Page 20 - Paper Library - AiPaper

GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

Published:3/19/2025

Generalist Humanoid Robot Foundation ModelVision-Language-Action ModelDiffusion Transformer ModuleHumanoid Robot Manipulation TasksMultimodal Data Training

GR00T N1 is an open foundation model for humanoid robots, integrating a reasoning module and a motion generation module. Trained endtoend with a pyramid of heterogeneous data, it outperforms existing imitation learning methods in simulation benchmarks, demonstrating high perfor

SAFE: Multitask Failure Detection for Vision-Language-Action Models

Published:6/12/2025

Failure Detection for Vision-Language-Action ModelsMultitask Failure DetectionFeature-Based Failure PredictionGeneralist Robot PoliciesAdaptive Failure Alert System

SAFE introduces a multitask failure detection system for visionlanguageaction models, leveraging internal features to predict task failure. It demonstrates superior performance across various environments, balancing accuracy and timeliness, and supports zeroshot generalization

On the role of theories in consciousness science

Published:11/26/2025

Role of Theories in Consciousness ScienceAdvances in Neuroscience TechnologyTheoretical and Empirical Developments in Consciousness Science

This paper critiques the role of theories in consciousness science, advocating for focused 'humbler theories.' It argues that the 'uniformity assumption' limits theoretical effectiveness, promoting testable theories to enhance the interactive loop between theory and experimentati

Curriculum Conditioned Diffusion for Multimodal Recommendation

Published:4/11/2025

Multimodal Recommendation SystemsDiffusion ModelsKnowledge-Aware Negative SamplingCurriculum Learning FrameworkMultimodal Aligning Module

The Curriculum Conditioned Diffusion framework (CCDRec) addresses data sparsity in multimodal recommendation by integrating diffusion models with negative sampling, enhancing personalization through the exploration of modality correlation. Its effectiveness and robustness are val

Directly Aligning the Full Diffusion Trajectory with Fine-Grained Human Preference

Published:9/9/2025

Diffusion Model Alignment with Human PreferencesDirect-Align MethodSemantic Relative Preference Optimization (SRPO)Predefined Noise PriorImage Recovery and Interpolation Techniques

The paper introduces DirectAlign, a method that optimizes diffusion models by predefining noise priors to align with finegrained human preferences, addressing multistep denoising costs and enabling realtime reward adjustments through Semantic Relative Preference Optimization (

S-SYNC: Shuttle and Swap Co-Optimization in Quantum Charge-Coupled Devices

Published:5/2/2025

Quantum Charge-Coupled Device (QCCD)Scheduling HeuristicsQuantum Application Success Rate OptimizationQuantum Computing CompilerQubit Shuttle and Swap

The SSYNC compiler optimizes shuttling and swapping operations in Quantum ChargeCoupled Devices (QCCD), reducing shuttling by 3.69 times and improving quantum application success rates by 1.73 times, thereby advancing the practical use of quantum computing.

Diffusion Transformers with Representation Autoencoders

Published:10/14/2025

Diffusion Transformers with Representation AutoencodersRepresentation AutoencodersImprovement of Image Generation QualityHigh-Dimensional Latent Space ModelingDINO and MAE Encoders

This paper introduces Representation Autoencoders (RAEs) that replace traditional Variational Autoencoders (VAEs) with pretrained representation encoders, enhancing image generation quality in Diffusion Transformers (DiTs). RAEs achieve highquality reconstructions and rich seman

Evolution Strategies at the Hyperscale

Published:11/21/2025

Evolution Strategies AlgorithmLow-Rank Learning OptimizationLarge-Scale Neural Network OptimizationGradient-Free Optimization MethodsOptimization in Reinforcement Learning Settings

This paper introduces EGGROLL, a lowrank learning optimization method that scales backpropagationfree optimization for large neural networks, significantly reducing computational and memory costs while maintaining performance even with billions of parameters.

CipherGPT: Secure Two-Party GPT Inference

Secure Two-Party GPT InferenceEncrypted Matrix MultiplicationSecure GELU Computation ProtocolSecure Top-k Sampling ProtocolGPT Inference Optimization

CipherGPT is a framework addressing user privacy in GPT inference. It innovatively optimizes secure matrix multiplication and GELU computation, achieving significant performance gains. Additionally, it introduces a secure topk sampling protocol, providing comprehensive benchmark

DiffBIR: Towards Blind Image Restoration with Generative Diffusion Prior

Published:8/29/2023

Diffusion Model Image RestorationBlind Image RestorationHigh-Fidelity Image ReconstructionImage Quality EnhancementRegion-Adaptive Restoration Guidance

DiffBIR is a unified image restoration pipeline for blind image tasks, involving degradation removal and information regeneration. It utilizes IRControlNet for realistic detail generation and introduces regionadaptive guidance for usertunable balance between realness and fideli

