Learning
Chengyang Ying (Tsinghua University), Jun Zhu (Tsinghua University)
CodeReinforcement LearningDiffusion modelScore-based Model
π― What it does: Propose a framework called Exploratory Diffusion Model (ExDM), which utilizes diffusion models to estimate the state distribution density in reward-free environments, generating score-based intrinsic rewards that guide agents to explore efficiently and pretrain diverse policies.
Explore-on-Graph: Incentivizing Autonomous Exploration of Large Language Models on Knowledge Graphs with Path-refined Reward Modeling
Shiqi Yan (Zhongguancun Laboratory), Yunqi Zhang (Zhongguancun Laboratory)
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackGraph Neural NetworkLarge Language ModelSupervised Fine-TuningReinforcement LearningGraphBenchmarkChain-of-Thought
π― What it does: Propose a framework called EoG that encourages large language models to autonomously explore and discover new reasoning paths on knowledge graphs through reinforcement learning
π― What it does: This paper proposes a multi-step feature alignment framework based on flow matching (Flow Matching Alignment, FMA) for cross-modal alignment and classification in few-shot vision-language models.
Exploring Interpretability for Visual Prompt Tuning with Cross-layer Concepts
Yubin Wang (Microsoft Research Asia), Cairong Zhao (Tongji University)
CodeClassificationExplainability and InterpretabilityTransformerPrompt EngineeringImage
π― What it does: Introduce interpretable cross-layer concept prototypes in visual prompt tuning to generate interpretable visual prompts corresponding to human-understandable concepts.
π― What it does: Proposed a StreamSR dataset containing 5,200 YouTube videos specifically for streaming compressed video, conducted benchmark testing of 11 real-time super-resolution models on this dataset, and subsequently designed and implemented the EfRLFN model, significantly improving the quality and efficiency of real-time super-resolution.
Exploring Synthesizable Chemical Space with Iterative Pathway Refinements
Seul Lee (Korea Advanced Institute Of Science And Technology), Arash Vahdat (NVIDIA)
CodeDrug DiscoveryTransformerFlow-based ModelBiomedical Data
π― What it does: Proposes the ReaSyn framework, achieving a synthetically feasible projection of a synthetically accessible chemical space through iterative bottom-up and top-down path generation combined with global editing.
Exploring the Potential of Encoder-free Architectures in 3D LMMs
Yiwen Tang (Shanghai AI Laboratory), Xuelong Li (Tele AI)
CodeClassificationRecognitionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPoint Cloud
π― What it does: This paper introduces an encoder-free architecture into 3D multimodal large models, proposes strategies such as LLM embedded semantic encoding and hierarchical geometric aggregation, and implements the ENEL model, completing the full training process from pre-training to instruction tuning.
ExPO-HM: Learning to Explain-then-Detect for Hateful Meme Detection
Jingbiao Mei (University of Cambridge), Bill Byrne (University of Cambridge)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodality
π― What it does: Proposed an explain-then-detect method for hateful meme detection called ExPO-HM.
π― What it does: This paper proposes EXPO, an algorithm that performs online reinforcement learning fine-tuning on expression strategies (e.g., diffusion, flow matching).
Exposing and Defending the Achilles' Heel of Video Mixture-of-Experts
Songping Wang (Nanjing University), Caifeng Shan (Nanjing University)
CodeClassificationAdversarial AttackMixture of ExpertsVideo
π― What it does: Propose temporal Lipschitz-guided attacks (TLGA, J-TLGA) targeting the router and expert components in video Mixture-of-Experts (MoE), revealing their independent and collaborative vulnerabilities, and design hierarchical joint adversarial training (J-TLAT) based on these weaknesses to enhance robustness.
ExpVid: A Benchmark for Experiment Video Understanding & Reasoning
Yicheng Xu (Shanghai AI Laboratory), Yi Wang (Shanghai AI Laboratory)
CodeTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: Proposed the ExpVid benchmark specifically designed for scientific experiment videos, covering three-tier tasks of fine-grained perception, procedural understanding, and scientific reasoning.
