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.

UniWorld: High-Resolution Semantic Encoders for Unified Visual Understanding and Generation

Critique-GRPO: Advancing LLM Reasoning with Natural Language and Numerical Feedback

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

Multi Layered Autonomy and AI Ecologies in Robotic Art Installations

CoRe-MMRAG: Cross-Source Knowledge Reconciliation for Multimodal RAG

Comparative Analysis of AI Agent Architectures for Entity Relationship Classification

MINT: Multimodal Instruction Tuning with Multimodal Interaction Grouping

Random-key genetic algorithms: Principles and applications

Music Interpretation and Emotion Perception: A Computational and Neurophysiological Investigation

FlashMLA-ETAP: Efficient Transpose Attention Pipeline for Accelerating MLA Inference on NVIDIA H20 GPUs

MedEBench: Revisiting Text-instructed Image Editing on Medical Domain

Tug-of-war between idiom's figurative and literal meanings in LLMs

Optimizing Sensory Neurons: Nonlinear Attention Mechanisms for Accelerated Convergence in Permutation-Invariant Neural Networks for Reinforcement Learning

It Takes a Good Model to Train a Good Model: Generalized Gaussian Priors for Optimized LLMs

Recover Experimental Data with Selection Bias using Counterfactual Logic

Children's Voice Privacy: First Steps And Emerging Challenges

ClinBench-HPB: A Clinical Benchmark for Evaluating LLMs in Hepato-Pancreato-Biliary Diseases

Bench4KE: Benchmarking Automated Competency Question Generation

Object Centric Concept Bottlenecks

Grid-LOGAT: Grid Based Local and Global Area Transcription for Video Question Answering

AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning

Large Language Models are Locally Linear Mappings

LayerIF: Estimating Layer Quality for Large Language Models using Influence Functions

Mind the Gap: A Practical Attack on GGUF Quantization

Comparing the Effects of Persistence Barcodes Aggregation and Feature Concatenation on Medical Imaging

The Arabic AI Fingerprint: Stylometric Analysis and Detection of Large Language Models Text

Implicit Inversion turns CLIP into a Decoder

ATI: Any Trajectory Instruction for Controllable Video Generation

What happens when generative AI models train recursively on each others' generated outputs?

Trans-EnV: A Framework for Evaluating the Linguistic Robustness of LLMs Against English Varieties

Adversarial bandit optimization for approximately linear functions

What LLMs Miss in Recommendations: Bridging the Gap with Retrieval-Augmented Collaborative Signals

Collision- and Reachability-Aware Multi-Robot Control with Grounded LLM Planners

EgoZero: Robot Learning from Smart Glasses

Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning

SEM: Enhancing Spatial Understanding for Robust Robot Manipulation

Polar Sparsity: High Throughput Batched LLM Inferencing with Scalable Contextual Sparsity

Algorithmic Tradeoffs in Fair Lending: Profitability, Compliance, and Long-Term Impact

SourceDetMamba: A Graph-aware State Space Model for Source Detection in Sequential Hypergraphs

HyperDet: Source Detection in Hypergraphs via Interactive Relationship Construction and Feature-rich Attention Fusion

Exploring Sparsity for Parameter Efficient Fine Tuning Using Wavelets

THELMA: Task Based Holistic Evaluation of Large Language Model Applications-RAG Question Answering

FactsR: A Safer Method for Producing High Quality Healthcare Documentation

Crowd Scene Analysis using Deep Learning Techniques

Improved Uncertainty Quantification in Physics-Informed Neural Networks Using Error Bounds and Solution Bundles

Flow-GRPO: Training Flow Matching Models via Online RL

LLM Code Customization with Visual Results: A Benchmark on TikZ

The Cognitive Foundations of Economic Exchange: A Modular Framework Grounded in Behavioral Evidence

CARL: Camera-Agnostic Representation Learning for Spectral Image Analysis

Biased by Design: Leveraging AI Inherent Biases to Enhance Critical Thinking of News Readers

A Survey on (M)LLM-Based GUI Agents

Revisiting Uncertainty Quantification Evaluation in Language Models: Spurious Interactions with Response Length Bias Results

Token-Driven GammaTune: Adaptive Calibration for Enhanced Speculative Decoding

VecTrans: Enhancing Compiler Auto-Vectorization through LLM-Assisted Code Transformations

Detecting Dataset Bias in Medical AI: A Generalized and Modality-Agnostic Auditing Framework

Language-Enhanced Representation Learning for Single-Cell Transcriptomics

SemHiTok: A Unified Image Tokenizer via Semantic-Guided Hierarchical Codebook for Multimodal Understanding and Generation

PFDial: A Structured Dialogue Instruction Fine-tuning Method Based on UML Flowcharts

ROGRAG: A Robustly Optimized GraphRAG Framework

Knowledge Retention for Continual Model-Based Reinforcement Learning

The Role of Visual Modality in Multimodal Mathematical Reasoning: Challenges and Insights

Dynamic Benchmarking of Reasoning Capabilities in Code Large Language Models Under Data Contamination

M3HF: Multi-agent Reinforcement Learning from Multi-phase Human Feedback of Mixed Quality

Anomaly Detection in Complex Dynamical Systems: A Systematic Framework Using Embedding Theory and Physics-Inspired Consistency

