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.

HPC Digital Twins for Evaluating Scheduling Policies, Incentive Structures and their Impact on Power and Cooling

NLKI: A lightweight Natural Language Knowledge Integration Framework for Improving Small VLMs in Commonsense VQA Tasks

Interact-Custom: Customized Human Object Interaction Image Generation

A Self-Supervised Mixture-of-Experts Framework for Multi-behavior Recommendation

MIDAS: Multimodal Interactive Digital-humAn Synthesis via Real-time Autoregressive Video Generation

From Tabula Rasa to Emergent Abilities: Discovering Robot Skills via Real-World Unsupervised Quality-Diversity

Dynamic Triangulation-Based Graph Rewiring for Graph Neural Networks

STDiff: A State Transition Diffusion Framework for Time Series Imputation in Industrial Systems

LLMs Can't Handle Peer Pressure: Crumbling under Multi-Agent Social Interactions

Graph-R1: Incentivizing the Zero-Shot Graph Learning Capability in LLMs via Explicit Reasoning

Modality-Specific Speech Enhancement and Noise-Adaptive Fusion for Acoustic and Body-Conduction Microphone Framework

Humans Perceive Wrong Narratives from AI Reasoning Texts

SpecVLM: Enhancing Speculative Decoding of Video LLMs via Verifier-Guided Token Pruning

Pareto Actor-Critic for Communication and Computation Co-Optimization in Non-Cooperative Federated Learning Services

Learning to Drive Ethically: Embedding Moral Reasoning into Autonomous Driving

Generative AI Against Poaching: Latent Composite Flow Matching for Wildlife Conservation

Privacy-Aware Detection of Fake Identity Documents: Methodology, Benchmark, and Improved Algorithms (FakeIDet2)

Beyond the Rosetta Stone: Unification Forces in Generalization Dynamics

Steering Towards Fairness: Mitigating Political Bias in LLMs

Dynamic Context Compression for Efficient RAG

Irredundant $k$-Fold Cross-Validation

Prompt Engineering and the Effectiveness of Large Language Models in Enhancing Human Productivity

A Highly Clean Recipe Dataset with Ingredient States Annotation for State Probing Task

Entropy-Memorization Law: Evaluating Memorization Difficulty of Data in LLMs

The Joys of Categorical Conformal Prediction

Adversarial Manipulation of Reasoning Models using Internal Representations

Agent-to-Agent Theory of Mind: Testing Interlocutor Awareness among Large Language Models

A Hybrid Artificial Intelligence Method for Estimating Flicker in Power Systems (Changes are marked)

GLProtein: Global-and-Local Structure Aware Protein Representation Learning

Program Semantic Inequivalence Game with Large Language Models

DSO: Aligning 3D Generators with Simulation Feedback for Physical Soundness

Improving Quantization with Post-Training Model Expansion

Safe and Efficient Social Navigation through Explainable Safety Regions Based on Topological Features

A Simple Approach to Constraint-Aware Imitation Learning with Application to Autonomous Racing

Federated nnU-Net for Privacy-Preserving Medical Image Segmentation

ExPath: Targeted Pathway Inference for Biological Knowledge Bases via Graph Learning and Explanation

Enhancing Automated Loop Invariant Generation for Complex Programs with Large Language Models

RevPRAG: Revealing Poisoning Attacks in Retrieval-Augmented Generation through LLM Activation Analysis

Categorical Data Clustering via Value Order Estimated Distance Metric Learning

Application of AI to formal methods - an analysis of current trends

Reconsidering the Performance of GAE in Link Prediction

See then Tell: Enhancing Key Information Extraction with Vision Grounding

Enhancing Natural Language Inference Performance with Knowledge Graph for COVID-19 Automated Fact-Checking in Indonesian Language

Puppet-Master: Scaling Interactive Video Generation as a Motion Prior for Part-Level Dynamics

FFHFlow: Diverse and Uncertainty-Aware Dexterous Grasp Generation via Flow Variational Inference

SoAy: A Solution-based LLM API-using Methodology for Academic Information Seeking

Investigating the Robustness of Counterfactual Learning to Rank Models: A Reproducibility Study

Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off

Network Formation and Dynamics Among Multi-LLMs

NetGPT: Generative Pretrained Transformer for Network Traffic

OLKAVS: An Open Large-Scale Korean Audio-Visual Speech Dataset

Explainability of Text Processing and Retrieval Methods: A Survey

The Ramon Llull's Thinking Machine for Automated Ideation

RLMR: Reinforcement Learning with Mixed Rewards for Creative Writing

LLM-Based Agents for Competitive Landscape Mapping in Drug Asset Due Diligence

MSARL: Decoupling Reasoning and Tool Use with Multi-Small-Agent Reinforcement Learning

Automated Algorithmic Discovery for Gravitational-Wave Detection Guided by LLM-Informed Evolutionary Monte Carlo Tree Search

Can Large Language Models Develop Strategic Reasoning? Post-training Insights from Learning Chess

Technology as uncharted territory: Contextual integrity and the notion of AI as new ethical ground

Possible Principles for Aligned Structure Learning Agents

OptiMUS-0.3: Using Large Language Models to Model and Solve Optimization Problems at Scale

Prompt-to-Product: Generative Assembly via Bimanual Manipulation

OnGoal: Tracking and Visualizing Conversational Goals in Multi-Turn Dialogue with Large Language Models

