Evaluating Chain-of-Thought Monitorability

When AI systems make complex decisions, the final output alone makes it difficult to fully grasp the process behind it. Therefore, monitoring the model's internal reasoning process, Chain-of-thought (CoT), is far more effective at detecting malfunctions such as deception or reward hacking.

This research introduces a framework and 13 evaluation tools (24 environments) for systematically measuring the Monitorability of CoT. The evaluation methods consist of three types: intervention, process, and outcome-property.

The results showed that most state-of-the-art reasoning models exhibited considerably high levels of monitorability. In particular, monitoring performance tended to improve as models 'thought' longer and the Chain-of-thought grew longer. Additionally, current levels of reinforcement learning (RL) optimization were found not to significantly degrade monitorability.

Meanwhile, a trade-off between reasoning effort and model size was also observed. A smaller model using high reasoning effort may be easier to monitor than a larger model operating with low reasoning effort. However, this comes with a 'monitorability tax' that increases reasoning cost.