Frontier Model Forum

All else equal, prefer architectures that require CoTs for serial reasoning, until interpretability methods can provide similar levels of insight.

Develop and publish monitorability evaluations. Standardized evaluations and metrics can improve assessments for how monitorable an AI model’s reasoning process is. These tests would measure the clarity, coherence, and faithfulness of a model’s Chain...

[...] if the evaluation results indicate [...] an exploitable vulnerability that may lead to a significant increase in biorisks, this information should not be published.