LithoBench - Open-Ended: leaderboard

LithoBench Open-Ended scores vision-language models on 6,000 open-ended questions about the rock types visible in GF-2 satellite image patches, from identification to geological reasoning (LithoBench, 2026). The OPEN_J score, from 0 to 10 and higher is better, blends ROUGE-L and BLEU-4 overlap with the reference answer and a GPT-5.4 judge rating.

Metric: OPEN_J score (0-10) = 0.4 x ROUGE-L x 10 + 0.2 x BLEU-4 x 10 + 0.4 x the GPT-5.4 judge's overall score (0-10), comparing the answer with the reference answer, averaged over the 6,000 open-ended questions (1,200 per capability level) of LithoBench, remote-sensing lithology interpretation from GF-2 image patches of 12 lithological categories; temperature 0, 1,024 output tokens; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 11 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3.5 397B A17B3.99#264
2Kimi K2.53.92#266
3GLM-4.6V3.89#710
4Intern-S1-Pro3.88#585
5InternVL3.5-241B-A28B3.8#733
6Grok 43.78#319
7Gemini 3 Pro3.68#145
8InternVL3.5-8B-HF3.67
9Qwen 3.5 9B3.23#828
10Qwen 3.5 27B2.49#375
11Claude Sonnet 4.62.29#173

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

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