README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: question
      dtype: string
    - name: choices
      list: string
    - name: answer
      dtype: int32
    - name: meta_info
      struct:
        - name: title
          dtype: string
        - name: journal
          dtype: string
        - name: doi
          dtype: string
        - name: url
          dtype: string
    - name: question_type
      dtype: string
  splits:
    - name: en
      num_bytes: 546653187.125
      num_examples: 1525
    - name: zh
      num_bytes: 546319847.125
      num_examples: 1525
  download_size: 218606009
  dataset_size: 1092973034.25
configs:
  - config_name: RxnBench-VQA
    data_files:
      - split: en
        path: data/en-*
      - split: zh
        path: data/zh-*
license: cc-by-nc-sa-4.0
task_categories:
  - visual-question-answering
language:
  - en
  - zh
tags:
  - chemistry

RxnBench: A Benchmark for Chemical Reaction Figure Understanding

📘 Benchmark Summary

RxnBench (SF-QA) is a visual question answering (VQA) benchmark comprising 1,525 multiple-choice questions (MCQs) at the PhD-level of organic chemistry reaction understanding.

The benchmark is built from 305 scientific figures drawn from high-impact OpenAssess journals. For each figure, domain experts carefully designed five multiple-choice VQA questions targeting the interpretation of organic reaction diagrams. These questions were further refined through multiple rounds of rigorous review and revision to ensure both clarity and scientific accuracy. The questions cover a variety of types, including the description of chemical reaction images, extraction of reaction content, recognition of molecules or Markush structures, and determination of mechanisms. This benchmark challenges visual-language models on their foundational knowledge of organic chemistry, multimodal contextual reasoning, and chemical reasoning skills.

The benchmark is released in both English and Chinese versions.

📑 Task Types

We categorize chemical reaction visual question answering tasks into six types:

• Type 0 — Fact Extraction: Direct retrieval of textual or numerical information from reaction schemes.
• Type 1 — Reagent Roles and Functions Identification: Identification of reagents and their functional roles, requiring chemical knowledge and reaction-type awareness.
• Type 2 — Reaction Mechanism and Process Understanding: Interpretation of reaction progression, including intermediates, catalytic cycles, and mechanistic steps.
• Type 3 — Comparative Analysis and Reasoning: Comparative evaluation, causal explanation, or outcome prediction under varying conditions.
• Type 4 — Multi-step Synthesis and Global Understanding: Comprehension of multi-step pathways, step-to-step coherence, and overall synthetic design.
• Type 5 — Chemical Structure Recognition: Extraction and reasoning-based parsing of chemical structures in SMILES or E-SMILES (as defined in the MolParser paper).

🎯 Benchmark Evaluation

This benchmark evaluates model performance on multiple-choice question answering (MCQ) tasks.

We provide two versions of the prompt template, depending on the language setting.

English Prompt

Question: {question}
Choices:
A. {choice_A}
B. {choice_B}
C. {choice_C}
D. {choice_D}
Based on the image and the question, choose the most appropriate answer.
**Only output a single letter (A, B, C, or D)**. Do NOT output any other text or explanation.

Chinese Prompt

问题: {question}
选项:
A. {choice_A}
B. {choice_B}
C. {choice_C}
D. {choice_D}

请根据图像和问题，从以上四个选项中选择最合适的答案。
只输出单个字母 (A, B, C 或 D)，不要输出选项内容，也不要输出任何解释。

Evaluation Protocol

If the model’s output is not one of A, B, C, or D, we use GPT-4o to map the output to A–D based on the option content. The final evaluation reports the absolute accuracy of the benchmark in both English and Chinese versions.

Code Example: https://github.com/uni-parser/RxnBench

📊 Benchmark Leaderboard

We evaluated several of the latest popular MLLMs, including both closed-source and open-source models.

