RuleWorld: leaderboard

Metric: Exact-match accuracy (%; x100 of the 0-1 score; mean of the six task-representation cells (single-rule, parallel multi-rule over 2-8 rules, and 2-4 hop multi-hop QA, each with rules in first-order logic and 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 December 2026. 10 models tracked.

Top models

#ModelScore
1Claude Sonnet 4.677.63
2GPT-5.572.83
3GPT-5.167.96
4DeepSeek V3.2 (Non-reasoning)66.87
5Llama 3.1 70B Instruct54.9
6Qwen 2.5 72B Instruct53.07
7Llama 3.3 70B Instruct51.62
8Qwen 3 32B51.32
9Qwen 2.5 32B Instruct50.77
10Qwen 2.5 14B Instruct42.97

Interactive version: theaggregate.ai/benchmark?slug=ruleworld · How It Works · Data refreshed daily, snapshot 2026-09-26.