PluRule: leaderboard
Community-rule moderation benchmark (PluRule, Kachwala et al., Indiana University, ACL 2026): given a Reddit community's description and full rule list, a comment with its discussion thread, the submission with its images and anonymized author labels, a vision-language model picks which rule the comment breaks or 'No rules broken'. Each of the 2,834 test instances pairs a moderator-removed comment with a compliant one from the same post, so always answering 'No rules broken' scores 50; the score is accuracy in percent over the 5,668 comments, higher is better.
Metric: Accuracy (%; full context, 5,668 comments: each test instance pairs a rule-violating and a compliant comment; always answering 'no rules broken' scores 50). Source: arxiv.org. Saturation forecast: Around May 2028. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 (High) | 57.6 | #172 (GPT-5.2) |
| 2 | GPT-5.2 (Low) | 57.4 | #172 (GPT-5.2) |
| 3 | Qwen 3 VL 30B A3B Instruct | 52.3 | #870 |
| 4 | Qwen 3 VL 8B Instruct | 49.8 | #982 |
| 5 | Qwen 3 VL 30B A3B (Thinking) | 49.5 | #710 |
| 6 | Qwen 3 VL 4B Instruct | 48.4 | #1281 |
| 7 | Qwen 3 VL 4B (Thinking) | 45 | #1303 |
| 8 | Qwen 3 VL 8B (Thinking) | 44.9 | #887 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=plurule · How It Works · Data refreshed daily, snapshot 2026-11-12.