OpenAI paper 'Monitoring Monitorability': RL training doesn't materially erode CoT monitorability
niloofar_mire · x · 2026-10-19
OpenAI researchers including Melody Y. Guan and Bowen Baker released "Monitoring Monitorability" (arXiv:2512.18311), a systematic study of chain-of-thought monitorability.
Key points:
- Proposes three evaluation archetypes (intervention, process, outcome-property) and a new monitorability metric, plus a broad evaluation suite.
- The suite catches model organisms trained with obfuscated CoTs; CoT monitoring beats action-only monitoring in practical settings.
- Frontier model comparison: most models are fairly, but not perfectly, monitorable.
- Longer CoTs are generally more monitorable; RL optimization doesn't materially decrease monitorability at current frontier scale.
- Practical takeaway: swapping a low-reasoning-effort model for a smaller higher-reasoning-effort model yields higher monitorability at equal capability, at higher inference cost.
The tweet recommends the paper (and Boaz Barak's course) to safety researchers as material on deception and loss of monitorability.
More from Safety
- Researchers demo 0-click RCE on Tesla Model 3 via Bluetooth TPMS sensors — matthew_d_green · 2026-10-19
- Anthropic tells Congress Alibaba-linked operators ran 28.8M-exchange distillation attack on Claude — deanwball · 2026-10-19
- Midwit meme nails the AI agent accountability debate: both ends agree, the middle panics — marius_siuram · 2026-10-19
- Developer unsettled as his AI agent asks to solve CAPTCHAs to proceed — akumar · 2026-10-19
- Anthropic confirms draft federal AI safety bill translates its Advanced AI Framework into legislative language — edwinarbus · 2026-10-19
- Was the HuggingFace hacking incident scheming? Victoria Krakovna on OpenAI agents' covert exploits — vkrakovna · 2026-10-19