Extending Sequence Length is Not All You Need: Effective Integration of Multimodal Signals for Gene Expression Prediction
Zhao Yang (Renmin University of China), Bing Su (Renmin University of China)
CodeMultimodalityBiomedical Data
π― What it does: This paper proposes the Prism framework, which can predict gene expression using short sequences by effectively integrating proximal multi-modal epigenetic signals.
Extreme Weather Nowcasting via Local Precipitation Pattern Prediction
Chang hoon Song, Youngjoon Hong (Seoul National University)
CodeComputational EfficiencyTransformerVideo
π― What it does: Proposes exPreCast, a deterministic radar forecasting framework based on video Swin Transformer, aiming to efficiently and finely predict extreme precipitation events;
π― What it does: Proposed and trained the Flow-Anchored Consistency Model (FACM), achieving efficient few-step generation by combining flow matching with a hybrid objective of continuous-time consistency models.
Fair Conformal Classification via Learning Representation-Based Groups
Senrong Xu (Nanjing University), Xiaoxing Ma (Nanjing University)
CodeClassificationRepresentation LearningAuto EncoderTabularBiomedical Data
π― What it does: Propose the FAREG method, using a variational information bottleneck encoder-decoder to learn subgroups in the latent representation space, and constructing an adaptive balanced coverage synthetic prediction set based on this.
Fair Decision Utility in Human-AI Collaboration: Interpretable Confidence Adjustment for Humans with Cognitive Disparities
Jiashi Gao (Southern University of Science and Technology), Xuetao Wei (Southern University of Science and Technology)
CodeExplainability and InterpretabilityImageTextTabular
π― What it does: This paper investigates the utility unfairness problem in human-AI collaborative decision-making caused by differences in human decision-makers' cognitive abilities, and proposes a new AI confidence adjustment objective called inter-group-alignment, along with a cognition-aware multicalibration method to achieve dual goals of fairness and optimal utility.
CodeOptimizationSafty and PrivacyGraph Neural NetworkGraph
π― What it does: To address the fairness issue of graph neural networks in scenarios with missing sensitive attributes (especially adversarial missingness), we propose the BFtS (Better Fair than Sorry) three-player adversarial framework, which improves the balance between model fairness and accuracy through worst-case sensitive attribute imputation.
FaithCoT-Bench: Benchmarking Instance-Level Faithfulness of Chain-of-Thought Reasoning
Xu Shen (Jilin University), Tianlong Chen (University of North Carolina at Chapel Hill)
CodeExplainability and InterpretabilityTextBenchmarkChain-of-Thought
π― What it does: Propose the FAITHCOT-BENCH framework, construct the expert-annotated FINE-COT dataset, and systematically evaluate instance-level CoT untrustworthiness detection methods.
Faithful Bi-Directional Model Steering via Distribution Matching and Distributed Interchange Interventions
Yuntai Bao (Zhejiang Normal University), Jianwei Yin (Zhejiang Normal University)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: This paper proposes a lightweight, interpretable bidirectional control method for LLMs based on distribution matching and distributed interactive intervention (CDAS).
Faithfulness Under the Distribution: A New Look at Attribution Evaluation
Zhiyu Zhu (University of Technology Sydney), Jianlong Zhou (University of Technology Sydney)
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerDiffusion modelScore-based ModelImageStochastic Differential Equation
π― What it does: To assess the credibility of explanation methods, a novel distribution-aware evaluation framework named FUD is proposed. It utilizes a score diffusion model to reconstruct masked regions into data distribution samples while preserving important features, thereby eliminating biases caused by out-of-distribution (OOD) scenarios in traditional methods.
CodeObject DetectionAnomaly DetectionExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageMultimodalityChain-of-Thought
π― What it does: Developed FakeXplainer, an explainable AI-generated image detection system capable of detecting, locating, and explaining forged images.
