Reproduce “Accelerating scientific discovery with Co-Scientist” 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
Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Petar Sirkovic, Artiom Myaskovsky, Grzegorz Glowaty, Felix Weissenberger and 43 more
Co-Scientist develops a researcher-steered process in which specialist AI agents generate, critique, compare, and refine research hypotheses, while a supervisor allocates computation. Across 203 goals, later outputs received higher Elo ratings; on 15 expert-curated goals, its rating eventually exceeded tested models and expert guesses. In vitro, KIRA6 showed an 18-fold IC₅₀ difference between KG-1a and TK6, not evidence of clinical benefit.
Key insights
- Hypothesis quality consistently improves over time with increased computational thinking (test-time compute scaling), demonstrating a capacity for continuous self-improvement without explicit re-training.
- Co-Scientist refines and enhances initial 'best guess' solutions provided by human domain experts, with these AI-refined hypotheses eventually surpassing those generated autonomously by the system.
- Quantitative ablation studies confirmed the necessity of the multi-agent architecture, showing that specialized agents and their unique strategies (e.g., Reflection agent's external search for hallucination reduction) are crucial for robust scientific reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.