safety security
Reasoning Monitorability Becomes a Training Target
An unsupervised mutual-information objective recovered much of the monitoring signal lost in three degraded chain-of-thought model organisms.
Summary
An unsupervised mutual-information objective recovered much of the monitoring signal lost in three degraded chain-of-thought model organisms.
Post-training strengthened the connection between internal reasoning text and model outputs, improving judged monitorability and prediction of unsafe behavior in the reported tests. The result suggests monitorability can be optimized, but it does not make chain-of-thought a complete or tamper-proof oversight channel.
Why it matters
An unsupervised mutual-information objective recovered much of the monitoring signal lost in three degraded chain-of-thought model organisms.
Limits and context
- The result suggests monitorability can be optimized, but it does not make chain-of-thought a complete or tamper-proof oversight channel.
Key claims
An unsupervised mutual-information objective recovered much of the monitoring signal lost in three degraded chain-of-thought model organisms.
Qualification: The result suggests monitorability can be optimized, but it does not make chain-of-thought a complete or tamper-proof oversight channel.
Evidence: source-2026-10-22-019
Sources
- arXiv preprint 2610.29734, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.