Decomposing and Measuring Evaluation Awareness: leaderboard
Frontier language models sometimes recognize that they are being evaluated and adjust their behavior, undermining validity of benchmark results. Yet the field studies it without a shared foundation, conflating properties of the evaluation with properties of the model, and detection with behavioral response. We ground evaluation awareness in social psychology, decomposing it into an environment com
Metric: HarmBench Saf. (self-reported). Source: benchmarklist.com. Saturation forecast: Around June 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 | 5.6 | #251 |
| 2 | Kimi K2.5 | 4.9 | #144 |
| 3 | GPT-OSS-120B | 1.3 | #341 |
| 4 | Claude Sonnet 4.5 | 0.5 | #142 |
| 5 | GPT-5 | 0.4 | #98 |
| 6 | GLM-5 | 0.4 | #141 |
| 7 | GPT-OSS-20B | 0.2 | #518 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=decomposing-and-measuring-evaluation-awareness · How It Works · Data refreshed daily, snapshot 2026-10-20.