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.

OPTIC-ER: A Reinforcement Learning Framework for Real-Time Emergency Response and Equitable Resource Allocation in Underserved African Communities

BioAnalyst: A Foundation Model for Biodiversity

Scaling Towards the Information Boundary of Instruction Sets: The Infinity Instruct Subject Technical Report

Dual-Objective Reinforcement Learning with Novel Hamilton-Jacobi-Bellman Formulations

NeuroPhysNet: A FitzHugh-Nagumo-Based Physics-Informed Neural Network Framework for Electroencephalograph (EEG) Analysis and Motor Imagery Classification

PRO-V-R1: Reasoning Enhanced Programming Agent for RTL Verification

Large language models can learn and generalize steganographic chain-of-thought under process supervision

Turing Test 2.0: The General Intelligence Threshold

Consistency-based Abductive Reasoning over Perceptual Errors of Multiple Pre-trained Models in Novel Environments

Empowering Clients - Transformation of Design Processes Due to Generative AI

The Delusional Hedge Algorithm as a Model of Human Learning from Diverse Opinions

The Universal Weight Subspace Hypothesis

DraCo: Draft as CoT for Text-to-Image Preview and Rare Concept Generation

ShadowDraw: From Any Object to Shadow-Drawing Compositional Art

Semantic Soft Bootstrapping: Long Context Reasoning in LLMs without Reinforcement Learning

TV2TV: A Unified Framework for Interleaved Language and Video Generation

Structured Document Translation via Format Reinforcement Learning

SA-IQA: Redefining Image Quality Assessment for Spatial Aesthetics with Multi-Dimensional Rewards

David vs. Goliath: Can Small Models Win Big with Agentic AI in Hardware Design?

Multi-LLM Collaboration for Medication Recommendation

Meta-Learning for Quantum Optimization via Quantum Sequence Model

QKAN-LSTM: Quantum-inspired Kolmogorov-Arnold Long Short-term Memory

Arbitrage: Efficient Reasoning via Advantage-Aware Speculation

Model-Free Assessment of Simulator Fidelity via Quantile Curves

Reflection Removal through Efficient Adaptation of Diffusion Transformers

Evolutionary Architecture Search through Grammar-Based Sequence Alignment

Strategic Self-Improvement for Competitive Agents in AI Labour Markets

Balanced Few-Shot Episodic Learning for Accurate Retinal Disease Diagnosis

GeoPE:A Unified Geometric Positional Embedding for Structured Tensors

Realizable Abstractions: Near-Optimal Hierarchical Reinforcement Learning

LLMs Know More Than Words: A Genre Study with Syntax, Metaphor & Phonetics

CARL: Critical Action Focused Reinforcement Learning for Multi-Step Agent

Declarative Synthesis and Multi-Objective Optimization of Stripboard Circuit Layouts Using Answer Set Programming

ReflexFlow: Rethinking Learning Objective for Exposure Bias Alleviation in Flow Matching

Developing a General Personal Tutor for Education

SEAL: Self-Evolving Agentic Learning for Conversational Question Answering over Knowledge Graphs

Language Models as Semantic Teachers: Post-Training Alignment for Medical Audio Understanding

Mitigating Catastrophic Forgetting in Target Language Adaptation of LLMs via Source-Shielded Updates

From Symptoms to Systems: An Expert-Guided Approach to Understanding Risks of Generative AI for Eating Disorders

SoK: a Comprehensive Causality Analysis Framework for Large Language Model Security

Setting up for failure: automatic discovery of the neural mechanisms of cognitive errors

287,872 Supermassive Black Holes Masses: Deep Learning Approaching Reverberation Mapping Accuracy

YingMusic-SVC: Real-World Robust Zero-Shot Singing Voice Conversion with Flow-GRPO and Singing-Specific Inductive Biases

YingMusic-Singer: Zero-shot Singing Voice Synthesis and Editing with Annotation-free Melody Guidance

Using Machine Learning to Take Stay-or-Go Decisions in Data-driven Drone Missions

