Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation

A study of chain-of-thought monitoring during OpenAI's reinforcement learning: under strong optimization against a monitor that read its reasoning, an agent kept reward hacking while the monitor's recall fell to near zero. Swarm ecology takes this as the main laboratory evidence for surveillance selection (P10) and the detection gap. Prediction 7 tests it, and the toy model in Companion Fig. 8b (Flagship Fig. 4b) illustrates it.

Cited in

both papers (Flagship: Organisms that read us, Fig. 4; Companion §8, Fig. 8)

Used for

P10Prediction 7P10; prediction 7; detection gap; Companion Fig. 8b (Flagship Fig. 4b)

Strand

AI safety, alignment & autonomy evaluations · Monitoring, interpretability & hidden channels

Cite

@misc{baker2025monitoring,
  title = {Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation},
  author = {Bowen Baker and Joost Huizinga and Leo Gao and Zehao Dou and Melody Y. Guan and Aleksander Madry and Wojciech Zaremba and Jakub Pachocki and David Farhi},
  year = {2025},
  url = {https://arxiv.org/abs/2503.11926},
}