Page 4 - Paper Library - AiPaper

Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editing

Published:9/24/2025

Detoxification of Large Language ModelsAutoregressive Reward Guided Representation EditingToxicity Transition ModelingDynamic Editing StrategyEffectiveness and Efficiency Optimization

This study introduces the Autoregressive Reward Guided Representation Editing (ARGRE) framework for detoxifying large language models. ARGRE models toxicity transitions in the latent space, identifies nontoxic directions, and interpolates between toxic and nontoxic representati

UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion

Published:1/20/2025

Exposure Fusion TechniqueHigh Dynamic Range ImagingUltra High Dynamic Range DatasetGuided Inpainting for ImagesNatural Tone Mapping Generation

UltraFusion is the first exposure fusion technique that merges images with 9stop differences, modeling it as a guided inpainting problem. It uses underexposed images for filling overexposed highlights and generates natural tone mapping, outperforming existing methods.

De-collapsing User Intent: Adaptive Diffusion Augmentation with Mixture-of-Experts for Sequential Recommendation

Diffusion Model-based Recommendation SystemsUser Intent Reconstruction in Sequential RecommendationSparse Data Augmentation MethodsMixture-of-Experts ArchitectureAdaptive Diffusion Augmentation Framework

The ADARec framework addresses data sparsity in sequential recommendation by utilizing a MixtureofExperts architecture to decouple coarse and finegrained user intents, effectively reconstructing intent hierarchies. Experiments show it outperforms existing methods on standard b

扇形覆冰特高压八分裂导线舞动特性分析

Analysis of Motion Characteristics of Sector Ice-Covered ConductWind Tunnel Experiments for UHV Transmission LinesABAQUS Finite Element ModelingAerodynamic Parameter Study of Transmission Lines in Heavy Ice RAnalysis of Wind Speed Influence on Conductor Motion

This study combines wind tunnel experiments and numerical simulations to analyze the galloping characteristics of sectorshaped iced eightbundle conductors in heavy ice regions, establishing a finite element model to assess aerodynamic parameters and the impact of wind condition

UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models

Published:5/18/2025

Benchmark for Large Language Model EditingNeighborhood Multi-hop Chain Sampling AlgorithmOpen-Domain Knowledge GraphModel Editing Performance EvaluationKnowledge Editing Sample Generation

The paper introduces UniEdit, a unified benchmark for large language model editing using opendomain knowledge. It employs a Neighborhood Multihop Chain Sampling algorithm to ensure comprehensive evaluation and coverage, revealing strengths and weaknesses across various models f

Rethinking Memory in AI: Taxonomy, Operations, Topics, and Future Directions

Published:5/2/2025

Memory Representation TaxonomyMemory Operations DynamicsMemory Systems in Large Language ModelsLong-Context Memory ResearchPersonalized Memory in LLM-based Agents

This survey analyzes memory mechanisms in AI, introducing a unified taxonomy categorizing memory as parametric and contextual, and defining six key operations. These operations inform research directions on longterm and multisource memory, providing a structured dynamic perspec

Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning

Published:2/5/2025

Adaptive Attention SparsityAcceleration of Long-Context Large Language ModelsTop-p Sampling Based Sparse AttentionDynamic Budgeting Attention MechanismSelf-Attention Operation Acceleration

The study introduces the framework that employs sampling for adaptive attention sparsity to accelerate longcontext LLMs. Results demonstrate up to 98% token pruning, achieving 15.4x speedup in selfattention and 3.9x in endtoend latency without compromising

PvP: Data-Efficient Humanoid Robot Learning with Proprioceptive-Privileged Contrastive Representations

Published:12/15/2025

Humanoid Whole-Body ControlHumanoid Robot LearningContrastive Learning FrameworkState Representation Learning MethodsData-Efficient Reinforcement Learning

The paper introduces the PvP framework, addressing sample inefficiency in humanoid robot control by leveraging the complementarity of proprioceptive and privileged states. It improves learning efficiency without manual data augmentation, significantly enhancing performance in vel

DiffTMR: Diffusion-based Hierarchical Alignment for Text-Molecule Retrieval

Published:10/25/2025

Diffusion Model for Molecular RetrievalHierarchical Alignment ApproachText-Molecule Retrieval FrameworkDynamic Perturbation Embedding MechanismCross-Modal Hierarchical Alignment

DiffTMR is an innovative textmolecule retrieval framework that reframes retrieval as a reverse denoising process, addressing traditional methods' limitations in outofdistribution detection and diversity maintenance. By integrating hierarchical diffusion alignment and dynamic p

Wavelet Enhanced Adaptive Frequency Filter for Sequential Recommendation

Published:11/10/2025

Sequential Recommender SystemsAdaptive Frequency FilteringWavelet Feature EnhancementDynamic User Preference ModelingFrequency Domain Analysis Methods

This paper introduces the Wavelet Enhanced Adaptive Frequency Filter (WEARec) for sequential recommendation, enhancing traditional frequencydomain methods by dynamically filtering and wavelet feature enhancement, effectively improving performance and efficiency in capturing user