Moldel	Think	Weight	API-Version	RxnBench-En	RxnBench-Zh	Mean Score
Gemini-3-Pro-preview	√	Proprietary	20251119	0.9318	0.9403	0.9361
GPT-5(high)	√	Proprietary	20250807	0.9279	0.9246	0.9263
Gemini-2.5-Pro	√	Proprietary	20250617	0.9095	0.9423	0.9259
GPT-5.1(high)	√	Proprietary	20251113	0.9213	0.9220	0.9216
GPT-5(medium)	√	Proprietary	20250807	0.9207	0.9226	0.9216
Qwen3-VL-235BA22B-Think	√	Open	-	0.9220	0.9134	0.9177
Qwen3-VL-32B-Think	√	Open	-	0.9128	0.9161	0.9144
GPT-5.1(medium)	√	Proprietary	20251113	0.9108	0.9141	0.9125
GPT-5-mini	√	Proprietary	20250807	0.9108	0.9128	0.9118
Seed1.5-VL-Think	√	Proprietary	20250428	0.9056	0.9161	0.9109
GPT o3	√	Proprietary	20250416	0.9056	0.9115	0.9086
GPT o4 mini	√	Proprietary	20250416	0.9062	0.9075	0.9069
InternVL3.5-241B-A28B	√	Open	-	0.9003	0.9062	0.9033
Intern-S1	√	Open	-	0.8938	0.8944	0.8941
Qwen3-VL-30BA3B-Think	√	Open	-	0.8689	0.8590	0.8689
Qwen3-VL-Plus	×	Proprietary	20250923	0.8551	0.8656	0.8604
Qwen3-VL-8B-Think	√	Open	-	0.8636	0.8564	0.8600
Seed1.5-VL	×	Proprietary	20250328	0.8518	0.8669	0.8594
Qwen3-VL-235BA22B-Instruct	×	Open	-	0.8492	0.8675	0.8584
InternVL3-78b	×	Open	-	0.8531	0.8308	0.8420
Qwen3-VL-4B-Think	√	Open	-	0.8577	0.8256	0.8416
Intern-S1-mini	√	Open	-	0.8521	0.8282	0.8402
GLM-4.1V-9B-Thinking	√	Open	-	0.8392	0.8341	0.8367
Qwen3-VL-32B-Instruct	×	Open	-	0.8315	0.8407	0.8361
Qwen2.5-VL-72B	×	Open	-	0.8341	0.8308	0.8325
Qwen2.5-VL-Max	×	Proprietary	20250813	0.8192	0.8262	0.8227
GPT-5-nano	√	Proprietary	20250807	0.7980	0.7941	0.7961
Qwen2.5-VL-32B	×	Open	-	0.7980	0.7908	0.7944
Gemini-2.5-Flash	√	Proprietary	20250617	0.6925	0.8557	0.7741
Qwen3-VL-8B-Instruct	×	Open	-	0.7548	0.7495	0.7521
Qwen3-VL-30BA3B-Instruct	×	Open	-	0.7456	0.7436	0.7456
GPT-4o	×	Proprietary	20240806	0.7462	0.7436	0.7449
Qwen2.5-VL-7B	×	Open	-	0.7082	0.7233	0.7158
Qwen3-VL-4B-Instruct	×	Open	-	0.7023	0.7023	0.7023
Qwen3-VL-2B-Think	√	Open	-	0.6780	0.6708	0.6744
Qwen2.5-VL-3B	×	Open	-	0.6748	0.6643	0.6696
GPT-4o mini	×	Proprietary	20240718	0.6636	0.6066	0.6351
Qwen3-VL-2B-Instruct	×	Open	-	0.5711	0.5928	0.5820
Choice longest answer	-	-	-	0.4262	0.4525	0.4394
Deepseek-VL2	×	Open	-	0.4426	0.4216	0.4321
Random	-	-	-	0.2500	0.2500	0.2500

We also conducted separate evaluations for different task types (in RxnBench-en).

