Reproduce “Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents” 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
Ziyang Cai, Christos Ziakas, Vasilis Kontonis, Tim Pearce, Siddhartha Sen, Akshay Krishnamurthy, Shivam Garg, Dimitris Papailiopoulos
Fork-and-flush organizes autoresearch by periodically preserving an agent’s workspace, clearing its chat history, and launching parallel continuations, then carrying forward the highest-scoring branch. Across 13 tasks under matched compute budgets, it achieved a cross-task min–max normalized mean score of 0.78, versus 0.47 for single-run and 0.54 for best-of-4. It led on 12 tasks; LWE-rec was the exception.
Key insights
- Candidate artifacts from the same run tend to occupy localized regions of the solution space, consistent with runs becoming attached to distinct approaches or “idea basins.”
- Flushing conversation history while preserving workspace artifacts can encourage strategic reconsideration without losing implementation progress.
- On a MAX-3-SAT analysis, run-to-run outcome variation was greatest at intermediate task difficulty and lower when tasks were very easy or beyond the model’s capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.