RuleWeaver - Same-Source - Rubric Score: 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): DeepSeek-V4-Flash's all-or-nothing instance rubric score for the answer; 0-100, higher is better.
Metric: Judge rubric score (0-100). Source: arxiv.org. Saturation forecast: Around September 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (Non-reasoning) | 53.83 | #33 (GPT-5.5) |
| 2 | GPT-5.4 (Non-reasoning) | 48.21 | #83 (GPT-5.4) |
| 3 | Claude Opus 4.6 (Non-reasoning) | 48.06 | #67 (Claude Opus 4.6) |
| 4 | Claude Sonnet 4.6 (Non-reasoning) | 43.96 | #94 (Claude Sonnet 4.6) |
| 5 | GLM-5 (Non-reasoning) | 37.54 | #149 (GLM-5) |
| 6 | DeepSeek V4 Pro (Non-reasoning) | 34.69 | #112 (DeepSeek V4 Pro) |
| 7 | Kimi K2.6 (Non-reasoning) | 32.5 | #113 (Kimi K2.6) |
| 8 | Gemini 3.1 Pro (Preview) (Low) | 31.77 | #64 (Gemini 3.1 Pro (Preview)) |
| 9 | MiniMax-M2.7 | 29.69 | #270 |
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-rubric-score · How It Works · Data refreshed daily, snapshot 2026-11-05.