Sliding Window Attention Training for Efficient Large Language Models

The Built-In Robustness of Decentralized Federated Averaging to Bad Data

Aligning Compound AI Systems via System-level DPO

D2S-FLOW: Automated Parameter Extraction from Datasheets for SPICE Model Generation Using Large Language Models

FragFM: Hierarchical Framework for Efficient Molecule Generation via Fragment-Level Discrete Flow Matching

Human Misperception of Generative-AI Alignment: A Laboratory Experiment

ATRI: Mitigating Multilingual Audio Text Retrieval Inconsistencies by Reducing Data Distribution Errors

Dehumanizing Machines: Mitigating Anthropomorphic Behaviors in Text Generation Systems

Why Safeguarded Ships Run Aground? Aligned Large Language Models' Safety Mechanisms Tend to Be Anchored in The Template Region

Sleepless Nights, Sugary Days: Creating Synthetic Users with Health Conditions for Realistic Coaching Agent Interactions

Visual Attention Never Fades: Selective Progressive Attention ReCalibration for Detailed Image Captioning in Multimodal Large Language Models

Data-Juicer 2.0: Cloud-Scale Adaptive Data Processing for and with Foundation Models

Improving thermal state preparation of Sachdev-Ye-Kitaev model with reinforcement learning on quantum hardware

Neural Honeytrace: A Robust Plug-and-Play Watermarking Framework against Model Extraction Attacks

Model Alignment Search

Verbosity-Aware Rationale Reduction: Effective Reduction of Redundant Rationale via Principled Criteria

Data Laundering: Artificially Boosting Benchmark Results through Knowledge Distillation

HashAttention: Semantic Sparsity for Faster Inference

RAC3: Retrieval-Augmented Corner Case Comprehension for Autonomous Driving with Vision-Language Models

From Intention To Implementation: Automating Biomedical Research via LLMs

MMedPO: Aligning Medical Vision-Language Models with Clinical-Aware Multimodal Preference Optimization

T2I-FactualBench: Benchmarking the Factuality of Text-to-Image Models with Knowledge-Intensive Concepts

Mixture of Experts for Node Classification

CAdam: Confidence-Based Optimization for Online Learning

CatNet: Controlling the False Discovery Rate in LSTM with SHAP Feature Importance and Gaussian Mirrors

Engagement-Driven Content Generation with Large Language Models

A Survey of Event Causality Identification: Taxonomy, Challenges, Assessment, and Prospects

RoundTable: Investigating Group Decision-Making Mechanism in Multi-Agent Collaboration

Objective drives the consistency of representational similarity across datasets

Improving Radiology Report Conciseness and Structure via Local Large Language Models

Enabling LLM Knowledge Analysis via Extensive Materialization

Generating Synthetic Electronic Health Record Data: a Methodological Scoping Review with Benchmarking on Phenotype Data and Open-Source Software

Generative Emotion Cause Explanation in Multimodal Conversations

LoGU: Long-form Generation with Uncertainty Expressions

Expand and Compress: Exploring Tuning Principles for Continual Spatio-Temporal Graph Forecasting

Nudging: Inference-time Alignment of LLMs via Guided Decoding

Practical Adversarial Attacks on Stochastic Bandits via Fake Data Injection

Created by

Haebom

저자

Qirun Zeng, Eric He, Richard Hoffmann, Xuchuang Wang, Jinhang Zuo

개요

본 논문은 기존의 비현실적인 가정(라운드별 보상 조작 및 무한대의 섭동)에 의존하는 확률적 밴딧에 대한 적대적 공격 연구의 한계를 지적하고, 더욱 현실적인 위협 모델인 "가짜 데이터 주입(Fake Data Injection)"을 제안합니다. 이 모델은 제한된 수의 경계가 있는 가짜 피드백 샘플만 학습자의 기록에 주입할 수 있다는 제약을 반영하여 실제 상호작용을 모방합니다. 본 논문에서는 이 모델 하에서 보상 값에 대한 크기 제약과 데이터 주입 시점 및 빈도에 대한 시간적 제약을 명시적으로 고려한 효율적인 공격 전략을 설계합니다. 이론적 분석을 통해 이러한 공격이 UCB와 Thompson Sampling 알고리즘을 속여 거의 모든 라운드에서 목표 암을 선택하게 만들면서도 하위 선형의 공격 비용만 발생시킬 수 있음을 보여줍니다. 합성 및 실제 데이터셋에 대한 실험을 통해 제안된 전략의 효과를 검증하고, 실제 적대적 시나리오에서 널리 사용되는 확률적 밴딧 알고리즘의 취약성을 드러냅니다.

시사점, 한계점

•

시사점:

◦

현실적인 제약 조건을 고려한 새로운 적대적 공격 모델(가짜 데이터 주입) 제시

◦

UCB 및 Thompson Sampling 알고리즘의 실제 적대적 상황에 대한 취약성 규명

◦

제한된 자원으로도 효과적인 적대적 공격이 가능함을 증명

◦

확률적 밴딧 알고리즘의 보안 강화에 대한 필요성 제기

•

한계점:

◦

제안된 공격 모델의 일반성: 특정 유형의 밴딧 알고리즘과 데이터 분포에 대한 효과성 검증이 필요

◦

실제 시스템에서의 적용 가능성: 가짜 데이터 주입의 실제 구현 및 탐지에 대한 추가 연구 필요

◦

공격 전략의 최적화: 더욱 효율적이고 강력한 공격 전략 개발 필요

Made with Slashpage