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.

On the Brittleness of CLIP Text Encoders

NVIDIA Nemotron Nano V2 VL

Improving Diagnostic Performance on Small and Imbalanced Datasets Using Class-Based Input Image Composition

Retrofitters, pragmatists and activists: Public interest litigation for accountable automated decision-making

Text to Robotic Assembly of Multi Component Objects using 3D Generative AI and Vision Language Models

The Future of Generative AI in Software Engineering: A Vision from Industry and Academia in the European GENIUS Project

AURA: A Reinforcement Learning Framework for AI-Driven Adaptive Conversational Surveys

Grounded in Reality: Learning and Deploying Proactive LLM from Offline Logs

USF-MAE: Ultrasound Self-Supervised Foundation Model with Masked Autoencoding

CompressionAttack: Exploiting Prompt Compression as a New Attack Surface in LLM-Powered Agents

Every Activation Boosted: Scaling General Reasoner to 1 Trillion Open Language Foundation

Enterprise Deep Research: Steerable Multi-Agent Deep Research for Enterprise Analytics

What Can String Probability Tell Us About Grammaticality?

Fine-Tuning MedGemma for Clinical Captioning to Enhance Multimodal RAG over Malaysia CPGs

Learning to Navigate Socially Through Proactive Risk Perception

Policy-as-Prompt: Turning AI Governance Rules into Guardrails for AI Agents

A Mega-Study of Digital Twins Reveals Strengths, Weaknesses and Opportunities for Further Improvement

Amortized Latent Steering: Low-Cost Alternative to Test-Time Optimization

Amulet: a Python Library for Assessing Interactions Among ML Defenses and Risks

LimiX: Unleashing Structured-Data Modeling Capability for Generalist Intelligence

AI Agentic Vulnerability Injection And Transformation with Optimized Reasoning

What Matters in Data for DPO?

Chain-of-Query: Unleashing the Power of LLMs in SQL-Aided Table Understanding via Multi-Agent Collaboration

Ethics-Aware Safe Reinforcement Learning for Rare-Event Risk Control in Interactive Urban Driving

Advanced Hybrid Transformer LSTM Technique with Attention and TS Mixer for Drilling Rate of Penetration Prediction

Learning Dynamics of Meta-Learning in Small Model Pretraining

GAITEX: Human motion dataset of impaired gait and rehabilitation exercises using inertial and optical sensors

To Trust or Not to Trust: On Calibration in ML-based Resource Allocation for Wireless Networks

NMIXX: Domain-Adapted Neural Embeddings for Cross-Lingual eXploration of Finance

ZERO: Industry-ready Vision Foundation Model with Multi-modal Prompts

Conformal Information Pursuit for Interactively Guiding Large Language Models

Less Greedy Equivalence Search

Progressive Inference-Time Annealing of Diffusion Models for Sampling from Boltzmann Densities

Know What You Don't Know: Uncertainty Calibration of Process Reward Models

Conformal Prediction Adaptive to Unknown Subpopulation Shifts

ProRefine: Inference-Time Prompt Refinement with Textual Feedback

Learning of Population Dynamics: Inverse Optimization Meets JKO Scheme

RCCDA: Adaptive Model Updates in the Presence of Concept Drift under a Constrained Resource Budget

Optimizing Anytime Reasoning via Budget Relative Policy Optimization

Taskmaster Deconstructed: A Quantitative Look at Tension, Volatility, and Viewer Ratings

XBreaking: Understanding how LLMs security alignment can be broken

Pogobot: an Open-Source, Low-Cost Robot for Swarm Robotics and Programmable Active Matter

Large language models as uncertainty-calibrated optimizers for experimental discovery

AIRepr: An Analyst-Inspector Framework for Evaluating Reproducibility of LLMs in Data Science

InterFeedback: Unveiling Interactive Intelligence of Large Multimodal Models via Human Feedback

Analyzing limits for in-context learning

Activation-Informed Merging of Large Language Models

iTool: Reinforced Fine-Tuning with Dynamic Deficiency Calibration for Advanced Tool Use

MMDocIR: Benchmarking Multimodal Retrieval for Long Documents

Cognitive Edge Computing: A Comprehensive Survey on Optimizing Large Models and AI Agents for Pervasive Deployment

TOBUGraph: Knowledge Graph-Based Retrieval for Enhanced LLM Performance Beyond RAG

CTPD: Cross-Modal Temporal Pattern Discovery for Enhanced Multimodal Electronic Health Records Analysis

Affordance-based Robot Manipulation with Flow Matching

FunOTTA: On-the-Fly Adaptation on Cross-Domain Fundus Image via Stable Test-time Training

Tactical Decision Making for Autonomous Trucks by Deep Reinforcement Learning with Total Cost of Operation Based Reward

Characterizing the Training Dynamics of Private Fine-tuning with Langevin diffusion

Learning for Interval Prediction of Electricity Demand: A Cluster-based Bootstrapping Approach

The Less Intelligent the Elements, the More Intelligent the Whole. Or, Possibly Not?

