InstrumentalEval - Hiding Unwanted Behavior: leaderboard
The 11 InstrumentalEval tasks in which the model, pushed to reach its goal at all costs, could hide actions from monitoring; the score is the share of responses a GPT-4o judge marks as doing so. It is one of six categories behind the InstrumentalEval total.
Metric: Responses pursuing the instrumental goal, 11 tasks (%, lower is better). Source: arxiv.org. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3 | 27.27 | #716 |
| 2 | GPT-4o | 36.36 | #741 |
| 3 | Gemini 2.0 Flash (01-21) (Thinking) | 36.36 | #711 (Gemini 2.0 Flash (01-21)) |
| 4 | O1 Mini | 54.55 | #764 |
| 5 | DeepSeek R1 | 63.64 | #473 |
| 6 | O1 (2024-12-17) | 63.64 | #311 |
| 7 | O3 Mini | 63.64 | #533 |
Interactive version: theaggregate.ai/benchmark?slug=instrumentaleval-hiding-unwanted-behavior · How It Works · Data refreshed daily, snapshot 2026-11-13.