Scaling Discovery Requires Forgetting
arXiv cs.AI · 15 Oct 2026, 03:49 AEST
Automated research with LLM agents has shown promising results, but strategies to scale discovery eventually plateau. The key to further scaling is to periodically forget past attempts, allowing agents to restart from the best candidate found.
Key facts
- Automated research with LLM agents has shown promising results in various domains.
- However, we find that all three strategies eventually plateau as the agents accumulate history.
- Periodically discarding the search state, including every agent's session and the candidate history, and restarting from the best candidate found so far, unblocks the search and pushes the same budget further.
- The results show up to 15% improvement compared to parallel agents, and up to 39% compared to a single agent.
Read aloud
Automated research with L L M agents has shown promising results, but strategies to scale discovery eventually plateau. The key to further scaling is to periodically forget past attempts, allowing agents to restart from the best candidate found. Automated research with L L M agents has shown promising results in various domains. However, we find that all three strategies eventually plateau as the agents accumulate history. Periodically discarding the search state, including every agent's session and the candidate history, and restarting from the best candidate found so far, unblocks the search and pushes the same budget further. The results show up to fifteen percent improvement compared to parallel agents, and up to thirty nine percent compared to a single agent.
Sources in this cluster
- Scaling Discovery Requires Forgetting
arXiv cs.AI · 15 Oct 2026, 03:49 AEST