Prim - Discovery: leaderboard

Prim's primitive-discovery score: from the problem alone, the model must state its mathematical primitive, the one key idea that organizes the solution. An LLM judge grades it against a human-verified gold primitive over 182 Humanity's Last Exam mathematics problems; higher is better (Prim, 2026).

Metric: Discovery score (%): how well the model states the problem's mathematical primitive (the key organizing idea) from the problem alone, judged against the human-verified gold primitive by GPT-5.4-High on a graded rubric, over 182 problems; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 4 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.4High82.42#138
2GPT-5.4 MiniHigh59.34#405
3GPT-5.4 NanoHigh41.76#651
4GPT-OSS-20BHigh34.62#1166
5Qwen 3.5 27B28.57#378
6Qwen 3.6 27B24.73#352
7Qwen 3.5 9B13.19#829
8DeepSeek-R1-Distill-Qwen-7B7.14#7912
9Qwen 3.5 4B6.59#1235
10DeepSeek R1 0528 Qwen3 8B6.59#2190
11DeepSeek R1 Distill Qwen 32B6.04#1813
12DeepSeek R1 Distill Qwen 14B4.4#2706

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=prim-discovery · How It Works · Data refreshed daily, snapshot 2026-11-20.