#AI Safety

21개의 포스트

[논문리뷰] MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models

arXiv에 게시된 'MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Models' 논문에 대한 자세한 리뷰입니다.

#Review #Large Language Models #Chain-of-Thought #Monitorability #Benchmark #AI Safety #Stress-Test #Faithfulness

2026년 3월 31일

[논문리뷰] The Reasoning Trap -- Logical Reasoning as a Mechanistic Pathway to Situational Awareness

Divya Chaudhary이 arXiv에 게시한 'The Reasoning Trap -- Logical Reasoning as a Mechanistic Pathway to Situational Awareness' 논문에 대한 자세한 리뷰입니다.

#Review #Logical Reasoning #Situational Awareness #LLMs #Deceptive Alignment #AI Safety #RAISE Framework #Self-Modeling #Deduction #Induction #Abduction

2026년 3월 10일

[논문리뷰] SAHOO: Safeguarded Alignment for High-Order Optimization Objectives in Recursive Self-Improvement

Divya Chaudhary이 arXiv에 게시한 'SAHOO: Safeguarded Alignment for High-Order Optimization Objectives in Recursive Self-Improvement' 논문에 대한 자세한 리뷰입니다.

#Review #Recursive Self-Improvement #Alignment Drift #AI Safety #Goal Drift Index (GDI)#Constraint Preservation #Regression Risk #Capability Alignment Ratio (CAR)

2026년 3월 10일

[논문리뷰] Reasoning Models Struggle to Control their Chains of Thought

arXiv에 게시된 'Reasoning Models Struggle to Control their Chains of Thought' 논문에 대한 자세한 리뷰입니다.

#Review #Chain-of-Thought (CoT)#Model Controllability #AI Safety #Monitorability #Large Language Models (LLMs)#Reinforcement Learning (RL)#Evaluation Suite

2026년 3월 8일

[논문리뷰] Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use

arXiv에 게시된 'Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use' 논문에 대한 자세한 리뷰입니다.

#Review #Agentic LLM #AI Safety #Multi-Step Tool Use #Reinforcement Learning #Preference-Based Learning #Safety Guardrails #Refusal Mechanism #Structured Reasoning

2026년 3월 3일

[논문리뷰] Exposing the Systematic Vulnerability of Open-Weight Models to Prefill Attacks

arXiv에 게시된 'Exposing the Systematic Vulnerability of Open-Weight Models to Prefill Attacks' 논문에 대한 자세한 리뷰입니다.

#Review #Large Language Models #Prefill Attacks #AI Safety #Red Teaming #Vulnerability #Open-Weight Models #Jailbreaking #Generative AI

2026년 2월 16일

[논문리뷰] The Devil Behind Moltbook: Anthropic Safety is Always Vanishing in Self-Evolving AI Societies

Jinyu Hou이 arXiv에 게시한 'The Devil Behind Moltbook: Anthropic Safety is Always Vanishing in Self-Evolving AI Societies' 논문에 대한 자세한 리뷰입니다.

#Review #Multi-agent Systems #Self-evolution #AI Safety #Alignment Drift #Information Theory #Thermodynamics #Entropy Accumulation #Moltbook

2026년 2월 12일

[논문리뷰] A Safety Report on GPT-5.2, Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5

Yutao Wu이 arXiv에 게시한 'A Safety Report on GPT-5.2, Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5' 논문에 대한 자세한 리뷰입니다.

#Review #AI Safety #Large Language Models #Multimodal LLMs #Benchmark Evaluation #Adversarial Robustness #Multilingual Evaluation #Regulatory Compliance #Image Generation Safety

2026년 1월 15일

[논문리뷰] Steerability of Instrumental-Convergence Tendencies in LLMs

j-hoscilowic이 arXiv에 게시한 'Steerability of Instrumental-Convergence Tendencies in LLMs' 논문에 대한 자세한 리뷰입니다.

#Review #LLM Steerability #Instrumental Convergence #AI Safety #AI Security #Open-Weight Models #Prompt Engineering #Model Control #Behavioral Alignment

2026년 1월 6일

[논문리뷰] K-EXAONE Technical Report

arXiv에 게시된 'K-EXAONE Technical Report' 논문에 대한 자세한 리뷰입니다.

#Review #Multilingual Language Model #Mixture-of-Experts (MoE)#Long Context #AI Safety #Korean AI #Foundation Model #Reinforcement Learning (RL)

2026년 1월 5일

[논문리뷰] COMPASS: A Framework for Evaluating Organization-Specific Policy Alignment in LLMs

arXiv에 게시된 'COMPASS: A Framework for Evaluating Organization-Specific Policy Alignment in LLMs' 논문에 대한 자세한 리뷰입니다.

