ReactBench (Reaction Diagrams) - Localization: leaderboard

ReactBench chemical reaction diagram questions that ask which nodes are sources, sinks or intermediates from their position in the diagram topology. Exact-match accuracy in percent, higher is better. It is unrelated to the ReactBench React UI coding benchmark.

Metric: Answer accuracy (%) on the spatial element localization questions (identify source, sink and intermediate nodes from topological position), 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 December 2027. 24 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 32B (Thinking)63.59#573
2Qwen 3 VL 32B Instruct60.24#606
3Gemini 3.1 Pro (Preview)59.16#99
4Claude Sonnet 4.655.69#173
5GPT-5.555.21#54
6Qwen 3 VL 8B Instruct52.34#1029
7Claude 3.5 Sonnet50.78#752
8Qwen 3 VL 8B (Thinking)49.22#910
9Qwen-VL-Max (ReactBench checkpoint unspecified)49.22
10Qwen 2.5 VL 72B46.47#781
11Gemini 1.5 Pro45.75#1063
12DeepSeek-VL2-Tiny45.39#6534
13GPT-4o44.07#751
14InternVL2.5-78B41.08#1031
15InternVL2.5-8B-MPO38.2#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-localization · How It Works · Data refreshed daily, snapshot 2026-11-21.