# Improving Evaluation Realism with Inference-Time Compute and Deployment Scaffolds

arXiv: https://arxiv.org/abs/2609.02302 · HTML: https://mlsanity.com/paper/2609.02302/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Reasoning for agents, planning and code, one of 17 directions in LLM reasoning.
- Builds on 9 of its 20 references that are in the map.
- First posted on arXiv 2 Sep 2026.

## Placed in

- LLM reasoning / [Reasoning for agents, planning and code](https://mlsanity.com/lr/reasoning-for-agents-planning-and-code/)

## Builds on (most cited first, 9 of 9)

- [Self-Refine: Iterative Refinement with Self-Feedback](https://mlsanity.com/paper/2303.17651/) (2023)
- [Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters](https://mlsanity.com/paper/2408.03314/) (2024)
- [WildChat: 1M ChatGPT Interaction Logs in the Wild](https://mlsanity.com/paper/2405.01470/) (2024)
- [Frontier Models are Capable of In-context Scheming](https://mlsanity.com/paper/2412.04984/) (2024)
- [Large Language Models Often Know When They Are Being Evaluated](https://mlsanity.com/paper/2505.23836/) (2025)
- [Evaluating whether AI models would sabotage AI safety research](https://mlsanity.com/paper/2604.24618/) (2026)
- [UK AISI Alignment Evaluation Case-Study](https://mlsanity.com/paper/2604.00788/) (2026)
- [LURE: Live-Usage Replay Evaluations for Reducing Evaluation Awareness](https://mlsanity.com/paper/2605.26438/) (2026)
- [Gram: Assessing sabotage propensities via automated alignment auditing](https://mlsanity.com/paper/2605.30322/) (2026)

## Abstract (arXiv)

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.