SciEval (K-12) - Evidence Match Rate: leaderboard

Automatic evaluation of K-12 science lessons on SciEval (2026): given an NGSS-aligned lesson and one EQuIP rubric criterion, the model returns a 0 to 3 score with supporting evidence, compared with expert annotations on the held-out test set. Share of predicted evidence sentences that match an expert evidence sentence by embedding similarity, in percent; higher is better.

Metric: Evidence match rate (%): share of predicted evidence sentences whose best sentence-embedding cosine similarity to an expert evidence sentence exceeds 0.65, on the held-out test set of SciEval, a K-12 science instructional-material evaluation benchmark: given an NGSS lesson as page-marked PDF text and one EQuIP rubric criterion, the model returns a 0-3 score and the supporting evidence as JSON, simplified prompt, temperature 0; failed predictions are left out; higher is better. Source: arxiv.org. Saturation forecast: Around 2034. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 4B 2507 Instruct36.04#2221
2Gemini 2.0 Flash18.27#723
3Llama 3.1 8B Instruct17.25#4600
4Llama 3.2 3B Instruct14.97#7241
5GPT-4o Mini8.63#1667

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

Interactive version: theaggregate.ai/benchmark?slug=scieval-k-12-evidence-match-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.