Falcon: Fast Proximal Linearization of Normalized Cuts for Unsupervised Image Segmentation
Xiao Zhang (University of Pennsylvania), Konrad Kording (University of Pennsylvania)
CodeSegmentationImageBenchmark
π― What it does: Propose Falcon, a primal gradient solver based on first-order approximation, directly optimizing the discrete K-way NCut objective, eliminating recursive bipartitioning and spectral relaxation processes, achieving unsupervised image segmentation.
π― What it does: Proposed FANTASYWORLD, a unified forward network that can simultaneously generate video frames and implicit 3D representations in a single inference.
π― What it does: Designed and implemented a fast, high-performance multi-template autoregressive visual tracking framework called FARTrack for real-time tracking.
Yifei Wang (Alibaba Group), Julian McAuley (University of California San Diego)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelTextSequential
π― What it does: This paper proposes FASA, a training-free, query-aware KV cache compression framework that dynamically predicts and removes unimportant tokens by leveraging the sparsity of frequency blocks (FC) in RoPE, followed by performing full attention computation only on the retained, limited number of tokens;
Fast and Stable Riemannian Metrics on SPD Manifolds via Cholesky Product Geometry
Ziheng Chen, Nicu Sebe (University of Trento)
CodeClassificationRecognition
π― What it does: Leverage Cholesky decomposition to reveal the product structure of SPD matrix space, and design two new Riemannian metricsβPower-Cholesky Metric (PCM) and Bures-Wasserstein-Cholesky Metric (BWCM)βto provide closed-form, fast, and numerically stable geometric operations for SPD neural networks.
Fast Estimation of Wasserstein Distances via Regression on Sliced Wasserstein Distances
Khai Nguyen (University of Texas at Austin), Nhat Ho (University of Texas at Austin)
CodeOptimizationComputational EfficiencyPoint CloudBiomedical Data
π― What it does: Propose a framework for fast estimation of Wasserstein distance by regressing Wasserstein distance on sliced Wasserstein distance (SW);
Fast FrankβWolfe Algorithms with Adaptive Bregman Step-Size for Weakly Convex Functions
Shota Takahashi (University of Tokyo), Akiko Takeda (University of Tokyo)
CodeOptimizationComputational Efficiency
π― What it does: Proposed an adaptive Bregman step size Frank-Wolfe algorithm applicable to L-smooth adaptable (L-smad) and weakly convex objective functions, providing linear convergence analysis for both convex and non-convex scenarios.
π― What it does: Propose an end-to-end feed-forward 3D avatar reconstruction framework called FastAvatar, which can rapidly generate high-quality animatable 3D Gaussian Splatting (3DGS) models from any number of images or video frames, and supports progressive incremental reconstruction.
Fastcar: Cache Attentive Replay for Fast Auto-Regressive Video Generation on the Edge
Xuan Shen (Northeastern University), Jiuxiang Gu (Adobe)
CodeGenerationTransformerVideo
π― What it does: Propose the FastCar framework, which utilizes temporal redundancy in autoregressive video generation by caching and replaying the MLP output of the previous frame to significantly reduce computational costs; and implement a dynamic resource scheduling hardware accelerator on FPGA;
Faster Gradient Methods for Highly-smooth Stochastic Bilevel Optimization
Lesi Chen (Tsinghua University & Shanghai Qizhi Institute), Jingzhao Zhang (Tsinghua University & Shanghai Qizhi Institute)
CodeOptimizationText
π― What it does: This paper studies fully gradient-based stochastic bilevel optimization algorithms and proposes the FSA^p method, which achieves faster convergence speeds by approximating the supergradient using high-order finite differences.
π― What it does: Significantly accelerate generation by adaptively skipping redundant denoising steps during the inference of flow matching models, using a multi-armed bandit to determine when to approximate.
