Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation研究推理链监测能否识别奖励投机,以及优化过程如何削弱这种可见性。核对原文署名 ↗
Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety将推理链监测作为安全研究机会,并分析其可靠性可能降低的条件。核对原文署名 ↗