LitXBench - Composition Extraction: leaderboard

Composition extraction on LitXBench (2026): from the OCR text of 19 alloy papers the model outputs each synthesized composition as an atomic-percent string parsed by pymatgen. F1 from 0 to 1; higher is better.

Metric: F1 (0-1) for extracting the experimentally synthesized compositions from the OCR text of LitXAlloy's 19 alloy papers, mean of three runs; the model outputs an atomic-percent string parsed by pymatgen; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.1 Pro (Preview)0.97#99
2GPT-5 Mini0.97#364
3GPT-5.2High0.97#185
4Gemini 3 Flash (Preview)0.94#152
5Claude Opus 4.60.91#113
6Claude Haiku 4.50.78#583

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

Interactive version: theaggregate.ai/benchmark?slug=litxbench-composition-extraction · How It Works · Data refreshed daily, snapshot 2026-11-21.