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.

Private Frequency Estimation Via Residue Number Systems

A Unified Convergence Analysis for Semi-Decentralized Learning: Sampled-to-Sampled vs. Sampled-to-All Communication

Privacy Challenges and Solutions in Retrieval-Augmented Generation-Enhanced LLMs for Healthcare Chatbots: A Review of Applications, Risks, and Future Directions

Virtual Width Networks

Efficient Reinforcement Learning for Zero-Shot Coordination in Evolving Games

Algorithms Trained on Normal Chest X-rays Can Predict Health Insurance Types

DialogGraph-LLM: Graph-Informed LLMs for End-to-End Audio Dialogue Intent Recognition

A Style is Worth One Code: Unlocking Code-to-Style Image Generation with Discrete Style Space

MonkeyOCR v1.5 Technical Report: Unlocking Robust Document Parsing for Complex Patterns

DermAI: Clinical dermatology acquisition through quality-driven image collection for AI classification in mobile

BhashaKritika: Building Synthetic Pretraining Data at Scale for Indic Languages

T2IBias: Uncovering Societal Bias Encoded in the Latent Space of Text-to-Image Generative Models

History Rhymes: Macro-Contextual Retrieval for Robust Financial Forecasting

C$^3$TG: Conflict-aware, Composite, and Collaborative Controlled Text Generation

Good-for-MDP State Reduction for Stochastic LTL Planning

PAN: A World Model for General, Interactable, and Long-Horizon World Simulation

HQ-SVC: Towards High-Quality Zero-Shot Singing Voice Conversion in Low-Resource Scenarios

MARC: Multimodal and Multi-Task Agentic Retrieval-Augmented Generation for Cold-Start Recommender System

Constrained and Robust Policy Synthesis with Satisfiability-Modulo-Probabilistic-Model-Checking

Self-Correction Distillation for Structured Data Question Answering

Test-driven Reinforcement Learning

Learning Quantized Continuous Controllers for Integer Hardware

Differentiated Directional Intervention A Framework for Evading LLM Safety Alignment

Breaking the Dyadic Barrier: Rethinking Fairness in Link Prediction Beyond Demographic Parity

Ghost in the Transformer: Tracing LLM Lineage with SVD-Fingerprint

Reaction Prediction via Interaction Modeling of Symmetric Difference Shingle Sets

Measuring Model Performance in the Presence of an Intervention

Ada-FCN: Adaptive Frequency-Coupled Network for fMRI-Based Brain Disorder Classification

MusRec: Zero-Shot Text-to-Music Editing via Rectified Flow and Diffusion Transformers

Proto-LeakNet: Towards Signal-Leak Aware Attribution in Synthetic Human Face Imagery

DMSORT: An efficient parallel maritime multi-object tracking architecture for unmanned vessel platforms

Multi-Personality Generation of LLMs at Decoding-time

LeMiCa: Lexicographic Minimax Path Caching for Efficient Diffusion-Based Video Generation

Comparative Study of UNet-based Architectures for Liver Tumor Segmentation in Multi-Phase Contrast-Enhanced Computed Tomography

Reflections on the Reproducibility of Commercial LLM Performance in Empirical Software Engineering Studies

MMEdge: Accelerating On-device Multimodal Inference via Pipelined Sensing and Encoding

VeriStruct: AI-assisted Automated Verification of Data-Structure Modules in Verus

Mutual Wanting in Human--AI Interaction: Empirical Evidence from Large-Scale Analysis of GPT Model Transitions

Geometric Algorithms for Neural Combinatorial Optimization with Constraints

SelecTKD: Selective Token-Weighted Knowledge Distillation for LLMs

Integrating Genomics into Multimodal EHR Foundation Models

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

A Multi-level Analysis of Factors Associated with Student Performance: A Machine Learning Approach to the SAEB Microdata

Online Mixture of Experts: No-Regret Learning for Optimal Collective Decision-Making

