Paper Library - AiPaper

Large Reasoning Embedding Models: Towards Next-Generation Dense Retrieval Paradigm

RL Training for Large Language ModelsSequence Policy OptimizationLarge Language Model Fine-TuningLLM-based Recommendation SystemsRetrieval-Augmented Reasoning

LREM integrates reasoning before generating query embeddings, enhancing deep semantic understanding and retrieval accuracy for difficult queries. Trained via supervised finetuning and reinforcement learning, it’s deployed on China’s largest ecommerce platform.

Understanding Generative Recommendation with Semantic IDs from a Model-scaling View

Status:completed

Generative Recommendation SystemsSemantic ID QuantizationLLM-based Recommendation SystemsModel-Scaling AnalysisCollaborative Filtering and Semantic Fusion

This study reveals scaling bottlenecks in semantic IDbased generative recommendation due to limited encoding capacity. Directly using large language models outperforms by up to 20%, challenging assumptions about LLMs’ effectiveness in collaborative filtering and suggesting a pro

Divide and Conquer: Grounding LLMs as Efficient Decision-Making Agents via Offline Hierarchical Reinforcement Learning

Status:completed

RL Training for Large Language ModelsSequence Policy OptimizationLong-Context ModelingLLM-guided motion planningHierarchical Reinforcement Learning Framework

GLIDER employs offline hierarchical reinforcement learning to transform LLMs into efficient decision agents by decomposing longhorizon tasks into structured subtasks, enhancing exploration and adaptability with strong transferability, validated on ScienceWorld and ALFWorld bench

DeepSeek-OCR: Contexts Optical Compression

Status:completed

Long-Context CompressionOptical 2D MappingVision EncoderText Optical Character Recognition (OCR)Large-Scale Document Training Data Generation

DeepSeekOCR uses 2D optical mapping for efficient longtext compression, achieving 97% OCR accuracy at 10x compression and 60% at 20x. It surpasses existing OCR models with fewer vision tokens and enables largescale training data generation.

ExGRPO: Learning to Reason from Experience

Status:completed

RL Training for Large Language ModelsReinforcement Learning for Math ReasoningSequence Policy Optimization

ExGRPO identifies key experience metrics to prioritize valuable reasoning data, improving reinforcement learning efficiency and reasoning performance in large language models, with stable training across diverse model scales.

DrivingWorld: Constructing World Model for Autonomous Driving via Video GPT

Status:completed

Driving World ModelsLong-Horizon Video GenerationVideo Generation ModelsSpatio-Temporal Fusion Mechanisms

DrivingWorld introduces a video GPT model with spatiotemporal fusion, combining nextstate and nexttoken prediction to enhance autonomous driving video generation, achieving over 40 seconds of highfidelity, coherent video with novel masking and reweighting to reduce drift.

Spacetime-GR: A Spacetime-Aware Generative Model for Large Scale Online POI Recommendation

Status:completed

Large-Scale Online POI RecommendationGenerative Recommendation ModelSpatiotemporal Encoding ModuleMultimodal POI EmbeddingsGeographic-Aware Hierarchical Indexing

This paper introduces SpacetimeGR, a spatiotemporally aware generative model for largescale POI recommendation, employing geographic hierarchical indexing, a novel spatiotemporal encoding module, and multimodal POI embeddings; experiments demonstrate its superiority and success

AutoComm: A Framework for Enabling Efficient Communication in Distributed Quantum Programs

Status:completed

Distributed Quantum Computing Compiler FrameworkQuantum Communication Pattern OptimizationBurst Quantum Communication Detection and OptimizationQuantum Program OptimizationCommunication Overhead Reduction in Distributed Quantum Programs

AutoComm identifies burst communication patterns in distributed quantum programs and optimizes them, cutting communication overhead and latency by 72.9% and 69.2% respectively, enhancing distributed quantum computing efficiency.

Catalog-Native LLM: Speaking Item-ID Dialect with Less Entanglement for Recommendation

Status:completed

LLM-based Recommendation SystemsSequential Recommender SystemsMixture-of-Experts ModelItem-ID Representation LearningMultimodal Recommendation Signal Fusion

IDIOMoE models itemID interactions as a native language dialect, splitting experts within pretrained LLMs to reduce interference between text and item signals, enhancing recommendation accuracy and generalization across datasets.

Margin Matching Preference Optimization: Enhanced Model Alignment with Granular Feedback

Status:completed

Large Language Model Fine-TuningSequence Policy OptimizationRL Training for Large Language ModelsLLM Confidence Calibration

MMPO enhances LLM alignment by incorporating relative quality margins into training using soft target probabilities based on the BradleyTerry model, outperforming baselines on human and AI feedback datasets and achieving stateoftheart results on RewardBench with improved robu

A Split-Merge DP-means Algorithm to Avoid Local Minima

Status:completed

Extension of DP-means Clustering AlgorithmSplit-Merge TechniqueLocal Minima AvoidanceNonparametric Bayesian ClusteringHard Clustering Approximation

This paper extends DPmeans with a splitmerge technique to avoid local minima by dynamically splitting dense clusters, enabling nearoptimal cluster counts with improved robustness demonstrated on multiple datasets.

