Scaling Discovery Requires Forgetting
Resetting agents’ search history while keeping their best candidate yielded gains of up to 39% over a single agent.
The authors report that longer runs, parallel agents and communicating teams all eventually plateau as accumulated history anchors agents to earlier attempts. Their method periodically clears every agent’s session and the candidate history, then restarts from the best candidate found. Tested on four tasks spanning heuristic programming, kernel optimization and model training, it improved results by up to 15% over parallel agents at the same budget.
ArXiv · AI/CL/LG's note
The authors report that longer runs, parallel agents and communicating teams all eventually plateau as accumulated history anchors agents to earlier attempts. Their method periodically clears every agent’s session and the candidate history, then restarts from the best candidate found. Tested on four tasks spanning heuristic programming, kernel optimization and model training, it improved results by up to 15% over parallel agents at the same budget.
ArXiv · AI/CL/LG's note
score 5