ReactBench (Reaction Diagrams): leaderboard

ReactBench (2026) tests how multimodal models reason about the structure of chemical reaction diagrams drawn from journals and patents, with 1,618 expert-annotated questions scored by exact match. The score is the unweighted mean accuracy over four task dimensions, in percent, higher is better. It is unrelated to the ReactBench React UI coding benchmark.

Metric: Answer accuracy (%) on the four task dimensions (unweighted mean of localization, extraction, tracing and reasoning accuracy), over the 1,618 expert-annotated question-answer pairs on more than 1,300 real chemical reaction diagrams from chemistry journals and patents of ReactBench (single-line, multi-line, tree and graph topologies), multiple-choice or short open answers matched exactly against the key after template extraction, no LLM judge; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 24 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 32B Instruct73.91#606
2Gemini 3.1 Pro (Preview)73.88#99
3Qwen 3 VL 32B (Thinking)70.89#573
4GPT-5.569.59#54
5Qwen 2.5 VL 72B68.66#781
6Claude 3.5 Sonnet68.6#752
7Claude Sonnet 4.667#173
8Qwen-VL-Max (ReactBench checkpoint unspecified)65.03
9GPT-4o63.83#751
10Qwen 3 VL 8B Instruct63.02#1029
11Gemini 1.5 Pro62.17#1063
12Qwen 3 VL 8B (Thinking)60.24#910
13Qwen 2.5 VL 7B57.32#1862
14InternVL2.5-78B53.74#1031
15InternVL2.5-8B-MPO50.38#2012

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

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