Scaling Discovery Requires Forgetting

The short version

The problem

Artificial-intelligence agents can stop finding better solutions even when they still have time and computing budget. Their growing record of past attempts can make them stick to small, cautious changes, and adding more agents does not necessarily fix that.

The idea

The paper periodically gives the agents a clean slate: it erases their old conversations and records of explored solutions but saves the best solution found so far. In a team, everyone restarts from that shared best solution, so useful progress is kept while the agents can try fresh directions.

What it got

Across four programming and model-training tasks, teams using this approach performed best on every task. They improved results by as much as 15% over parallel agents given the same budget.