Mixture of Contexts for Long Video Generation

FakeParts: a New Family of AI-Generated DeepFakes

Enabling Equitable Access to Trustworthy Financial Reasoning

Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning

Understanding, Protecting, and Augmenting Human Cognition with Generative AI: A Synthesis of the CHI 2025 Tools for Thought Workshop

Inference-Time Alignment Control for Diffusion Models with Reinforcement Learning Guidance

ChainReaction! Structured Approach with Causal Chains as Intermediate Representations for Improved and Explainable Causal Video Question Answering

Train-Once Plan-Anywhere Kinodynamic Motion Planning via Diffusion Trees

ExpertSim: Fast Particle Detector Simulation Using Mixture-of-Generative-Experts

WoW-Bench: Evaluating Fine-Grained Acoustic Perception in Audio-Language Models via Marine Mammal Vocalizations

ProactiveEval: A Unified Evaluation Framework for Proactive Dialogue Agents

Research Challenges in Relational Database Management Systems for LLM Queries

Quantum Verifiable Rewards for Post-Training Qiskit Code Assistant

AI Agentic Vulnerability Injection And Transformation with Optimized Reasoning

JADES: A Universal Framework for Jailbreak Assessment via Decompositional Scoring

Learning Primitive Embodied World Models: Towards Scalable Robotic Learning

Multi-Agent Penetration Testing AI for the Web

Uncertainty Aware-Predictive Control Barrier Functions: Safer Human Robot Interaction through Probabilistic Motion Forecasting

Exploring Machine Learning and Language Models for Multimodal Depression Detection

Speech Emotion Recognition via Entropy-Aware Score Selection

Surfel-based 3D Registration with Equivariant SE(3) Features

Evaluating Compositional Generalisation in VLMs and Diffusion Models

Safer Skin Lesion Classification with Global Class Activation Probability Map Evaluation and SafeML

Unleashing Uncertainty: Efficient Machine Unlearning for Generative AI

Signs of Struggle: Spotting Cognitive Distortions across Language and Register

Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection

Looking Beyond the Obvious: A Survey on Abstract Concept Recognition for Video Understanding

SKGE-SWIN: End-To-End Autonomous Vehicle Waypoint Prediction and Navigation Using Skip Stage Swin Transformer

Occlusion Robustness of CLIP for Military Vehicle Classification

SeqVLM: Proposal-Guided Multi-View Sequences Reasoning via VLM for Zero-Shot 3D Visual Grounding

Provable Benefits of In-Tool Learning for Large Language Models

${C}^{3}$-GS: Learning Context-aware, Cross-dimension, Cross-scale Feature for Generalizable Gaussian Splatting

Rethinking Testing for LLM Applications: Characteristics, Challenges, and a Lightweight Interaction Protocol

EEGDM: Learning EEG Representation with Latent Diffusion Model

Generative Annotation for ASR Named Entity Correction

MobileCLIP2: Improving Multi-Modal Reinforced Training

Task Allocation for Autonomous Machines using Computational Intelligence and Deep Reinforcement Learning

Estimating Worst-Case Frontier Risks of Open-Weight LLMs

Created by

Haebom

저자

Eric Wallace, Olivia Watkins, Miles Wang, Kai Chen, Chris Koch

개요

본 논문은 오픈소스 대규모 언어 모델(gpt-oss) 배포의 최악의 시나리오 위험을 연구합니다. 생물학 및 사이버 보안 두 영역에서 gpt-oss의 능력을 극대화하기 위해 악의적인 미세 조정(MFT) 기법을 도입하여, 생물학적 위험을 극대화하기 위해 위협 생성 관련 작업으로, 사이버 보안 위험을 극대화하기 위해 CTF 챌린지를 해결하는 에이전트 코딩 환경에서 gpt-oss를 미세 조정했습니다. 웹 브라우징이 포함된 강화 학습 환경에서 훈련을 진행했습니다. MFT 모델을 오픈 및 클로즈드-웨이트 LLM과 비교하여 위험 평가를 수행했습니다. 결과적으로, MFT gpt-oss는 생물학적 위험과 사이버 보안 위험 모두에서 Preparedness High 수준 미만인 OpenAI o3 모델에 비해 성능이 떨어졌습니다. 오픈-웨이트 모델과 비교했을 때, gpt-oss는 생물학적 능력을 약간 향상시킬 수 있지만, 상당한 발전을 이루지는 못했습니다. 이러한 결과는 모델 공개 결정에 기여했으며, MFT 접근 방식이 향후 오픈-웨이트 모델 공개로 인한 피해를 추정하는 데 유용한 지침이 될 수 있기를 기대합니다.

시사점, 한계점

•

시사점: 악의적인 미세 조정(MFT) 기법을 활용하여 오픈소스 LLM의 잠재적 위험을 평가하는 새로운 방법론을 제시했습니다. MFT를 통해 실제 위험 수준을 보다 정확하게 예측할 수 있는 가능성을 보여주었습니다. 본 연구의 결과는 오픈소스 모델 배포에 대한 결정을 지원하는 데 유용한 정보를 제공합니다.

•

한계점: MFT는 모든 가능한 악용 시나리오를 포괄하지 못할 수 있습니다. 평가에 사용된 모델의 범위가 제한적일 수 있습니다. 생물학적 및 사이버 보안 위험 평가의 객관성을 더욱 높일 필요가 있습니다. MFT를 통해 얻은 결과가 실제 세계 위험을 완벽하게 반영한다고 단정 지을 수 없습니다.

Made with Slashpage