🧪 ChemO Dataset

📄 Paper: ChemLabs on ChemO: A Multi-Agent System for Multimodal Reasoning on IChO 2025

ChemO Version 1.0 is now publicly available! 🎉

The ChemO dataset has been officially released after meticulous proofreading and preparation. This benchmark is built from the International Chemistry Olympiad (IChO) 2025 and represents a new frontier in automated chemical problem-solving.

🌟 Key Features

• 🏆 Olympic-Level Benchmark - Challenging problems from IChO 2025 for advanced AI reasoning
• 🔬 Multimodal Symbolic Language - Addresses chemistry's unique combination of text, formulas, and molecular structures
• 📊 Two Novel Assessment Methods:
  • AER (Assessment-Equivalent Reformulation) - Converts visual output requirements (e.g., drawing molecules) into computationally tractable formats
  • SVE (Structured Visual Enhancement) - Diagnostic mechanism to separate visual perception from core chemical reasoning capabilities

📦 What's Included

The current release includes:

• ✅ Original Problems - Complete problem sets with additional chapter markers for Problems and Solutions sections (no other modifications to the original content)
• ✅ Well-structured JSON Files - Clean, organized data designed for:
  • 🤖 MLLM Benchmarking - Olympic-level chemistry reasoning evaluation
  • 🔗 Multi-Agent System Testing - Hierarchical agent collaboration assessment
  • 🎯 Multimodal Reasoning - Text, formula, and molecular structure understanding
• ⏳ Original CDXML Files - Coming soon (see note below)

📋 Dataset Structure

The ChemO dataset consists of 9 problems from IChO 2025, with each problem provided as a structured JSON file (1.json ~ 9.json in JSON/). The JSON/images/ directory contains all referenced images indexed in the JSON files.

📚 Data Source

All problems are sourced from ICHO 2025: https://www.icho2025.ae/problems

📝 Note on CDXML Files

Due to compatibility issues across different ChemDraw versions, the CDXML files for molecular structures are not included in the initial v1.0 release. We are actively working to resolve these compatibility challenges and will supplement the dataset with CDXML files in a future update.

🚀 State-of-the-Art Results

Our ChemLabs multi-agent system combined with SVE achieves 93.6/100 on ChemO, surpassing the estimated human gold medal threshold and establishing a new benchmark in automated chemical problem-solving.

🤝 Community

We appreciate your patience and look forward to your feedback as we continue to improve this resource for the community.

📄 Citation

If you use ChemO in your research, please cite our paper:

@article{qiang2025chemlabs,
  title={ChemLabs on ChemO: A Multi-Agent System for Multimodal Reasoning on IChO 2025},
  author={Xu, Qiang and Bai, Shengyuan and Chen, Leqing and Liu, Zijing and Li, Yu},
  journal={arXiv preprint arXiv:2511.16205},
  year={2025}
}