Learning to Be Monitored: Causal Bounds and Intervention-Aware Post-Training Recipes for Chain-of-Thought Monitorability
Fu et al.
The standard metric for chain-of-thought monitorability is shown to be an upper bound on the true causal detection rate; the authors add a lower bound and train models whose reasoning is easier to monitor.
Source published · In the Daily Brief of 27 October 2026
Filed under Reasoning and inference, AI safety
Chain-of-thought monitoring tries to catch harmful behaviour from a model’s written reasoning. Intervention-based tests ask whether a monitor can see the effect of an intervention in the reasoning when the target behaviour occurs, usually through a minimal-criterion sensitivity \(\mathrm{TPR}_{\mathrm{mc}}\). The authors point out that behaviour observed under an intervention is not always caused by it.
They prove that \(\mathrm{TPR}_{\mathrm{mc}}\) upper-bounds the rate at which a monitor detects behaviour actually caused by the intervention, and derive a matching lower bound. Using paired control and intervention tasks, they then train models by off-policy self-distillation, reinforcement learning and on-policy self-distillation, and all three generally improve monitorability under the tighter evaluation.