Qualcomm AI Research and David Krueger find a post-training step can restore chain-of-thought monitorability

arXiv · Research preprint

Read the source (opens in a new tab)
What happened
Usman Anwar, Tim Bakker and Christos Louizos, whose work was done at Qualcomm AI Research, and David Krueger of Mila posted a preprint on arXiv on October 21, 2026.
Key points
  • They call chain-of-thought monitoring "one of the few practical tools available for overseeing capable AI systems, but it is fragile," degraded by compression, optimization against a monitor and training on synthetic documents.
  • Their unsupervised post-training objective "recovers most or all of the lost monitorability" in three model organisms whose reasoning had been degraded.
  • On models without degraded reasoning, it "improves prediction of unsafe behavior by CoT monitors and activation probes."
  • They argue "monitorability should be treated as a trainable attribute of the policy," with a dedicated post-training stage.
Evidence
Research preprint, not peer reviewed, tested on open-weight models including GPT-OSS-120B and Qwen2.5-7B.
Uncertain
The authors say their evaluations "use controlled model organisms," so whether the gains extend to "naturally occurring monitoring failures in stronger agents" remains to be established.

See it in the latest record