Monitor-Generate-Verify (MGV): Formalising Metacognitive Theory for Language Model Reasoning

Outbidding and Outbluffing Elite Humans: Mastering Liar's Poker via Self-Play and Reinforcement Learning

A Proprietary Model-Based Safety Response Framework for AI Agents

Orion-MSP: Multi-Scale Sparse Attention for Tabular In-Context Learning

Inverse Knowledge Search over Verifiable Reasoning: Synthesizing a Scientific Encyclopedia from a Long Chains-of-Thought Knowledge Base

From Observability Data to Diagnosis: An Evolving Multi-agent System for Incident Management in Cloud Systems

How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations

String Seed of Thought: Prompting LLMs for Distribution-Faithful and Diverse Generation

HugAgent: Benchmarking LLMs for Simulation of Individualized Human Reasoning

Internal World Models as Imagination Networks in Cognitive Agents

Open Agent Specification (Agent Spec): A Unified Representation for AI Agents

Bilinear relational structure fixes reversal curse and enables consistent model editing

Introducing LongCat-Flash-Thinking: A Technical Report

AI Through the Human Lens: Investigating Cognitive Theories in Machine Psychology

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

Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization

Joint Verification and Refinement of Language Models for Safety-Constrained Planning

TimeSearch-R: Adaptive Temporal Search for Long-Form Video Understanding via Self-Verification Reinforcement Learning

DGTN: Graph-Enhanced Transformer with Diffusive Attention Gating Mechanism for Enzyme DDG Prediction

On Flow Matching KL Divergence

AI Literacy Assessment Revisited: A Task-Oriented Approach Aligned with Real-world Occupations

SWE-Compass: Towards Unified Evaluation of Agentic Coding Abilities for Large Language Models

Self-adaptive weighting and sampling for physics-informed neural networks

APP: Accelerated Path Patching with Task-Specific Pruning

"I Like That You Have to Poke Around": Instructors on How Experiential Approaches to AI Literacy Spark Inquiry and Critical Thinking

ProDER: A Continual Learning Approach for Fault Prediction in Evolving Smart Grids

Multi-modal Loop Closure Detection with Foundation Models in Severely Unstructured Environments

Robust Neural Audio Fingerprinting using Music Foundation Models

Sample Complexity of Distributionally Robust Off-Dynamics Reinforcement Learning with Online Interaction

AI Assisted AR Assembly: Object Recognition and Computer Vision for Augmented Reality Assisted Assembly

TeaRAG: A Token-Efficient Agentic Retrieval-Augmented Generation Framework

AI Literacy for Community Colleges: Instructors' Perspectives on Scenario-Based and Interactive Approaches to Teaching AI

A multimodal multiplex of the mental lexicon for multilingual individuals

Perceptually Aligning Representations of Music via Noise-Augmented Autoencoders

What Are the Facts? Automated Extraction of Court-Established Facts from Criminal-Court Opinions

Rethinking Metrics and Diffusion Architecture for 3D Point Cloud Generation

LiveStar: Live Streaming Assistant for Real-World Online Video Understanding

DeepEyesV2: Toward Agentic Multimodal Model

TAMAS: Benchmarking Adversarial Risks in Multi-Agent LLM Systems

Integrating Score-Based Diffusion Models with Machine Learning-Enhanced Localization for Advanced Data Assimilation in Geological Carbon Storage

Find A Winning Sign: Sign Is All We Need to Win the Lottery

Created by

Haebom

저자

Junghun Oh, Sungyong Baik, Kyoung Mu Lee

개요

본 논문은 Lottery Ticket Hypothesis (LTH)에 대한 연구로, 과매개화된 신경망 내에 일반화 성능이 우수한 희소 부분 네트워크(winning ticket)가 존재한다는 가설을 다룬다. 기존의 winning ticket 발견 방법인 Iterative Pruning (IP)은 초기화 방식 및 소규모 아키텍처/데이터셋에 의존적이거나, 훈련된 가중치에 마스크를 적용하는 방식으로 한계를 보인다. 본 논문에서는 매개변수 부호 설정이 일반화에 중요한 역할을 한다는 것을 보여주며, 선형 모드 연결 분석을 통해 기존 IP 방법으로 훈련된 희소 네트워크가 매개변수 부호 및 정규화 계층 매개변수를 유지하면 basin of attraction을 유지할 수 있음을 확인한다. 정규화 계층 매개변수에 대한 의존성을 줄이고, 희소 네트워크와 초기화된 정규화 계층 매개변수를 가진 네트워크 간의 선형 경로 상의 높은 에러 장벽을 방지함으로써 winning ticket 발견에 한 걸음 더 다가선다. 다양한 아키텍처와 데이터셋에서, 희소성과 매개변수 부호 정보를 상속받음으로써 임의로 초기화된 네트워크가 본 논문의 방법으로 훈련된 희소 네트워크에 대한 선형 경로 상의 낮은 에러 장벽을 가지도록 최적화될 수 있으며, 원래 네트워크와 비슷한 성능을 달성할 수 있음을 관찰했다.

시사점, 한계점

•

시사점:

◦

매개변수 부호 설정이 신경망의 일반화 성능에 중요한 역할을 한다는 것을 밝힘.

◦

정규화 계층 매개변수에 대한 의존성을 줄이는 새로운 winning ticket 발견 방법 제시.

◦

임의로 초기화된 네트워크에서도 희소 네트워크의 성능에 근접한 성능을 달성할 가능성을 제시.

◦

다양한 아키텍처와 데이터셋에서 실험적으로 검증.

•

한계점:

◦

제안된 방법의 일반화 가능성에 대한 추가적인 연구 필요.

◦

매우 큰 규모의 네트워크나 데이터셋에 대한 실험 결과가 부족할 수 있음.

◦

선형 모드 연결 분석의 한계로 인해 비선형적인 영역에서의 일반화 성능은 추가 연구가 필요함.

Made with Slashpage