K12-Bench - Experiment Evidence: leaderboard

The Experiment Evidence task of K12-Bench (2026), built from the K12-KGraph knowledge graph of Chinese K-12 mathematics, physics, chemistry and biology textbooks, as four-option multi-select questions answered zero-shot: link concepts to the experiments that verify them, in both directions. Exact match in percent (every option must be chosen correctly); higher is better.

Metric: Experiment Evidence exact match (%, all-or-nothing over the selected option set) on four-option multi-select questions generated from the K12-KGraph curriculum knowledge graph of Chinese K-12 mathematics, physics, chemistry and biology textbooks; zero-shot, answer-only, temperature 0: link concepts and the experiments that verify them, in both directions (1,391 items); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Flash47.4#180
2Gemini 2.5 Flash47.1#466
3Gemma 4 31B (IT)43.3#354
4Qwen 3 32B41.7#983
5GPT-5.241.6#190
6GLM-4.7 Flash39.3#1349
7Ministral-3-14B-Instruct38.3#1151
8GPT-5 Mini32.9#373
9GPT-4o32.4#754
10Llama 3 8B Instruct5.2#4858

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=k12-bench-experiment-evidence · How It Works · Data refreshed daily, snapshot 2026-11-22.