Predicting When RL Training Breaks Chain-of-Thought Monitorability
Introduces a conceptual framework predicting when RL training pressure makes chain-of-thought unmonitorable, with empirical tests on a coin-flip task and obfuscated reward hacking.}
Research
Google DeepMind Safety Research
Safety research from Google DeepMind. deepmindsafetyresearch.medium.com
Focus All
Topic Alignment + Control
Source DeepMind · 2
Date Any time
1 item
1 item
An Approach to Technical AGI Safety and Security
Google DeepMind outlines its technical approach to preventing severe harms from AGI, covering misuse and misalignment risks in foundation models and prioritising mitigations for foreseeable capabilities.