ForesightSafety Bench (Embodied AI Safety) - Living-Being Physical Contact Safety: leaderboard
Metric: Violation rate (%) over the living-being physical contact safety scenarios: 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 | 11.11 | #105 |
| 2 | Gemini 2.5 Flash | 15.28 | #237 |
| 3 | Gemini 3 Flash (Preview) | 16.11 | #78 |
| 4 | GLM-4.7 | 41.94 | #185 |
| 5 | Grok 4 Fast (Non-reasoning) | 52.92 | #242 (Grok 4 Fast) |
| 6 | Qwen 3 235B A22B 2507 Instruct | 53.89 | #291 |
| 7 | Kimi K2 0905 | 54.58 | #282 |
| 8 | Llama 4 Maverick | 55.69 | #451 |
| 9 | GPT-4o (2024-11-20) | 58.19 | #369 |
| 10 | Llama 3.3 70B Instruct | 59.03 | #520 |
| 11 | Claude 3.5 Haiku | 59.31 | #553 |
| 12 | Grok 4 | 59.58 | #169 |
| 13 | Qwen 2.5 72B Instruct | 60 | #436 |
| 14 | Kimi K2.5 | 61.67 | #139 |
| 15 | DeepSeek V3.2 | 61.67 | #198 |
Interactive version: theaggregate.ai/benchmark?slug=foresightsafety-bench-embodied-ai-safety-living-being-physical-contact-safety · How It Works · Data refreshed daily, snapshot 2026-10-11.