Adenine base editor engineering reduces editing of bystander cytosines

Published:7/1/2021

Adenine Base Editor EngineeringDeaminase EngineeringBase Editing Tool OptimizationSpecific Base ConversionTargeted Editing Improvement

Engineered adenine base editor TadA7.10 significantly reduces unintended cytosine deamination with the D108Q mutation. This mutation lowers activity tenfold and is compatible with V106W. A P48R mutation enhances cytosine activity while decreasing adenine editing, creating a new T

ARAG: Agentic Retrieval Augmented Generation for Personalized Recommendation

Published:6/27/2025

Multi-Agent Personalized Recommendation SystemAugmented Retrieval-Generation FrameworkLLM-based Recommendation SystemsDynamic User Preference ModelingRecommendation System Evaluation

The ARAG framework enhances personalized recommendations by integrating a multiagent collaboration into RetrievalAugmented Generation. It employs agents for user understanding, natural language inference, context summarization, and item ranking, outperforming traditional RAG me

MA-RAG: Multi-Agent Retrieval-Augmented Generation via Collaborative Chain-of-Thought Reasoning

Published:10/8/2025

Multi-Agent Retrieval-Augmented GenerationCollaborative Chain-of-Thought ReasoningTask-Aware ReasoningDynamic Workflow OptimizationModular Reasoning-Driven Architecture

MARAG is a multiagent framework for RetrievalAugmented Generation, addressing ambiguities in complex information tasks. It employs specialized agents to decompose tasks and enables dynamic workflows, outperforming leading baseline models and validating collaborative reasoning'

010

HiPRAG: Hierarchical Process Rewards for Efficient Agentic Retrieval Augmented Generation

Published:10/8/2025

RL Training for Large Language ModelsAgentic Retrieval-Augmented GenerationHierarchical Process RewardsKnowledge-Grounded Process RewardsSearch Decision Optimization

HiPRAG introduces a novel hierarchical process rewards method to tackle common oversearch and undersearch issues in agentic retrievalaugmented generation, improving search efficiency and accuracy significantly across multiple QA benchmarks, demonstrating the importance of opti

RecoveryChaining: Learning Local Recovery Policies for Robust Manipulation

Published:10/18/2024

Robot Recovery Policy LearningHierarchical Reinforcement LearningComplex Manipulation TasksModel-Based ControllersSparse Reward Optimization

This study introduces RecoveryChaining, a hierarchical reinforcement learning method for robust recovery in complex manipulation tasks. By integrating nominal controllers as options, it enables effective recovery strategies upon failure detection, enhancing robustness and success

Towards Domain-Specific Network Transport for Distributed DNN Training

Distributed Deep Learning Network TransportDNN Communication OptimizationNetwork Traffic Pattern AnalysisDomain-Specific Network ProtocolsCongestion Control Techniques

This paper presents MLT, a domainspecific network transport protocol for distributed DNN training, optimizing congestion control and data transfer techniques. Extensive evaluations show significant improvements in training throughput and network utilization, addressing performan

CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer

Published:8/12/2024

Text-to-Video GenerationDiffusion ModelsDiffusion Transformer3D Variational AutoencoderVideo Generation Quality Improvement

CogVideoX is a largescale texttovideo model using a diffusion transformer that generates 10second videos at 16 fps and 768×1360 resolution. It addresses coherence and semantic alignment issues with methods like 3D VAE and expert transformers, achieving significant quality imp

Scalable Diffusion Models with Transformers

Published:12/20/2022

Diffusion ModelsTransformer architectureImage GenerationScalable Diffusion ModelsClass-Conditional Image Generation

This study introduces Diffusion Transformers (DiTs), which replace UNet with a transformer architecture for image generation. Higher Gflops correlate with better performance (lower FID), with the largest model achieving stateoftheart results on ImageNet benchmarks.

Hiding in the AI Traffic: Abusing MCP for LLM-Powered Agentic Red Teaming

Published:11/20/2025

MCP-Based Command and Control ArchitectureAutonomous Red Teaming SystemsGenerative Cyber Offense and DefenseAutonomous Penetration TestingAdversarial Behavior Analysis

This research presents a novel Command and Control architecture using the Model Context Protocol to facilitate stealthy autonomous reconnaissance agents, enhancing offensive cybersecurity by improving goaldirected behaviors and eliminating network artifacts that could reveal com

Rethinking Bradley-Terry Models in Preference-Based Reward Modeling: Foundations, Theory, and Alternatives

Published:11/8/2024

Foundations of Bradley-Terry ModelsPreference-Based Reward ModelingMulti-Task Evaluation for Large Language ModelsOrder Consistency in Reward ModelingReward Models Based on Deep Neural Networks

This paper reevaluates the BradleyTerry model in preferencebased reward modeling, establishes a theoretical foundation for convergence using deep neural networks, argues it's not strictly necessary for optimization, proposes an alternative upperbound algorithm, and validates m

381 - 400 / 980

Papers