Page 17 - Paper Library - AiPaper

Learning Decomposed Contextual Token Representations from Pretrained and Collaborative Signals for Generative Recommendation

Published:8/23/2025

Generative Recommendation SystemsContextual Token Representation LearningLarge Language Model OptimizationSequence-to-Sequence ModelingUser Interaction Modeling

The DECOR framework addresses limitations in generative recommenders, enhancing token adaptability while preserving pretrained semantics. It employs contextualized token composition and decomposed embedding fusion, demonstrating superior performance on real datasets compared to s

User-LLM: Efficient LLM Contextualization with User Embeddings

Published:2/21/2024

User Embedding-Based LLM ContextualizationSelf-Supervised User Behavior EncoderMovie Recommendation DatasetResponse Generation from User TimelineCross-Attention LLM Integration

The UserLLM framework employs selfsupervised user embeddings for direct contextualization with LLMs. Through crossattention, it dynamically adapts responses to user behavior, achieving 78.1X speedup and a 16.33% performance boost on datasets like MovieLens.

REGEN: Learning Compact Video Embedding with (Re-)Generative Decoder

Published:3/12/2025

Diffusion Model Video EmbeddingCompact Video EncodingGenerative Video ReconstructionAccelerated Training for Video Generation ModelsTemporal Compression Video Embedding

The study introduces a novel video embedding method focusing on synthesizing visually plausible reconstructions, achieving higher compression ratios. Utilizing an encodergenerator framework with a diffusion transformer, it shows superior encodingdecoding performance with up to

PersonaBench: Evaluating AI Models on Understanding Personal Information through Accessing (Synthetic) Private User Data

Published:2/28/2025

Synthetic User Data GenerationPersonalized AI Model EvaluationUnderstanding Private User InformationRetrieval-Augmented Generation (RAG) MethodEnhancing User Personalization Capabilities

This study introduces a synthetic data generation pipeline for creating realistic user profiles and private documents, leading to the PersonaBench benchmark. It highlights poor performance of current retrievalaugmented models in extracting personal information, underscoring the

Robust model predictive control for heat exchanger network

Published:8/28/2014

Robust Model Predictive ControlHeat Exchanger Network ControlNonlinear Control StrategiesSimulation Experiments in MATLAB/SimulinkControl Performance Optimization

This paper introduces a Robust Model Predictive Control (RMPC) strategy for optimizing heat exchanger network operation. Simulation experiments on three series countercurrent exchangers in MATLAB/Simulink demonstrate RMPC's effectiveness in reducing cooling medium consumption co

Planning with Diffusion for Flexible Behavior Synthesis

Published:5/20/2022

Diffusion Model PlanningModel-Based Reinforcement LearningTrajectory OptimizationLong-Horizon Decision MakingBehavior Synthesis

This paper presents a novel modelbased reinforcement learning approach that combines diffusion probabilistic modeling with trajectory optimization, enhancing consistency between modeling and decisionmaking. It demonstrates effective longhorizon decisionmaking and flexibility

Learnable Item Tokenization for Generative Recommendation

Published:5/12/2024

LLM-based Generative Recommendation SystemsLearnable Item TokenizationContrastive Learning-based Recommendation AlgorithmsResidual Quantized Variational AutoencoderRanking-Guided Generation Loss

The paper introduces LETTER, a learnable tokenizer addressing challenges of transforming recommendation data into LLMs’ language space. By integrating hierarchical semantics, collaborative signals, and code assignment diversity, its experimental validation on three datasets demon

027

HiD-VAE: Interpretable Generative Recommendation via Hierarchical and Disentangled Semantic IDs

Published:8/7/2025

Generative Recommendation SystemsHierarchical Semantic IDsInterpretable Generative RecommendationDisentangled Representation LearningUniqueness Loss Mechanism

HiDVAE is a proposed framework that enhances generative recommendation by learning hierarchically disentangled item representations, addressing traditional methods' flatness and entanglement issues, thereby improving recommendation accuracy and diversity.

TokenRec: Learning to Tokenize ID for LLM-based Generative Recommendation

Published:6/15/2024

LLM-based Recommendation SystemsGenerative Recommendation SystemsUser-Item ID TokenizationMasked Vector-Quantized TokenizerCapturing High-Order Collaborative Knowledge for LLMs

TokenRec is introduced as a novel framework for enhancing LLMbased recommendation systems by effectively tokenizing user and item IDs. Featuring the Masked VectorQuantized Tokenizer and generative retrieval, it captures highorder collaborative knowledge, improving recommendati

Omnidirectional 3D Scene Reconstruction from Single Image

Single Image 3D Scene ReconstructionDiffusion Model for 3D ReconstructionGeometric Consistency Optimization3D Gaussian Splatting RepresentationOmnidirectional Scene Reconstruction

The paper proposes Omni3D, a novel method for omnidirectional 3D scene reconstruction from a single image. By iteratively optimizing generated views and poses, it minimizes 3D reprojection errors, enhancing geometric consistency. Experiments show Omni3D significantly outperforms

MoE-Loco: Mixture of Experts for Multitask Locomotion

Published:3/11/2025

Multitask Locomotion LearningMixture of Experts FrameworkQuadrupedal and Bipedal LocomotionGradient Conflict MitigationRobot Task Migration and Skill Composition

The paper presents MoELoco, a Mixture of Experts framework for legged robots, enabling a single policy to navigate diverse terrains while mitigating gradient conflicts in multitask reinforcement learning, enhancing training efficiency and performance.

