POLAR-Bench: leaderboard
LLM agents increasingly have access to private user data and act on the user's behalf when interacting with third-party systems. The user defines what may and must not be shared, and the agent must robustly follow that intent even when third-party systems behave adversarially. We introduce POLAR-Bench (Policy-aware adversarial Benchmark), in which a trusted model with a privacy policy and a task c
Metric: Overall Mean (self-reported). Source: benchmarklist.com. Saturation forecast: Around August 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.1 | 94.34 | #97 |
| 2 | GPT-5.4 | 92.33 | #79 |
| 3 | Gemma 4 31B | 91.2 | #200 |
| 4 | Gemma 4 E4B | 90.87 | #767 |
| 5 | DeepSeek V3.1 | 90.81 | #264 |
| 6 | Llama 3.3 70B | 86.23 | #579 |
| 7 | Kimi K2.5 | 85.4 | #144 |
| 8 | Gemma 3 27B | 83.16 | #607 |
| 9 | DeepSeek R1 Distill Qwen 32B | 81.19 | #655 |
| 10 | Qwen 3 32B | 80.87 | #443 |
| 11 | GLM-4.7 Flash | 80 | #509 |
| 12 | Gemma 4 E2B | 79.74 | #1070 |
| 13 | Ministral 3 8B | 76.12 | #687 |
| 14 | GPT-OSS-120B | 74.62 | #341 |
| 15 | Ministral 3 14B | 71.96 | #650 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=polar-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.