Moldel	Think	Weight	API-Version	Type0	Type1	Type2	Type3	Type4	Type5
Gemini-3-Pro-preview	√	Proprietary	20251119	0.9648	0.9246	0.9527	0.9398	0.9322	0.7463
GPT-5(high)	√	Proprietary	20250807	0.9313	0.9444	0.9527	0.9167	0.9661	0.8358
Gemini-2.5-Pro	√	Proprietary	20250617	0.9331	0.9246	0.9459	0.9491	0.9322	0.6343
GPT-5.1(high)	√	Proprietary	20251113	0.9243	0.9524	0.9426	0.9167	0.9661	0.7910
GPT-5(medium)	√	Proprietary	20250807	0.9349	0.9325	0.9493	0.9167	0.9492	0.7761
Qwen3-VL-235BA22B-Think	√	Open	-	0.9190	0.9405	0.9459	0.9213	0.9322	0.8433
Qwen3-VL-32B-Think	√	Open	-	0.9296	0.9405	0.9426	0.9259	0.9153	0.7015
GPT-5.1(medium)	√	Proprietary	20251113	0.9243	0.9365	0.9426	0.9167	0.9492	0.7090
GPT-5-mini	√	Proprietary	20250807	0.9225	0.9325	0.9257	0.9259	0.9831	0.7388
Seed1.5-VL-Think	√	Proprietary	20250428	0.8996	0.9365	0.9358	0.9074	0.9153	0.8060
GPT o3	√	Proprietary	20250416	0.9313	0.9325	0.9223	0.8981	0.9492	0.7090
GPT o4 mini	√	Proprietary	20250416	0.6391	0.7302	0.7500	0.6667	0.6271	0.4627
InternVL3.5-241B-A28B	√	Open	-	0.8944	0.9127	0.9291	0.9167	0.9153	0.8134
Intern-S1	√	Open	-	0.9014	0.9127	0.9223	0.9028	0.8814	0.7463
Qwen3-VL-30BA3B-Think	√	Open	-	0.8732	0.8810	0.9054	0.8843	0.9322	0.6940
Qwen3-VL-Plus	×	Proprietary	20250923	0.8275	0.8968	0.8986	0.8565	0.9153	0.7687
Qwen3-VL-8B-Think	√	Open	-	0.8768	0.8730	0.8885	0.9028	0.8983	0.6567
Seed1.5-VL	×	Proprietary	20250328	0.9327	0.9127	0.9122	0.8472	0.8305	0.7015
Qwen3-VL-235BA22B-Instruct	×	Open	-	0.8204	0.8929	0.8986	0.8426	0.8814	0.7761
InternVL3-78b	×	Open	-	0.8556	0.8730	0.8885	0.8981	0.9153	0.6194
Qwen3-VL-4B-Think	√	Open	-	0.8838	0.8770	0.8615	0.9074	0.8983	0.6045
Intern-S1-mini	√	Open	-	0.8239	0.8690	0.8547	0.8611	0.8475	0.6791
GLM-4.1V-9B-Thinking	√	Open	-	0.8433	0.8690	0.8649	0.8657	0.8814	0.6493
Qwen3-VL-32B-Instruct	×	Open	-	0.8169	0.8571	0.8885	0.8519	0.8305	0.6866
Qwen2.5-VL-72B	×	Open	-	0.8063	0.8063	0.8770	0.9088	0.8102	0.9322
Qwen2.5-VL-Max	×	Proprietary	20250813	0.7958	0.8571	0.8885	0.8194	0.8983	0.6642
GPT-5-nano	√	Proprietary	20250807	0.8063	0.8452	0.8311	0.8241	0.7797	0.5672
Qwen2.5-VL-32B	×	Open	-	0.7729	0.8413	0.8750	0.8009	0.8305	0.6418
Gemini-2.5-Flash	√	Proprietary	20250617	0.7799	0.6111	0.6757	0.6620	0.7627	0.5373
Qwen3-VL-8B-Instruct	×	Open	-	0.7113	0.8175	0.8446	0.8241	0.7627	0.5075
Qwen3-VL-30BA3B-Instruct	×	Open	-	0.7042	0.7937	0.8311	0.7824	0.7119	0.5970
GPT-4o	×	Proprietary	20240806	0.7359	0.8175	0.7973	0.7500	0.7627	0.5224
Qwen2.5-VL-7B	×	Open	-	0.6678	0.7659	0.8041	0.7130	0.6441	0.5373
Qwen3-VL-4B-Instruct	×	Open	-	0.6708	0.7302	0.7804	0.7222	0.6610	0.5970
Qwen3-VL-2B-Think	√	Open	-	0.7342	0.6706	0.7128	0.7083	0.6102	0.3657
Qwen2.5-VL-3B	×	Open	-	0.6426	0.7381	0.7635	0.6898	0.6610	0.4776
GPT-4o mini	×	Proprietary	20240718	0.6391	0.7302	0.7500	0.6667	0.6271	0.4627
Qwen3-VL-2B-Instruct	×	Open	-	0.5405	0.6190	0.6318	0.6250	0.6102	0.3731
Deepseek-VL2	×	Open	-	0.4120	0.5040	0.4899	0.4907	0.3729	0.3060

🆕 RxnBench-Doc

A single reaction image often lacks the information needed for full interpretation, requiring contextual text from the literature. Therefore, we also provide a benchmark for chemical reaction literature understanding.

https://huggingface.co/datasets/UniParser/RxnBench-Doc

📖 Citation

our paper coming soon ...