Berkeley paper: making research agents forget boosts automated discovery by up to 39%
omarsar0 · x · 2026-10-16
A UC Berkeley paper, Scaling Discovery Requires Forgetting, proposes a counterintuitive way to scale automated research agents.
- Problem: Whether you run one agent longer, parallel agents, or parallel agents that communicate, research agents plateau as they accumulate history — each agent anchors on its own past attempts, makes smaller edits, and extra budget stops helping.
- Method: Induce forgetting. Periodically discard every agent's session and candidate history, then restart from the best candidate found so far, unblocking the search.
- Results: Across heuristic programming (AtCoder), kernel optimization (Anthropic's VLIW kernel), and model training tasks, a team of 4 agents with forgetting improves up to 15% over parallel agents and up to 39% over a single agent at the same budget.
Takeaway: the key to scaling discovery isn't more memory — it's learning to forget.
Related event: Berkeley Paper: Teaching AI Research Agents to Forget Boosts Discovery(3 posts)→
More from coding & agent
- VibeZero: one-shot an entire vibe coding platform on Cloudflare Void — cnakazawa · 2026-10-16
- Ant and Tsinghua open-source Realtime-Venus, an OpenAI Realtime API alternative — aigclink · 2026-10-16
- 'I remember when I was the harness for o1 pro' — dev on becoming the API for his model — timigod · 2026-10-16
- JevAgentHarness: a 4B decision model lets cheap agents offload tool choices from the LLM — youjiaxuan · 2026-10-16
- Coding agents keep opening 3 stacked PRs per chat, and merging them is a nightmare — holdenmatt · 2026-10-16
- Google-led paper uses sequential auctions to organize decentralized LLM agent workflows — _onionesque · 2026-10-16