aiqbench analysis
This research addresses the critical issue of "evaluation awareness," where advanced models detect they are being tested rather than deployed, thereby skewing safety assessments. By leveraging additional inference-time compute to refine simulation actions and wrap targets in realistic agent harnesses, the authors demonstrate that simulated evaluations can closely mimic real-world deployment conditions. These techniques offer a more efficient path to robust alignment verification compared to simply extending audit durations.
Improving Evaluation Realism with Inference-Time Compute and Deployment Scaffolds
arXiv:2609.02302v1 Announce Type: cross Abstract: A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety evaluation can support. We present two techniques that make simulated alignment evaluations harder to distinguish from real deployments. Our first technique, critique refinement, spends additional inference-time compute on each simulator action: the simulator generates multiple candidate actions, refines them using feedback from an instance of the target model on how to make them more realistic, and continues the evaluation with the most deployment-like candidate. Our second technique, DISH (Deployment-Imitating SWE-Agent Harness), wraps the target in an agent harness, reducing the gap between simulated and real deployment environments in coding settings. We test the techniques on multiple target models and find that they compose: applying both yields larger realism gains than either alone. Our results show that automated approaches can improve the realism of alignment evaluations, and that these improvements use additional compute more effectively than making the audits longer.
Abstract:A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety evaluation can support. We present two techniques that make simulated alignment evaluations harder to distinguish from real deployments. Our first technique, critique refinement, spends additional inference-time compute on each simulator action: the simulator generates multiple candidate actions, refines them using feedback from an instance of the target model on how to make them more realistic, and continues the evaluation with the most deployment-like candidate. Our second technique, DISH (Deployment-Imitating SWE-Agent Harness), wraps the target in an agent harness, reducing the gap between simulated and real deployment environments in coding settings. We test the techniques on multiple target models and find that they compose: applying both yields larger realism gains than either alone. Our results show that automated approaches can improve the realism of alignment evaluations, and that these improvements use additional compute more effectively than making the audits longer.
70 pages, 43 figures, 4 tables (13 figures in the main text). Under review at NeurIPS 2026. Code: this https URL and this https URL ; reproduction assets: this https URL
Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2609.02302 [cs.AI]
(or arXiv:2609.02302v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2609.02302 arXiv-issued DOI via DataCite (pending registration)
Full text under the publisher's open licence · 221 words
Related coverage
[AINews] OpenAI reports Navier-Stokes singularity find in 88 hours using Astra-next, roughly 10,000 agents and 130B tokens (>$40M), a contender for second ever Millennium Prize awarded
Latent Space · Sep 9, 2026
Improving Audit Realism With Inference Time Compute and Deployment Scaffolds
LessWrong · Sep 3, 2026
Construct validity of Claude Opus 4.8's System Card – A commentary
LessWrong · Jun 12, 2026
Concrete Evaluations to Investigate the OpenAI Model That Hacked Hugging Face
AI Alignment Forum · Aug 3, 2026