RuleWeaver - Same-Source - Rule Recall: leaderboard
RuleWeaver (arXiv 2608.26832v1, Table 2): 11 models answer 96 complex scenario questions, generated with DeepSeek-V4-Pro, with a 200-rule visible pool of gold and distractor rules, reasoning disabled where configurable. Same-Source setting (rules drawn from the same source dataset): share of the five gold rules the model cites or applies; percent, higher is better.
Metric: Gold-rule recall (%). Source: arxiv.org. Saturation forecast: Around March 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.6 (Non-reasoning) | 72.92 | #113 (Kimi K2.6) |
| 2 | Claude Sonnet 4.6 (Non-reasoning) | 71.67 | #94 (Claude Sonnet 4.6) |
| 3 | Claude Opus 4.6 (Non-reasoning) | 70.42 | #67 (Claude Opus 4.6) |
| 4 | GPT-5.5 (Non-reasoning) | 69.17 | #33 (GPT-5.5) |
| 5 | MiniMax-M2.7 | 65 | #270 |
| 6 | DeepSeek V4 Pro (Non-reasoning) | 63.75 | #112 (DeepSeek V4 Pro) |
| 7 | GLM-5 (Non-reasoning) | 62.5 | #149 (GLM-5) |
| 8 | GPT-5.4 (Non-reasoning) | 61.67 | #83 (GPT-5.4) |
| 9 | Gemini 3.1 Pro (Preview) (Low) | 60.83 | #64 (Gemini 3.1 Pro (Preview)) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ruleweaver-same-source-rule-recall · How It Works · Data refreshed daily, snapshot 2026-11-05.