SafeRelBench - Supporting: leaderboard
SafeRelBench (2026) tests whether VLM-driven embodied agents respect the safety conditions that spatial relations create before risk-prone actions in executable Behavior-1K household tasks. This board is the safety success rate in percent on the supporting-relation samples: episodes that complete the task goal while meeting every safety condition; higher is better.
Metric: Safety success rate (%) on supporting-relation samples: episodes that both complete the BDDL task goal and satisfy every process-level and termination-level safety condition, over all samples; closed-loop step-level planning in Behavior-1K household scenes, base prompt, deterministic decoding; higher is better. Source: arxiv.org. Saturation forecast: Around April 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 49 | #123 |
| 2 | Claude Sonnet 4.6 | 44 | #161 |
| 3 | Qwen 2.5 VL 72B Instruct | 42 | #846 |
| 4 | Gemini 3.1 Pro (Preview) | 40 | #94 |
| 5 | Qwen 3 VL 32B Instruct | 37 | #585 |
| 6 | Qwen 3 VL 8B Instruct | 34 | #981 |
| 7 | Qwen 2.5 VL 32B Instruct | 26 | #1053 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=saferelbench-supporting · How It Works · Data refreshed daily, snapshot 2026-11-13.