Revisiting k-means: New Algorithms via Bayesian Nonparametrics

Status:completed

Bayesian Nonparametric Clustering AlgorithmsDirichlet Process Mixture Modelsk-means Algorithm ImprovementsMulti-Dataset ClusteringSpectral Clustering and Normalized Cut

This paper revisits kmeans via Bayesian nonparametrics, proposing DPmeans to automatically determine cluster count and optimize a kmeanslike objective. Extensions include hierarchical clustering across datasets and spectral and normalized cut methods for improved performance.

Glyph: Scaling Context Windows via Visual-Text Compression

Status:completed

Long-Context ModelingVision-Language ModelsVisual-Text CompressionLLM Reasoning Capacity EnhancementLLM-guided motion planning

Glyph compresses long texts into images processed by visionlanguage models, achieving 34× token compression with maintained accuracy and improved efficiency, enabling milliontoken context scaling and enhancing multimodal document understanding.

CutQC: Using Small Quantum Computers for Large Quantum Circuit Evaluations

Status:completed

Hybrid Quantum Computing FrameworkLarge-Scale Quantum Circuit CuttingQuantum Circuit Evaluation on NISQ DevicesQuantum-Classical Hybrid Computation AccelerationUtilization of Small Quantum Computers

CutQC partitions large quantum circuits into smaller subcircuits executed on small quantum devices, using classical postprocessing to reconstruct outputs, enabling efficient evaluation of circuits beyond current NISQ limits with improved fidelity.

Resilient Distributed Datasets: A Fault-Tolerant Abstraction for In-Memory Cluster Computing

Status:completed

Resilient Distributed DatasetsIn-Memory Cluster ComputingFault-Tolerant Computing ModelLarge-Scale Iterative ComputationSpark Computing Framework

This paper introduces RDDs, a faulttolerant inmemory data abstraction enabling efficient iterative and interactive computing, implemented in Spark, boosting largescale cluster performance and scalability.

Plaintext-Ciphertext Matrix Multiplication and FHE Bootstrapping: Fast and Fused

Status:completed

Homomorphic Encryption Matrix MultiplicationMulti-Secret RLWE SchemePlaintext-Ciphertext Matrix Multiplication OptimizationCKKS Bootstrapping AccelerationBatch Bootstrapping Mechanism

This work develops multikey RLWEbased PCMM algorithms reducing complexity to efficient PPMM, integrates them with CKKS bootstrapping, and introduces MaMBo to accelerate privacypreserving model inference with minimal overhead.

One-Step Effective Diffusion Network for Real-World Image Super-Resolution

Status:completed

Real-World Image Super-ResolutionOne-Step Diffusion InferencePretrained Diffusion Model Based Image RestorationVariational Score Distillation RegularizationEfficient Diffusion Network Design

OSEDiff leverages pretrained diffusion models to perform realworld image superresolution in one step by starting diffusion from the lowquality image, removing noise uncertainty. Finetuned with variational score distillation, it efficiently achieves superior highquality resto

DaCapo: Automatic Bootstrapping Management for Efficient Fully Homomorphic Encryption

Status:completed

Automatic Bootstrapping Management for FHEFHE Performance OptimizationCompiler for Encrypted ComputationAutomatic Bootstrapping InsertionScale Management Based Latency Estimation

DaCapo automates bootstrapping in fully homomorphic encryption by analyzing ciphertext lifetimes and scale, selecting optimal insertion points to minimize latency. It reduces manual effort and improves performance, showing 1.21× speedup over handoptimized implementations.

Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5)

Status:completed

Unified Pretraining Recommendation FrameworkPersonalized Prompt LearningText-to-Text Recommendation ModelZero-Shot and Few-Shot RecommendationMulti-Task Recommendation System

P5 unifies diverse recommendation tasks as texttotext problems, using unified pretraining and personalized prompts to enable zero/fewshot prediction, enhancing knowledge transfer and generalization, paving the way for largescale recommendation models.

Autoregressive Video Generation without Vector Quantization

Status:completed

Autoregressive Video GenerationNon-Quantized ModelingGPT-Style Causal ModelingBidirectional Intra-frame ModelingEfficient Video Generation Model

NOVA reframes video generation as nonquantized autoregressive modeling combining temporal framewise and spatial setwise prediction. It outperforms prior models in efficiency, speed, fidelity, and generalizes well to longer videos and diverse zeroshot tasks, with a smaller mod

1 - 20 / 267

Papers