OpenAI：2025 AI思维链可监控性评估报告（英文版）（90页）.pdf

1、Monitoring MonitorabilityMelody Y.GuanMiles WangMicah CarrollZehao DouAnnie Y.WeiMarcus WilliamsBenjamin ArnavJoost HuizingaIan KivlichanMia GlaeseJakub PachockiBowen BakerOpenAIAbstractObservability into the decision making of modern AI systems may be required to safelydeploy increasingly capable a

2、gents.Monitoring the chain-of-thought(CoT)of today s reasoningmodels has proven effective for detecting misbehavior.However,this“monitorability”may befragile under different training procedures,data sources,or even continued system scaling.Tomeasure and track monitorability,we propose three evaluati

3、on archetypes(intervention,process,and outcome-property)and a new monitorability metric,and introduce a broad evaluationsuite.We demonstrate that these evaluations can catch simple model organisms trained to haveobfuscated CoTs,and that CoT monitoring is more effective than action-only monitoring in

4、practical settings.We compare the monitorability of various frontier models and find that mostmodels are fairly,but not perfectly,monitorable.We also evaluate how monitorability scaleswith inference-time compute,reinforcement learning optimization,and pre-training model size.We find that longer CoTs

5、 are generally more monitorable and that RL optimization does notmaterially decrease monitorability even at the current frontier scale.Notably,we find thatfor a model at a low reasoning effort,we could instead deploy a smaller model at a higherreasoning effort(thereby matching capabilities)and obtai

6、n a higher monitorability,albeit at ahigher overall inference compute cost.We further investigate agent-monitor scaling trends andfind that scaling a weak monitor s test-time compute when monitoring a strong agent increasesmonitorability.Giving the weak monitor access to CoT not only improves monito