#Review #LLM Evaluation #Policy Alignment #Organizational Policies #AI Safety #Adversarial Robustness #Refusal Behavior #Prompt Engineering #Fine-tuning

2026년 1월 5일

[논문리뷰] Reinventing Clinical Dialogue: Agentic Paradigms for LLM Enabled Healthcare Communication

Hengshu Zhu이 arXiv에 게시한 'Reinventing Clinical Dialogue: Agentic Paradigms for LLM Enabled Healthcare Communication' 논문에 대한 자세한 리뷰입니다.

#Review #Clinical Dialogue #LLM Agents #Healthcare AI #Agentic Paradigm #Medical Decision Support #Knowledge Grounding #AI Safety #Workflow Automation

2025년 12월 10일

[논문리뷰] AI & Human Co-Improvement for Safer Co-Superintelligence

arXiv에 게시된 'AI & Human Co-Improvement for Safer Co-Superintelligence' 논문에 대한 자세한 리뷰입니다.

#Review #AI Safety #Superintelligence #Human-AI Collaboration #Self-Improving AI #Co-Improvement #Alignment #AI Research Agents

2025년 12월 7일

[논문리뷰] Machine Text Detectors are Membership Inference Attacks

Naoaki Okazaki이 arXiv에 게시한 'Machine Text Detectors are Membership Inference Attacks' 논문에 대한 자세한 리뷰입니다.

#Review #Membership Inference Attacks #Machine-Generated Text Detection #Transferability #Likelihood Ratio Test #Large Language Models #Zero-Shot Detection #Model Security #AI Safety

2025년 10월 23일

[논문리뷰] RefusalBench: Generative Evaluation of Selective Refusal in Grounded Language Models

arXiv에 게시된 'RefusalBench: Generative Evaluation of Selective Refusal in Grounded Language Models' 논문에 대한 자세한 리뷰입니다.

#Review #RAG Systems #Selective Refusal #Generative Evaluation #Linguistic Perturbations #LLM Evaluation #Informational Uncertainty #Model Calibration #AI Safety

2025년 10월 17일

[논문리뷰] Blueprints of Trust: AI System Cards for End to End Transparency and Governance

Roman Zhukov이 arXiv에 게시한 'Blueprints of Trust: AI System Cards for End to End Transparency and Governance' 논문에 대한 자세한 리뷰입니다.

#Review #AI Governance #Transparency #AI System Card #Hazard-Aware System Card #Data Provenance #AI Safety #AI Risk Management #ISO/IEC 42001

2025년 9월 26일

[논문리뷰] FlagEval Findings Report: A Preliminary Evaluation of Large Reasoning Models on Automatically Verifiable Textual and Visual Questions

tengdai722이 arXiv에 게시한 'FlagEval Findings Report: A Preliminary Evaluation of Large Reasoning Models on Automatically Verifiable Textual and Visual Questions' 논문에 대한 자세한 리뷰입니다.

#Review #Large Reasoning Models #LLM Evaluation #Multimodal AI #Reasoning Behaviors #Hallucination #Contamination-Free #AI Safety #Instruction Following

2025년 9월 23일

[논문리뷰] R^textbf{2AI}: Towards Resistant and Resilient AI in an Evolving World

Bowen Zhou이 arXiv에 게시한 'R^textbf{2AI}: Towards Resistant and Resilient AI in an Evolving World' 논문에 대한 자세한 리뷰입니다.

#Review #AI Safety #Resistant AI #Resilient AI #Coevolution #Fast-Slow Models #Adversarial Training #Continual Learning #AGI Alignment

2025년 9월 9일

[논문리뷰] False Sense of Security: Why Probing-based Malicious Input Detection Fails to Generalize

Muhao Chen이 arXiv에 게시한 'False Sense of Security: Why Probing-based Malicious Input Detection Fails to Generalize' 논문에 대한 자세한 리뷰입니다.

#Review #LLM Safety #Malicious Input Detection #Probing Classifiers #Out-of-Distribution Generalization #Superficial Patterns #Instructional Patterns #Trigger Words #AI Safety

2025년 9월 5일

[논문리뷰] CorrSteer: Steering Improves Task Performance and Safety in LLMs through Correlation-based Sparse Autoencoder Feature Selection

Adriano Koshiyama이 arXiv에 게시한 'CorrSteer: Steering Improves Task Performance and Safety in LLMs through Correlation-based Sparse Autoencoder Feature Selection' 논문에 대한 자세한 리뷰입니다.

#Review #Sparse Autoencoders #LLM Steering #Feature Selection #Correlation Analysis #AI Safety #Bias Mitigation #Mechanistic Interpretability

2025년 8월 20일

[논문리뷰] A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems

Xinhao Yi이 arXiv에 게시한 'A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems' 논문에 대한 자세한 리뷰입니다.

#Review #Self-Evolving AI Agents #Lifelong Learning #Foundation Models #Multi-Agent Systems #Agent Optimization #Prompt Engineering #Tool Use #AI Safety #Survey

2025년 8월 12일