Daily Arxiv

전 세계에서 발간되는 인공지능 관련 논문을 정리하는 페이지 입니다.
본 페이지는 Google Gemini를 활용해 요약 정리하며, 비영리로 운영 됩니다.
논문에 대한 저작권은 저자 및 해당 기관에 있으며, 공유 시 출처만 명기하면 됩니다.

CoRT: Code-integrated Reasoning within Thinking

TransXSSM: A Hybrid Transformer State Space Model with Unified Rotary Position Embedding

Policy-Based Trajectory Clustering in Offline Reinforcement Learning

Understanding Human-AI Trust in Education

ConfPO: Exploiting Policy Model Confidence for Critical Token Selection in Preference Optimization

MasHost Builds It All: Autonomous Multi-Agent System Directed by Reinforcement Learning

CheMatAgent: Enhancing LLMs for Chemistry and Materials Science through Tree-Search Based Tool Learning

DeepVideo-R1: Video Reinforcement Fine-Tuning via Difficulty-aware Regressive GRPO

Robotic Policy Learning via Human-assisted Action Preference Optimization

LLM-D12: A Dual-Dimensional Scale of Instrumental and Relational Dependencies on Large Language Models

QuantMCP: Grounding Large Language Models in Verifiable Financial Reality

Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce

Multi-Modal Multi-Task Federated Foundation Models for Next-Generation Extended Reality Systems: Towards Privacy-Preserving Distributed Intelligence in AR/VR/MR

Peer-Ranked Precision: Creating a Foundational Dataset for Fine-Tuning Vision Models from DataSeeds' Annotated Imagery

Q-Ponder: A Unified Training Pipeline for Reasoning-based Visual Quality Assessment

Sample Complexity and Representation Ability of Test-time Scaling Paradigms

Context Is Not Comprehension

High Performance Space Debris Tracking in Complex Skylight Backgrounds with a Large-Scale Dataset

SALAD: Systematic Assessment of Machine Unlearing on LLM-Aided Hardware Design

iQUEST: An Iterative Question-Guided Framework for Knowledge Base Question Answering

Incentivizing Reasoning for Advanced Instruction-Following of Large Language Models

SealQA: Raising the Bar for Reasoning in Search-Augmented Language Models

Subgraph Gaussian Embedding Contrast for Self-Supervised Graph Representation Learning

Quantum AIXI: Universal Intelligence via Quantum Information

VTool-R1: VLMs Learn to Think with Images via Reinforcement Learning on Multimodal Tool Use

Qronos: Correcting the Past by Shaping the Future... in Post-Training Quantization

QuXAI: Explainers for Hybrid Quantum Machine Learning Models

Convert Language Model into a Value-based Strategic Planner

PhysNav-DG: A Novel Adaptive Framework for Robust VLM-Sensor Fusion in Navigation Applications

Token-Efficient RL for LLM Reasoning

MAYA: Addressing Inconsistencies in Generative Password Guessing through a Unified Benchmark

Elucidating the Design Space of Multimodal Protein Language Models

A Minimalist Approach to LLM Reasoning: from Rejection Sampling to Reinforce

On the Geometry of Receiver Operating Characteristic and Precision-Recall Curves

Expert Race: A Flexible Routing Strategy for Scaling Diffusion Transformer with Mixture of Experts

Content ARCs: Decentralized Content Rights in the Age of Generative AI

PLAY2PROMPT: Zero-shot Tool Instruction Optimization for LLM Agents via Tool Play

Computation Mechanism Behind LLM Position Generalization

CompMarkGS: Robust Watermarking for Compressed 3D Gaussian Splatting

Large Language Models for Outpatient Referral: Problem Definition, Benchmarking and Challenges

Social Bias Benchmark for Generation: A Comparison of Generation and QA-Based Evaluations

An energy-efficient learning solution for the Agile Earth Observation Satellite Scheduling Problem

Generative Uncertainty in Diffusion Models

EgoNormia: Benchmarking Physical Social Norm Understanding

Obliviate: Efficient Unmemorization for Protecting Intellectual Property in Large Language Models

From Features to Graphs: Exploring Graph Structures and Pairwise Interactions via GNNs

Object-Centric Latent Action Learning

Quality over Quantity: Boosting Data Efficiency Through Ensembled Multimodal Data Curation

TransMLA: Multi-Head Latent Attention Is All You Need

Implicit Language Models are RNNs: Balancing Parallelization and Expressivity

Lightweight Dataset Pruning without Full Training via Example Difficulty and Prediction Uncertainty

Prompt-based Depth Pruning of Large Language Models

Great Models Think Alike and this Undermines AI Oversight

Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting

Latent Action Learning Requires Supervision in the Presence of Distractors

SR-Reward: Taking The Path More Traveled

Heterogeneous Multi-Agent Reinforcement Learning for Distributed Channel Access in WLANs

SoK: Watermarking for AI-Generated Content

Engagement-Driven Content Generation with Large Language Models

PyGen: A Collaborative Human-AI Approach to Python Package Creation

DAWN: Designing Distributed Agents in a Worldwide Network

Efficient Length-Generalizable Attention via Causal Retrieval for Long-Context Language Modeling

Center-fixing of tropical cyclones using uncertainty-aware deep learning applied to high-temporal-resolution geostationary satellite imagery

LLM-Cure: LLM-based Competitor User Review Analysis for Feature Enhancement

Deploying Open-Source Large Language Models: A performance Analysis

Neural Networks Generalize on Low Complexity Data

M3-JEPA: Multimodal Alignment via Multi-gate MoE based on the Joint-Predictive Embedding Architecture