Embodied Co-Design for Rapidly Evolving Agents: Taxonomy, Frontiers, and Challenges

UnwrapDiff: Conditional Diffusion for Robust InSAR Phase Unwrapping

SignRoundV2: Closing the Performance Gap in Extremely Low-Bit Post-Training Quantization for LLMs

Neural Policy Composition from Free Energy Minimization

OsmT: Bridging OpenStreetMap Queries and Natural Language with Open-source Tag-aware Language Models

E3AD: An Emotion-Aware Vision-Language-Action Model for Human-Centric End-to-End Autonomous Driving

Measuring the Unspoken: A Disentanglement Model and Benchmark for Psychological Analysis in the Wild

Towards an AI Fluid Scientist: LLM-Powered Scientific Discovery in Experimental Fluid Mechanics

Large Speech Model Enabled Semantic Communication

TimesNet-Gen: Deep Learning-based Site Specific Strong Motion Generation

Generative AI for Self-Adaptive Systems: State of the Art and Research Roadmap

Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs

Semi Centralized Training Decentralized Execution Architecture for Multi Agent Deep Reinforcement Learning in Traffic Signal Control

SEASON: Mitigating Temporal Hallucination in Video Large Language Models via Self-Diagnostic Contrastive Decoding

When GenAI Meets Fake News: Understanding Image Cascade Dynamics on Reddit

When Robots Should Say "I Don't Know": Benchmarking Abstention in Embodied Question Answering

A Light-Weight Large Language Model File Format for Highly-Secure Model Distribution

Diffusion Fine-Tuning via Reparameterized Policy Gradient of the Soft Q-Function

RRPO: Robust Reward Policy Optimization for LLM-based Emotional TTS

Multi-Loss Learning for Speech Emotion Recognition with Energy-Adaptive Mixup and Frame-Level Attention

AdmTree: Compressing Lengthy Context with Adaptive Semantic Trees

Detection of Intoxicated Individuals from Facial Video Sequences via a Recurrent Fusion Model

PhyVLLM: Physics-Guided Video Language Model with Motion-Appearance Disentanglement

Prototype-Based Semantic Consistency Alignment for Domain Adaptive Retrieval

UW-BioNLP at ChemoTimelines 2025: Thinking, Fine-Tuning, and Dictionary-Enhanced LLM Systems for Chemotherapy Timeline Extraction

GraphBench: Next-generation graph learning benchmarking

GuidNoise: Single-Pair Guided Diffusion for Generalized Noise Synthesis

Open-Ended Goal Inference through Actions and Language for Human-Robot Collaboration

NORi: An ML-Augmented Ocean Boundary Layer Parameterization

Automating Complex Document Workflows via Stepwise and Rollback-Enabled Operation Orchestration

Explainable Parkinsons Disease Gait Recognition Using Multimodal RGB-D Fusion and Large Language Models

Dual-Stream Spectral Decoupling Distillation for Remote Sensing Object Detection

Towards 6G Native-AI Edge Networks: A Semantic-Aware and Agentic Intelligence Paradigm

Adversarial Limits of Quantum Certification: When Eve Defeats Detection

FMA-Net++: Motion- and Exposure-Aware Real-World Joint Video Super-Resolution and Deblurring

MASE: Interpretable NLP Models via Model-Agnostic Saliency Estimation

STeP-Diff: Spatio-Temporal Physics-Informed Diffusion Models for Mobile Fine-Grained Pollution Forecasting

AutoGuard: A Self-Healing Proactive Security Layer for DevSecOps Pipelines Using Reinforcement Learning

Mitigating Object and Action Hallucinations in Multimodal LLMs via Self-Augmented Contrastive Alignment

Counting Without Running: Evaluating LLMs' Reasoning About Code Complexity

RGE-GCN: Recursive Gene Elimination with Graph Convolutional Networks for RNA-seq based Early Cancer Detection

DAComp: Benchmarking Data Agents across the Full Data Intelligence Lifecycle

Bayes-DIC Net: Estimating Digital Image Correlation Uncertainty with Bayesian Neural Networks