Inductive Generative Recommendation via Retrieval-based Speculation

Published:10/4/2024

Generative Recommendation SystemsTraining-Free Acceleration MethodsOnline Recommendation System OptimizationSequential Recommender SystemsImage Generation

The paper introduces , a retrievalbased inductive generative recommendation framework that addresses the limitations of generative models in recommending unseen items by utilizing a drafter model for candidate generation and a generative model for verification, enhancing

LLM-Aligned Geographic Item Tokenization for Local-Life Recommendation

Published:11/18/2025

LLM-based Recommendation SystemsGeographic Item TokenizationLocal-Life RecommendationReinforcement Learning Geographic AlignmentHierarchical Geographic Item Tokenization

The LGSID framework enhances locallife recommendation by integrating RLbased geographic alignment and hierarchical item tokenization to capture spatial relationships, outperforming existing models in empirical studies.

Pre-training Generative Recommender with Multi-Identifier Item Tokenization

Published:4/6/2025

Generative Recommendation SystemsMulti-Identifier Item TokenizationCurriculum Recommender Pre-TrainingRQ-VAE as TokenizerLow-Frequency Item Semantic Modeling

The MTGRec framework enhances generative recommender pretraining through multiidentifier item tokenization, using RQVAE for multiple identifier association and a curriculum learning scheme to improve semantic modeling for lowfrequency items and token diversity.

Optimized Product Quantization for Approximate Nearest Neighbor Search

Published:6/1/2013

Optimized Product QuantizationApproximate Nearest Neighbor SearchHigh-Dimensional Vector EncodingQuantization Distortion MinimizationParametric and Non-Parametric Methods

This study presents an optimized product quantization method to enhance the accuracy of approximate nearest neighbor search (ANN). By minimizing quantization distortions, two optimization approaches are proposed: a nonparametric method and a parametric method assuring optimal so

Understanding Negative Sampling in Knowledge Graph Embedding

Published:1/31/2021

Knowledge Graph EmbeddingNegative Sampling MethodsNegative Sample GenerationKnowledge Representation in Recommendation SystemsLink Prediction and Node Classification

This paper discusses negative sampling in knowledge graph embedding, highlighting its importance in training. It categorizes negative sampling methods into static, dynamic, and custom clustering approaches, offering new insights to enhance knowledge representation in recommendati

Qwen2.5 Technical Report

Published:12/20/2024

Qwen 2.5 Large Language ModelMultistage Reinforcement LearningSupervised Fine-Tuning MethodsHuman Preference EnhancementLarge-Scale Pre-Training Datasets

The Qwen2.5 technical report introduces a new large language model series, expanding the pretraining dataset to 18 trillion tokens. It employs over 1 million finetuned samples and multistage reinforcement learning to enhance human preferences, demonstrating superior performanc

Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length

Published:12/4/2025

Real-Time Audio-Driven Avatar GenerationEfficient Inference of Diffusion ModelsLow-Latency Streaming GenerationTemporal Consistency Enhancement MechanismLarge-Scale Parameter Diffusion Model

Live Avatar' is an innovative algorithmsystem framework for highfidelity, infinitelength audiodriven avatar generation. It employs a 14billionparameter diffusion model and introduces a timestepforcing pipeline for lowlatency streaming, enhancing temporal consistency and m

Complex-valued Neural Operator for Solving 2D Wave Equation Based on Graph Neural Network

Published:1/1/2025

Complex-Valued Neural OperatorApplication of Graph Neural Networks2D Wave Equation SolvingGreen's Function MethodEM Simulation Acceleration

This work introduces a complexvalued neural operator (CVNeuralOp) using graph neural networks to solve the 2D wave equation. Inspired by Green's function method, it shows adaptability to various domain shapes and grid densities while outperforming the method of moments in compu

NSNO: Neumann Series Neural Operator for Solving Helmholtz Equations in Inhomogeneous Medium

Published:1/25/2024

Neumann Series Neural OperatorHelmholtz Equation SolvingEmbedded U-Net ArchitectureDeep Learning for PDEsInverse Scattering Problem Model

The Neumann Series Neural Operator (NSNO) is introduced for learning the solution operator of the Helmholtz equation, achieving 60% lower relative L2error and 50% reduced computational cost, especially beneficial in high wavenumber scenarios and inverse scattering problems.

321 - 340 / 980

Papers