OpenAI studies CoT monitoring limits and imperfect proxies
November 10 research examines how training affects control over reasoning text.
OpenAI observed that reinforcement learning can increase models’ control over chain-of-thought text, and that model-generated training data can pass this ability to a subsequent generation.
The researchers caution that this is an imperfect proxy for monitoring evasion and does not always mean more evasion. Vendor research, not independent validation, a product rollout or a safety guarantee. We did not reproduce the experiments.
The researchers caution that this is an imperfect proxy for monitoring evasion and does not always mean more evasion. Vendor research, not independent validation, a product rollout or a safety guarantee. We did not reproduce the experiments.
Usage conditions
- Hong Kong
- Research announcement, not local service availability; no local feature, web/app/API, language or pricing change is established.
- Macau
- Research announcement, not local service availability; no local feature, web/app/API, language or pricing change is established.
- Taiwan
- Research announcement, not local service availability; no local feature, web/app/API, language or pricing change is established.
Official source: What affects chain-of-thought monitorability across the model lifecycle?