Reproduce “ScienceFlow: A long-horizon agent for ML research, scientific discovery and beyond” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mingming Zhao, Jiqian Dong, Kangping Xu, Zadid Hasan, Chengrui Fan, Shan Jiang, Shuai Mao, Yating Ling and 11 more
ScienceFlow, developed by Noah's Ark Lab at Huawei, introduces an autonomous research agent that integrates recoverable executable states, adaptive research trajectories, and evidence-aware execution control for long-horizon tasks. This framework achieves a 70.22% Any-Medal rate on the full MLE-bench, outperforms baselines on SciModelingBench with a score of 54.41, and obtains competitive results in mathematical and engineering optimization.
Key insights
- Effective long-horizon autonomous research critically depends on the joint maintenance of complete executable states, adaptive research trajectories, and objective-aligned resource-aware execution.
- The ability to exactly restore an archived executable state, rather than just a textual summary, is fundamental for reliable research continuity, branching, and recovery from unproductive paths.
- Separating scientific decision-making (by research workers) from physical execution management (by a controller) optimizes resource allocation and ensures compute resources are aligned with validated research progress and budget constraints.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.