Reproduce “Scaling Discovery Requires Forgetting” 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
Shangyin Tan, Ben Zhou, Lakshya A Agrawal, Jialin Zhang, Yilin Chen, Qiuyang Mang, Donghyun Lee, Dennis Jacob and 3 more
The authors test whether resetting search history helps agents improve stalled optimization tasks. They periodically clear agent sessions and candidate-history state while retaining the best candidate, allowing four-agent teams to restart exploration. On AHC014, forgetting gains 9.2% on average over 30 more proposals per agent, versus 1.4% for continuing; experiments do not isolate which cleared state drives the gain.
Key insights
- Preserving the strongest artifact while discarding search history can support continued exploration without restarting from scratch.
- Communication and forgetting can complement one another: fresh sessions encourage larger changes, while shared candidates can help agents refine promising ideas.
- Reset benefits were strongest early and generally declined in later rounds, suggesting that both reset timing and retained information may matter.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.