Paired Completion: Flexible Quantification of Issue-framing at Scale with LLMs

The Fellowship of the LLMs: Multi-Agent Workflows for Synthetic Preference Optimization Dataset Generation

TimeBridge: Better Diffusion Prior Design with Bridge Models for Time Series Generation

Multi-group Uncertainty Quantification for Long-form Text Generation

Hey, That's My Model! Introducing Chain & Hash, An LLM Fingerprinting Technique

Privacy-Aware Spectrum Pricing and Power Control Optimization for LEO Satellite Internet-of-Things

IndoToxic2024: A Demographically-Enriched Dataset of Hate Speech and Toxicity Types for Indonesian Language

Incentivizing Quality Text Generation via Statistical Contracts

Visually Descriptive Language Model for Vector Graphics Reasoning

Securing Large Language Models: Threats, Vulnerabilities and Responsible Practices

Mitigating Object Hallucination in Large Vision-Language Models via Image-Grounded Guidance

Near-Optimal Algorithms for Constrained k-Center Clustering with Instance-level Background Knowledge

IoTGeM: Generalizable Models for Behaviour-Based IoT Attack Detection

Improved Algorithm for Deep Active Learning under Imbalance via Optimal Separation

ConvD: Attention Enhanced Dynamic Convolutional Embeddings for Knowledge Graph Completion

Noise Balance and Stationary Distribution of Stochastic Gradient Descent

The Packing Chromatic Number of the Infinite Square Grid is 15

Reinforcing Multimodal Understanding and Generation with Dual Self-rewards

A Proposal to Extend the Common Model of Cognition with Metacognition

The Optimization Paradox in Clinical AI Multi-Agent Systems

CHANCERY: Evaluating Corporate Governance Reasoning Capabilities in Language Models

DeePoly: A High-Order Accuracy Scientific Machine Learning Framework for Function Approximation and Solving PDE

Beamforming and Resource Allocation for Delay Optimization in RIS-Assisted OFDM Systems

Evaluation of LLMs for mathematical problem solving

The Automated but Risky Game: Modeling Agent-to-Agent Negotiations and Transactions in Consumer Markets

A Heuristic Algorithm Based on Beam Search and Iterated Local Search for the Maritime Inventory Routing Problem

A Vision for Auto Research with LLM Agents

AssistanceZero: Scalably Solving Assistance Games

Don't Lag, RAG: Training-Free Adversarial Detection Using RAG

Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search

Training-Free Safe Denoisers for Safe Use of Diffusion Models

CollabLLM: From Passive Responders to Active Collaborators

Position: Theory of Mind Benchmarks are Broken for Large Language Models

SageAttention2: Efficient Attention with Thorough Outlier Smoothing and Per-thread INT4 Quantization

Created by

Haebom

저자

Jintao Zhang, Haofeng Huang, Pengle Zhang, Jia Wei, Jun Zhu, Jianfei Chen

개요

SageAttention2는 선형 계층에 대한 양자화가 널리 사용됨에도 불구하고 어텐션 계산 가속에 대한 적용이 제한적이라는 점을 해결하기 위해 제안된 알고리즘입니다. SageAttention보다 효율성을 높이고 정확도를 유지하기 위해 하드웨어 친화적인 스레드 수준의 입자성을 가진 4비트 행렬 곱셈(Matmul)과 추가적인 정확도 향상 기술을 활용합니다. 주요 기술로는 (Q, K) 행렬을 INT4로, (P̃, V) 행렬을 FP8로 양자화하는 것, INT4 QKᵀ의 정확도를 높이기 위한 Q의 스무딩 기법, 그리고 FP8 P̃V의 정확도를 높이기 위한 2단계 누적 전략이 있습니다. RTX 4090에서 FlashAttention2와 xformers보다 각각 약 3배, 4.5배 빠른 OPS를 달성하며, Hopper GPU에서는 FlashAttention3(fp8)과 비슷한 속도를 보이면서 훨씬 높은 정확도를 제공합니다. 다양한 모델(언어, 이미지, 비디오 생성)에서 종단 간 지표 손실이 무시할 만하다는 것을 실험을 통해 확인하였습니다. 코드는 https://github.com/thu-ml/SageAttention 에서 이용 가능합니다.

GitHub - thu-ml/SageAttention: Quantized Attention achieves speedup of 2-5x and 3-11x compared to FlashAttention and xformers, without lossing end-to-end metrics across language, image, and video models.

Quantized Attention achieves speedup of 2-5x and 3-11x compared to FlashAttention and xformers, without lossing end-to-end metrics across language, image, and video models. - thu-ml/SageAttention

시사점, 한계점

•

시사점:

◦

기존 어텐션 메커니즘보다 훨씬 빠른 속도(RTX 4090에서 FlashAttention2 대비 3배, xformers 대비 4.5배)를 제공합니다.

◦

Hopper GPU에서 FlashAttention3(fp8)과 유사한 속도를 유지하면서 더 높은 정확도를 달성합니다.

◦

다양한 모델에서 종단 간 성능 저하가 거의 없음을 확인했습니다.

◦

하드웨어 친화적인 설계로 효율적인 어텐션 계산을 가능하게 합니다.

•

한계점:

◦

특정 하드웨어(RTX 4090, Hopper GPU)에 대한 성능 평가가 주를 이루어 다른 하드웨어에서의 성능은 추가적인 검증이 필요합니다.

◦

제시된 정확도 향상 기법의 일반성 및 다른 양자화 기법과의 비교 분석이 부족할 수 있습니다.

◦

4-bit와 FP8 양자화 방식에 대한 세부적인 설명이 추가적으로 필요할 수 있습니다.

Made with Slashpage