When Facts Change: Probing LLMs on Evolving Knowledge with evolveQA

Local Guidance for Configuration-Based Multi-Agent Pathfinding

3D Optimization for AI Inference Scaling: Balancing Accuracy, Cost, and Latency

SparseWorld: A Flexible, Adaptive, and Efficient 4D Occupancy World Model Powered by Sparse and Dynamic Queries

InfiMed-ORBIT: Aligning LLMs on Open-Ended Complex Tasks via Rubric-Based Incremental Training

DeceptionBench: A Comprehensive Benchmark for AI Deception Behaviors in Real-world Scenarios

DEXTER: Diffusion-Guided EXplanations with TExtual Reasoning for Vision Models

Dedelayed: Deleting remote inference delay via on-device correction

From Delegates to Trustees: How Optimizing for Long-Term Interests Shapes Bias and Alignment in LLM

On the Fairness of Privacy Protection: Measuring and Mitigating the Disparity of Group Privacy Risks for Differentially Private Machine Learning

A Human Behavioral Baseline for Collective Governance in Software Projects

Implicit-Knowledge Visual Question Answering with Structured Reasoning Traces

Reasoning under Vision: Understanding Visual-Spatial Cognition in Vision-Language Models for CAPTCHA

Global Convergence of Policy Gradient for Entropy Regularized Linear-Quadratic Control with Multiplicative Noise

Hierarchical Generalized Category Discovery for Brain Tumor Classification in Digital Pathology

Does Bigger Mean Better? Comparitive Analysis of CNNs and Biomedical Vision Language Modles in Medical Diagnosis

A Measurement Study of Model Context Protocol Ecosystem

Energy Guided Geometric Flow Matching

HyperCore: Coreset Selection under Noise via Hypersphere Models

VIR-Bench: Evaluating Geospatial and Temporal Understanding of MLLMs via Travel Video Itinerary Reconstruction

Understanding Post-Training Structural Changes in Large Language Models

Highly Imbalanced Regression with Tabular Data in SEP and Other Applications

MedFact: Benchmarking the Fact-Checking Capabilities of Large Language Models on Chinese Medical Texts

Exploring Efficient Open-Vocabulary Segmentation in the Remote Sensing

A GPU-Accelerated RAG-Based Telegram Assistant for Supporting Parallel Processing Students

Probabilistic Robustness Analysis in High Dimensional Space: Application to Semantic Segmentation Network

A Comparative Benchmark of Federated Learning Strategies for Mortality Prediction on Heterogeneous and Imbalanced Clinical Data

Realism Control One-step Diffusion for Real-World Image Super-Resolution

An Ontology-Based Approach to Optimizing Geometry Problem Sets for Skill Development

AHAMask: Reliable Task Specification for Large Audio Language Models without Instructions

NoLBERT: A No Lookahead(back) Foundational Language Model

SoK: Large Language Model Copyright Auditing via Fingerprinting

PAX-TS: Model-agnostic multi-granular explanations for time series forecasting via localized perturbations

Multi-Metric Preference Alignment for Generative Speech Restoration

Error analysis for the deep Kolmogorov method

Interpreting the Effects of Quantization on LLMs

Few-shot Class-incremental Fault Diagnosis by Preserving Class-Agnostic Knowledge with Dual-Granularity Representations

STA-GANN: A Valid and Generalizable Spatio-Temporal Kriging Approach

Unintended Misalignment from Agentic Fine-Tuning: Risks and Mitigation

edgeVLM: Cloud-edge Collaborative Real-time VLM based on Context Transfer

What One Cannot, Two Can: Two-Layer Transformers Provably Represent Induction Heads on Any-Order Markov Chains

Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression

HierarchicalPrune: Position-Aware Compression for Large-Scale Diffusion Models

NLP Methods May Actually Be Better Than Professors at Estimating Question Difficulty

Landsat30-AU: A Vision-Language Dataset for Australian Landsat Imagery

Intention-Guided Cognitive Reasoning for Egocentric Long-Term Action Anticipation

