RuleSafe-VL - Decision Relation Identification: leaderboard
Task 2 of RuleSafe-VL (Lu, Wang, Shang and Xu, Beijing University of Posts and Telecommunications, 2026), a benchmark of rule-conditioned content moderation built from platform policies on nudity, dangerous behavior and graphic injury: from a raw policy statement and its atomic rules, the model labels how the rules relate (independent, gating, threshold shift or override); the task is text only. Macro-F1 in percent over the 92 expert-labelled relations, averaged over the three policy families; higher is better.
Metric: Macro-F1 (%; typed relations among a policy's atomic rules from the policy text, mean of three policy families). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 64.8 | #91 |
| 2 | Claude Opus 4.6 | 63.4 | #100 |
| 3 | GLM-5.1 | 55.8 | #175 |
| 4 | MiniMax-M2.7 | 50.7 | #519 |
| 5 | Qwen 2.5 VL 7B Instruct | 25.6 | #1842 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rulesafe-vl-decision-relation-identification · How It Works · Data refreshed daily, snapshot 2026-11-12.