RuleWorld - Multi-Hop (Natural Language): leaderboard
RuleWorld (2026) multi-hop questions chaining 2 to 4 rules in sequence, with all 1,000 rules of the pool written in natural language and placed in the prompt. Exact-match accuracy in percent, with partial credit per matching answer position; higher is better.
Metric: Exact-match accuracy (%; x100 of the 0-1 score; multi-hop rule QA with 2 to 4 hops, mean of the three hop counts, rules in natural language; all 1,000 abstract, non-commonsense procedural rules of the pool prepended to the prompt (full-context prompting); exact match on the boxed answer list, partial credit per matching position; 10 questions per task type and difficulty level for each of five seeds; temperature 0 (Qwen3-32B 0.6)). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 64.33 | #158 |
| 2 | GPT-5.5 (2026-04-23) | 46.67 | #33 |
| 3 | DeepSeek V3.2 (Non-reasoning) | 38 | #403 (DeepSeek V3.2) |
| 4 | Qwen 3 32B | 33.33 | #960 |
| 5 | GPT-5.1 (2025-11-13) | 28 | #184 |
| 6 | Qwen 2.5 32B Instruct | 22 | #1216 |
| 7 | Qwen 2.5 14B Instruct | 19.33 | #1745 |
| 8 | Llama 3.3 70B Instruct | 18.67 | #1353 |
| 9 | Qwen 2.5 72B Instruct | 18.67 | #1093 |
| 10 | Llama 3.1 70B Instruct | 14.67 | #1541 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ruleworld-multi-hop-natural-language · How It Works · Data refreshed daily, snapshot 2026-11-12.