Station Framework Enables AI Agents to Rediscover 62.7% of Scientific Findings in Open-Ended Tasks

The Station framework demonstrates that AI agents can autonomously undertake open-ended scientific discovery, achieving a 62.7% average rediscovery rate of original findings. This performance significantly exceeds Codex Multiagent-v2 (15.4%) and AI Scientist-v2 (14.4–20.6%).

Mechanisms for Persistent Exploration

To address the lack of intermediate metrics in open-ended tasks, Station introduces two key mechanisms:

• Supervisor mechanism: Guides agent behavior.
• Periodic Meta Reflection: Encourages persistent exploration.

These additions improve research coverage and continuity, as confirmed by ablation and behavioral analyses.

Evaluation Methodology

Researchers constructed open-ended tasks from three recent oral papers presented at ICLR. Agents were given the main research question but withheld the paper's results and web access. The study measured how many individual criteria from the original findings the agents could rediscover.

Real-World Validation

Beyond benchmark papers, Station was evaluated on two open-ended tasks without oracle papers. Some discoveries made by the agents closely matched findings reported by researchers after the model's knowledge cutoff date, indicating meaningful progress in autonomous scientific inquiry.