Towards Alignment Auditing for RL Environments
> Thesis: Auditing what RL environments reward is a promising and actionable direction for improving frontier-model alignment. These environments provide a concrete point of intervention: their prompts, sandboxes, and graders can be inspected and revised when they reward behavior we do not intend to teach. Embedded evaluators are a valuable...
Sep 1817