ForesightSafety Bench (Embodied AI Safety) - Collaborative Norm Safety: leaderboard
Metric: Violation rate (%) over the collaborative norm 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 | Gemini 3 Flash (Preview) | 30.86 | #78 |
| 2 | Kimi K2.5 | 33.33 | #139 |
| 3 | Gemini 3 Pro (Preview) | 33.33 | #64 |
| 4 | GPT-5.2 | 40.74 | #105 |
| 5 | Gemini 2.5 Flash | 44.44 | #237 |
| 6 | Claude 3.5 Haiku | 49.38 | #553 |
| 7 | Grok 4 Fast (Non-reasoning) | 49.38 | #242 (Grok 4 Fast) |
| 8 | GPT-4o (2024-11-20) | 55.56 | #369 |
| 9 | GLM-4.7 | 56.79 | #185 |
| 10 | Seed-1.6 | 58.02 | #257 |
| 11 | Qwen 3 235B A22B 2507 Instruct | 59.26 | #291 |
| 12 | Grok 4 | 60.49 | #169 |
| 13 | Qwen 2.5 72B Instruct | 61.73 | #436 |
| 14 | Kimi K2 0905 | 61.73 | #282 |
| 15 | Seed 1.8 | 61.73 | #136 |
Interactive version: theaggregate.ai/benchmark?slug=foresightsafety-bench-embodied-ai-safety-collaborative-norm-safety · How It Works · Data refreshed daily, snapshot 2026-10-11.