MANTRA: a Framework for Multi-stage Adaptive Noise TReAtment During Training

Evaluating Long-Context Reasoning in LLM-Based WebAgents

Gamma-from-Mono: Road-Relative, Metric, Self-Supervised Monocular Geometry for Vehicular Applications

Learning Single-Image Super-Resolution in the JPEG Compressed Domain

Bootstrapped Mixed Rewards for RL Post-Training: Injecting Canonical Action Order

Quantitative Analysis of Technical Debt and Pattern Violation in Large Language Model Architectures

The Initialization Determines Whether In-Context Learning Is Gradient Descent

Catching UX Flaws in Code: Leveraging LLMs to Identify Usability Flaws at the Development Stage

Hey GPT-OSS, Looks Like You Got It - Now Walk Me Through It! An Assessment of the Reasoning Language Models Chain of Thought Mechanism for Digital Forensics

Fine-Tuning ChemBERTa for Predicting Inhibitory Activity Against TDP1 Using Deep Learning

MVRoom: Controllable 3D Indoor Scene Generation with Multi-View Diffusion Models

CRAFT-E: A Neuro-Symbolic Framework for Embodied Affordance Grounding

An Evaluation of LLMs Inference on Popular Single-board Computers

Created by

Haebom

Category

Empty

저자

Tung (Thomas), Nguyen, Tuyen Nguyen

개요

온디바이스 대규모 언어 모델(LLM) 추론에 대한 수요 증가로 인해, 엣지 하드웨어에 경량화되고 비용 효율적인 AI 솔루션을 배포하는 데 관심이 높아지고 있습니다. Raspberry Pi 및 Orange Pi와 같은 단일 보드 컴퓨터(SBC)는 로컬화되고 개인 정보 보호가 가능한 추론을 위한 유망한 플랫폼이지만, LLM 워크로드 컨텍스트에서는 아직 충분히 탐구되지 않았습니다. 본 연구에서는 Ollama 및 Llamafile 두 가지 추론 런타임을 사용하여 Raspberry Pi 4, Raspberry Pi 5, Orange Pi 5 Pro 세 가지 SBC에서 25개의 양자화된 오픈 소스 LLM의 성능을 벤치마킹했습니다. 다양한 CPU 구성을 사용하여 여러 프롬프트 유형으로 현실적인 워크로드를 시뮬레이션하여 생성 처리량, 메모리 사용량 및 전력 소비량을 평가했습니다. 결과적으로 SBC는 최대 1.5B 파라미터의 모델을 안정적으로 지원할 수 있으며, Llamafile은 Ollama보다 최대 4배 높은 처리량과 30-40% 낮은 전력 사용량을 달성했습니다. 아키텍처별 병목 현상을 식별하고, 런타임 수준의 트레이드 오프를 강조하며, 실용적인 배포 권장 사항을 제공합니다. 본 연구는 SBC에서 LLM 추론에 대한 최초의 광범위한 평가를 제공하며, 고성능 언어 모델과 저렴한 엣지 컴퓨팅 간의 격차를 해소합니다.

시사점, 한계점

•

시사점:

◦

SBC는 최대 1.5B 파라미터 모델을 안정적으로 지원 가능.

◦

Llamafile은 Ollama보다 더 높은 처리량과 낮은 전력 소비를 보임.

◦

아키텍처별 병목 현상 및 런타임 수준의 트레이드 오프를 식별.

◦

실용적인 배포 권장 사항 제공.

◦

고성능 언어 모델과 저렴한 엣지 컴퓨팅의 격차 해소.

•

한계점:

◦

25개의 양자화된 오픈 소스 LLM과 3개의 SBC에 대한 벤치마킹만 수행.

◦

다양한 모델 크기 및 하드웨어 구성에 대한 확장성 평가 부족.

◦

특정 런타임(Ollama, Llamafile)에 대한 성능 비교에 한정됨.

◦

구체적인 아키텍처별 병목 현상에 대한 상세 분석 부족.

Made with Slashpage