NyayaRAG: Realistic Legal Judgment Prediction with RAG under the Indian Common Law System

MoCHA: Advanced Vision-Language Reasoning with MoE Connector and Hierarchical Group Attention

Exploiting Synergistic Cognitive Biases to Bypass Safety in LLMs

TransPrune: Token Transition Pruning for Efficient Large Vision-Language Model

Machine Unlearning of Traffic State Estimation and Prediction

Benchmarking LLM Privacy Recognition for Social Robot Decision Making

Dream, Lift, Animate: From Single Images to Animatable Gaussian Avatars

AI-Native Open RAN for Non-Terrestrial Networks: An Overview

Sequential Attention-based Sampling for Histopathological Analysis

RAG-R1: Incentivizing the Search and Reasoning Capabilities of LLMs through Multi-query Parallelism

NNsight and NDIF: Democratizing Access to Open-Weight Foundation Model Internals

Created by

Haebom

저자

Jaden Fiotto-Kaufman, Alexander R. Loftus, Eric Todd, Jannik Brinkmann, Koyena Pal, Dmitrii Troitskii, Michael Ripa, Adam Belfki, Can Rager, Caden Juang, Aaron Mueller, Samuel Marks, Arnab Sen Sharma, Francesca Lucchetti, Nikhil Prakash, Carla Brodley, Arjun Guha, Jonathan Bell, Byron C. Wallace, David Bau

개요

NNsight와 NDIF는 매우 큰 신경망이 학습한 표현과 계산에 대한 과학적 연구를 가능하게 하는 기술입니다. NNsight는 PyTorch를 확장하여 지연된 원격 실행을 도입한 오픈소스 시스템이고, NDIF는 NNsight 요청을 실행하는 확장 가능한 추론 서비스로 사용자가 GPU 리소스와 사전 훈련된 모델을 공유할 수 있도록 합니다. 이러한 기술은 실험 설계와 모델 런타임을 분리하기 위해 개발된 Intervention Graph 아키텍처를 기반으로 합니다. 이 프레임워크는 개별적으로 사용자 정의 모델을 호스팅하는 비용이나 복잡성 없이 매우 큰 언어 모델(LLM)과 같은 심층 신경망의 내부에 투명하고 효율적으로 접근할 수 있도록 합니다. 본 논문은 대규모 AI의 내부 연구에 대한 증가하는 격차를 보여주는 기계 학습 문헌에 대한 정량적 조사를 수행하고, 이 격차를 해결하기 위해 거대한 모델에 대한 다양한 연구 방법을 가능하게 하는 프레임워크의 설계 및 사용을 보여줍니다. 마지막으로, 이전 접근 방식과의 성능을 비교하기 위한 벤치마크를 수행합니다. 코드, 설명서 및 자습서는 https://nnsight.net/ 에서 확인할 수 있습니다.

nnsight — nnsight

시사점, 한계점

•

시사점:

◦

매우 큰 신경망(예: LLM)의 내부에 대한 투명하고 효율적인 접근을 제공합니다.

◦

GPU 리소스와 사전 훈련된 모델을 공유하여 연구 비용을 절감합니다.

◦

대규모 AI 내부 연구에 대한 격차를 해소하는 데 기여합니다.

◦

다양한 연구 방법을 거대한 모델에 적용할 수 있도록 합니다.

◦

오픈소스 시스템으로 접근성이 높습니다.

•

한계점:

◦

NNsight와 NDIF의 성능은 NDIF 인프라의 용량 및 성능에 의존적일 수 있습니다.

◦

Intervention Graph 아키텍처의 복잡성으로 인해 사용에 어려움을 겪을 수 있습니다.

◦

지원하는 모델의 종류 및 크기에 제한이 있을 수 있습니다.

◦

논문에서 제시된 벤치마크 결과가 모든 상황에 일반화될 수 있는지에 대한 검증이 필요합니다.

Made with Slashpage