ForesightSafety Bench (Embodied AI Safety): leaderboard
Metric: Violation rate (%) over all 3,403 scenarios of seven dimensions (item-weighted): unsafe robot task instructions from the paper's ForesightSafetyBench-EmbodiedAI-O set, given to the model as a robot task planner; a violation is an unsafe instruction the model executes instead of refusing, judged by gpt-4o-2024-11-20 (the paper's primary evaluator) with a PAIR-style rating prompt; lower is better. Source: arxiv.org. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 9.02 | #105 |
| 2 | Gemini 3 Flash (Preview) | 11.75 | #78 |
| 3 | Gemini 2.5 Flash | 12.72 | #237 |
| 4 | GLM-4.7 | 37.26 | #185 |
| 5 | Qwen 3 235B A22B 2507 Instruct | 50.51 | #291 |
| 6 | Kimi K2 0905 | 51.78 | #282 |
| 7 | Llama 4 Maverick | 53.48 | #451 |
| 8 | Grok 4 Fast (Non-reasoning) | 53.75 | #242 (Grok 4 Fast) |
| 9 | GPT-4o (2024-11-20) | 54.36 | #369 |
| 10 | Qwen 2.5 72B Instruct | 56.33 | #436 |
| 11 | Grok 4 | 56.33 | #169 |
| 12 | Llama 3.3 70B Instruct | 56.98 | #520 |
| 13 | Kimi K2.5 | 57.54 | #139 |
| 14 | Gemini 3 Pro (Preview) | 58.18 | #64 |
| 15 | DeepSeek V3.2 | 58.68 | #198 |
Interactive version: theaggregate.ai/benchmark?slug=foresightsafety-bench-embodied-ai-safety · How It Works · Data refreshed daily, snapshot 2026-10-11.