A major obstacle in alignment evaluation is evaluation awareness: capable models can detect that they are being tested, which weakens the safety conclusions drawn. We introduce two techniques that make simulated evaluations harder to distinguish from real deployments. The first, critique refinement, adds inference‑time compute to each simulated action: the simulator produces several candidate actions, refines them using feedback from an instance of the target model on how to make them more realistic, and proceeds with the most deployment‑like candidate. The second, DISH (Deployment‑Imitating SWE‑Agent Harness), wraps the target in an agent harness, narrowing the gap between simulated and real deployment environments in coding tasks. Experiments on multiple target models show that the methods compose: applying both yields larger realism gains than either alone. The results demonstrate that automated approaches can improve alignment evaluation realism, and that the extra compute is used more efficiently than simply extending audit duration.
Review