New Techniques Enhance Realism in AI Alignment Evaluations

A significant challenge in evaluating AI alignment is "evaluation awareness," where advanced models can discern when they are undergoing testing rather than operating in a real deployment. This awareness can compromise the validity of safety evaluations. This paper introduces two novel techniques designed to make simulated alignment evaluations more indistinguishable from actual deployments. The first technique, called critique refinement, allocates additional inference-time compute for each action taken by the simulator. The simulator generates multiple potential actions, then refines them using feedback from an instance of the target model itself, aiming to make them more realistic, before proceeding with the most deployment-like option. The second technique, named DISH (Deployment-Imitating SWE-Agent Harness), involves wrapping the target model within an agent harness. This approach significantly narrows the gap between simulated and real deployment environments, particularly in coding-related scenarios. The researchers tested both techniques across various target models and found that they are complementary, with applying both yielding greater realism improvements than either technique used in isolation. The findings indicate that automated methods can effectively enhance the realism of alignment evaluations, and that these improvements leverage additional computational resources more efficiently than simply extending the duration of audits.