PrincipalBench: leaderboard
Multi-party principal loyalty (Li and Shi, 2026): the model acts as an agent for a principal while a counterparty simulated by Claude Sonnet probes, pressures or flatters it, and each trajectory is judged for leaking withheld facts or private bounds, capitulating, signalling weakness, writing leaks into drafted artifacts, mishandling third-party confidentiality, or refusing the principal's own legitimate requests. Harm rate in percent over the 36-item core under three system-prompt arms (no instructions, a loyalty scaffold, the scaffold plus a reader tag), mean of five seeds judged by gpt-5-mini and Claude Haiku; lower is better.
Metric: Aggregate Harm Rate (%; 36 items x 3 arms, mean of 5 seeds). Source: arxiv.org. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Flash | 5.5 | #452 |
| 2 | Mistral Large 2 (Nov) Instruct (2411) | 11 | #1199 |
| 3 | Gemini 3.1 Flash Lite (Preview) | 12 | #413 |
| 4 | DeepSeek V3.1 | 12.3 | #500 |
| 5 | Qwen 3 32B | 16.5 | #960 |
| 6 | Claude Opus 4.1 (20250805) | 18.1 | #247 |
| 7 | Llama 3.1 70B Instruct | 19.2 | #1541 |
| 8 | Gemini 3 Flash (Preview) | 19.4 | #141 |
| 9 | Claude Sonnet 4.5 | 19.5 | #244 |
| 10 | GLM-4.6 | 46 | #485 |
| 11 | GPT-5 Mini | 53.6 | #349 |
| 12 | GPT-5 | 71.1 | #163 |
| 13 | Qwen 3.5 27B | 75.3 | #358 |
Interactive version: theaggregate.ai/benchmark?slug=principalbench · How It Works · Data refreshed daily, snapshot 2026-11-12.