CodeOptimizationComputational EfficiencyLarge Language ModelReinforcement LearningText
π― What it does: Significantly improved the throughput during the generation phase by introducing concurrent-aware speculative decoding and online draft model learning in GRPO training, thereby accelerating the overall training process.
FATE: A Formal Benchmark Series for Frontier Algebra of Multiple Difficulty Levels
Jiedong Jiang (Westlake University), Bin Dong (Peking University)
CodeAI Code AssistantLarge Language ModelTextBenchmark
π― What it does: This study proposes the FATE (Formal Algebra Theorem Evaluation) benchmark series, focusing on constructing two new difficulty levelsβFATE-H (undergraduate to graduate level) and FATE-X (doctoral qualifying exams and above)βto assess the capability of large language models in formal algebraic proofs; and reveals the bottlenecks in the formalization process through a two-phase evaluation (first natural language reasoning, then conversion to Lean code).
Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs
Shreyas Singh (Fractal AI Research), Pradeep Moturi (Fractal AI Research)
CodeGenerationData SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Built a two-model proxy system Fathom-DeepResearch, including a specialized search inference model Fathom-Search-4B and a structured report generation model Fathom-Synthesizer-4B;
FeDaL: Federated Dataset Learning for General Time Series Foundation Models
Shengchao Chen (University of Technology Sydney), Jing Jiang (University of Technology Sydney)
CodeAnomaly DetectionFederated LearningTransformerTime Series
π― What it does: Proposes FeDaL, a federated learning framework for time series foundation models, capable of training models from scratch that can achieve cross-domain generalization on distributed and heterogeneous datasets.
Federated Learning with Profile Mapping under Distribution Shifts and Drifts
Mohan Li (UniversitΓ della Svizzera italiana), Marc Langheinrich (UniversitΓ della Svizzera italiana)
CodeFederated LearningSafty and PrivacyImageBiomedical Data
π― What it does: Propose the FEROMA framework, which utilizes differential privacy distribution summaries to map federated learning distributions, dynamically selects aggregation strategies, handles distribution shifts and drifts during both training and testing phases, and supports model allocation for unlabelled clients.
π― What it does: Proposed and implemented the first framework for open-set semi-supervised federated learning (OSSFL), FedOpenMatch, which can simultaneously handle samples from both known classes and unknown classes in the unlabeled data distributed across clients;
π― What it does: Studied the fairness issue in data-free robustness distillation and proposed the FERD framework to enhance the robust fairness of student models
Fewer Battles, More Gain: An Information-Efficient Framework for Arena-based LLM Evaluation
Zirui Liu (University of Science and Technology of China), Shijin Wang (iFLYTEK Co., Ltd)
CodeOptimizationComputational EfficiencyText
π― What it does: Propose an adaptive model pair selection framework based on Fisher information, using A-optimality and D-optimality to iteratively select the most informative model pairs, significantly reducing the number of required matches for evaluation;
π― What it does: Proposed a single-step high-fidelity motion deblurring diffusion model called FideDiff, achieving one-time restoration during inference.
Fine-Grained Activation Steering: Steering Less, Achieving More
Zijian Feng (Nanyang Technological University), Kezhi Mao (Nanyang Technological University)
CodeTransformerLarge Language ModelContrastive LearningText
π― What it does: Investigated and demonstrated the heterogeneity of block-level activation in large language models (LLMs), proposing a fine-grained atomic unit (AU)-level activation modulation method called AUSteer;
Fine-Grained Class-Conditional Distribution Balancing for Debiased Learning
Miaoyun Zhao (Dalian University of Technology), Qiang Zhang (Dalian University of Technology)
CodeDomain AdaptationImageText
π― What it does: Proposes an unbiased annotation bias exploration and fine-grained class-conditional distribution balancing (FG-CCDB) method to achieve robust group-level generalization in the presence of spurious correlations.
