Scaling Discovery Requires Forgetting
Restarting LLM research agents by forgetting search history improves discovery by up to 39 percent versus one agent.
LLM agents used for automated research eventually plateau whether one agent iterates longer, several run in parallel, or parallel agents share candidates, because accumulated history anchors them to fewer and smaller edits. The authors periodically discard every agent session and the candidate history, then restart from the best candidate found so far. Implemented in a communicating agent team and evaluated on heuristic programming, kernel optimization, and model training, forgetting improved results by up to 15% versus parallel agents and up to 39% versus a single agent.
- Longer runs, parallel agents, and communication all plateau as history accumulates.
- Agents make fewer and smaller edits once anchored to past attempts.
- Forgetting discards sessions and history, restarting from the best candidate.
- Gains reached 15% versus parallel agents and 39% versus one agent.
From the original article
1 min read at arxiv.orgAutomated research with LLM agents has shown promising results in various domains. Through iterative refinement and scored verifiers, LLM agents are capable of discovering new ideas. There are different ways to scale automated research: let one agent iterate for longer, run several agents in parallel, or further let the parallel agents communicate and build on one another's candidates.
By Shangyin Tan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.