ReBot: Scaling Robot Learning with Real-to-Sim-to-Real Robotic Video Synthesis

Published:3/16/2025

Vision-Language-Action ModelRobotic Video SynthesisReal-to-Sim-to-Real ApproachRobot Dataset ScalingRobotic Manipulation Tasks

ReBot enhances robot learning by proposing a realtosimtoreal video synthesis method, addressing data scaling challenges. It replays real robot movements in simulators and combines them with inpainted realworld backgrounds, significantly improving VLA model performance with s

SKILL-IL: Disentangling Skill and Knowledge in Multitask Imitation Learning

Published:5/6/2022

Multitask Imitation LearningSkill and Knowledge DisentanglementTransferable Skill LearningPolicy Network Memory DisentanglementRobot Navigation Tasks

The SKILLIL framework disentangles skill and knowledge in multitask imitation learning, improving training efficiency and generalization. It achieved a 30% increase in success rates in simulated environments and was validated in realworld navigation tasks.

Exploring Perception-Based Techniques for Redirected Walking in VR: A Comprehensive Survey

Published:5/22/2025

Perception-Based Redirected Walking TechniquesUser Experience in Virtual RealityExploration of Virtual EnvironmentsClassification of RDW AlgorithmsSurvey of Virtual Reality Techniques

This paper surveys perceptionbased redirected walking techniques in VR, reviewing 232 papers and analyzing 165. It introduces a new taxonomy categorizing RDW algorithms into Gains, Gain Application, Target Orientation Calculation, and Enhancements, emphasizing the importance of

Redirected Walking for Exploring Immersive Virtual Spaces With HMD: A Comprehensive Review and Recent Advances

Published:5/31/2022

Redirected Walking TechniquesExploration of Immersive Virtual SpacesVirtual-Physical Movement MappingRedirection Controller MethodsUser Movement Adjustment Strategies

This paper reviews Redirected Walking (RDW) techniques, addressing how to achieve immersive virtual experiences within limited physical space. It categorizes redirection manipulations, discusses controller methods, and incorporates emerging technologies like deep learning, summar

Predictive multiuser redirected walking using artificial potential fields

Published:8/8/2024

Multi-User Redirected WalkingArtificial Potential FieldsPredictive Redirected Walking SystemsClothoid Trajectory GenerationUser Experience in Virtual Environments

This paper introduces two novel predictive redirected walking systems using clothoidbased algorithms, effectively addressing multiuser navigation in limited physical spaces, and demonstrating improved user experience over traditional reactive methods.

BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models

Published:6/10/2025

Vision-Language-Action Model3D Manipulation LearningInput-Output Alignment2D Heatmap PredictionSample Efficiency Improvement

BridgeVLA introduces a novel visionlanguageaction model for 3D manipulation, addressing inefficiencies in existing models. By projecting 3D data to 2D images and using heatmaps for action prediction, it achieves stateoftheart performance on various benchmarks.

Prompting Science Report 4: Playing Pretend: Expert Personas Don't Improve Factual Accuracy

Published:1/1/2025

Impact of Expert Personas on Model PerformancePerformance Evaluation on Multiple-Choice QuestionsComparison of Domain-Specific and Low-Knowledge PersonasGPQA Diamond and MMLU-Pro BenchmarkingRelation between AI Model Performance and Persona Prompts

The study investigates whether assigning expert personas to large language models improves performance on difficult objective questions. Findings reveal no significant accuracy gains from expert personas, while mismatched and lowknowledge personas often degrade model performance

Large Language Models meet Collaborative Filtering: An Efficient All-round LLM-based Recommender System

Published:4/17/2024

LLM-based Recommendation SystemsCollaborative Filtering Recommender SystemsCold-Start Recommendation OptimizationCross-Domain Recommendation SystemUser/Item Embedding Generation

The ALLMRec system combines collaborative knowledge with large language models to excel in both cold and warm start scenarios, enhancing user experience while being modelagnostic and efficient.

Real-time Photorealistic Dynamic Scene Representation and Rendering with 4D Gaussian Splatting

Published:10/17/2023

4D Gaussian Splatting RepresentationDynamic Scene ReconstructionReal-Time RenderingSpatiotemporal ModelingMultiview Scene Synthesis

This paper introduces a 4D Gaussian Splatting method for dynamic scene reconstruction and rendering, addressing the challenge of generating highquality 3D scenes from 2D images. By viewing spacetime as a whole, it models geometry and dynamic appearance efficiently, outperforming

ReCamDriving: LiDAR-Free Camera-Controlled Novel Trajectory Video Generation

Published:12/3/2025

Camera-Controlled Trajectory Video Generation3D Geometric GuidanceMonocular Video Multi-Trajectory SupervisionParaDrive DatasetTwo-Stage Training Paradigm

This paper presents ReCamDriving, a visionbased framework for generating videos from novel trajectories using camera control. By leveraging dense 3D Gaussian Splatting as geometric guidance, it employs a twostage training to enhance controllability and structural consistency, a

61 - 80 / 972

Papers