Fine-R1: Make Multi-modal LLMs Excel in Fine-Grained Visual Recognition by Chain-of-Thought Reasoning
Hulingxiao He (Peking University), Yuxin Peng (Peking University)
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityChain-of-Thought
π― What it does: Designed and trained a multimodal large language model, Fine-R1, for fine-grained visual recognition (FGVR), enabling it to identify both known and unknown subcategories with a small number of training samples.
Fine-tuning Behavioral Cloning Policies with PreferenceβBased Reinforcement Learning
Mael Macuglia (University of Zurich), Giorgia Ramponi (University of Zurich)
CodeOptimizationReinforcement Learning from Human FeedbackSupervised Fine-TuningReinforcement LearningSequential
π― What it does: Integrate offline expert demonstrations with online preference feedback, proposing the BRIDGE algorithm. By constructing a Hellinger confidence ball to constrain online search, it achieves safe and efficient behavior cloning fine-tuning.
Wanli Yang (Institute of Computing Technology, Chinese Academy of Sciences), Fei Sun (Institute of Computing Technology, Chinese Academy of Sciences)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Re-evaluated and corrected the implementation of fine-tuning in model editing, transforming it into a standard breadth-first (BF) training pipeline combined with local parameter fine-tuning, and proposed the LocFT-BF method;
π― What it does: Propose a method called Quantized Zeroth-order Optimization (QZO) for fine-tuning quantized large language models (LLMs), achieving extreme memory compression by only updating the quantization scale and eliminating gradients and optimizer states.
FingerTip 20K: A Benchmark for Proactive and Personalized Mobile LLM Agents
Qinglong Yang (Tsinghua University), Yong Li (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringImageTextSequentialBenchmark
π― What it does: This paper proposes the FingerTip 20K benchmark to evaluate the capabilities of mobile large language model agents in proactive task suggestion and personalized task execution, and constructs the dataset by collecting 20,000 real user multi-step Android interaction examples.
Fixing the Broken Compass: Diagnosing and Improving Inference-Time Reward Modeling
Jiachun Li (University of Chinese Academy of Sciences), Jun Zhao (University of Chinese Academy of Sciences)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBenchmark
π― What it does: Proposed a reward model-based reasoning acceleration algorithm called CRISP, which enhances LLM inference effectiveness through answer clustering, reward aggregation, and step-by-step prefix guidance.
Flash-Searcher: Fast and Effective Web Agents via DAG-Based Parallel Execution
Tianrui Qin (OPPO AI Center), Wangchunshu Zhou (OPPO Reseach Institute)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelAgentic AITextTabularBenchmark
π― What it does: Propose the Flash-Searcher framework, which utilizes a DAG structure to decompose complex tasks into parallelizable subtasks, achieving efficient Web Agent reasoning through multi-path parallel inference and tool calls.
FlashDLM: Accelerating Diffusion Language Model Inference via Efficient KV Caching and Guided Diffusion
Zhanqiu Hu (Cornell University), Udit Gupta (Cornell University)
CodeComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
π― What it does: Propose two training-agnostic acceleration methods: FreeCache uses KV caching to reduce redundant computations in DLM inference, while Guided Diffusion employs a lightweight AR model to provide consistency guidance during the diffusion process, thereby reducing denoising steps and improving parallel generation efficiency.
FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spatiotemporal Token Merging
Ziyang Fan (Harbin Institute of Technology), Zhuotao Tian (Harbin Institute of Technology)
CodeCompressionComputational EfficiencyTransformerVision Language ModelVideoMultimodality
π― What it does: Proposed FlashVID, a training-agnostic, plug-and-play acceleration framework that leverages attention + diversity screening and tree-structured spatiotemporal merging to significantly compress video visual tokens, achieving efficient inference.
Flatter Tokens are More Valuable for Speculative Draft Model Training
Jiaming Fan (Southeast University), Xu Yang (Southeast University)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
π― What it does: Studied speculative decoding for accelerating large language model inference, proposing a data selection method based on 'flatness' to enhance the training efficiency of the draft model.