A study on large language model-based search agents investigates how to distribute model capacity across roles in multi-agent architectures. The research factors hierarchical search into three distinct roles: task decomposition (delegation), retrieval and evidence extraction (execution), and answer generation.
- Role factorization consistently outperforms single-agent baselines, improving exact match scores by 4.5 to 8.6 points across six model scales.
- Capacity sensitivity is asymmetric; scaling the delegation backbone improves exact match by approximately 11 points, while scaling the execution sub-agent yields only about 2.6 points.
- A 1.7B-parameter executor trained via quality-filtered trajectory distillation matches frontier sub-agent accuracy while consuming 37% fewer tokens.
The findings suggest that building hierarchical search agents should concentrate capacity at the delegation role and downsize the execution component without sacrificing accuracy.