LLM‑based multi‑agent systems improve complex reasoning by assigning specialized roles and enabling collaboration, yet the collaboration pattern must adapt to the backbone model's competence on the current query. Excessive coordination incurs extra token cost and redundant interactions, while insufficient coordination can degrade accuracy. To address this, we introduce a two‑stage competence‑aware orchestration framework. The first stage builds a task‑conditioned competence model that predicts the backbone's confidence for the given query. The second stage uses this signal to dynamically activate roles and construct a communication graph, ensuring that multi‑agent interaction occurs only when beneficial.
Across six public benchmarks the approach attains an average score of 93.81%. Experiments with six backbones of varying competence reveal that for high‑competence models the average accuracy rises from 87.22% to 89.97% while token usage drops by 36.26%; for low‑competence models the method adds a further 4.58‑point gain over existing techniques. The same trend holds for unseen backbones, demonstrating strong generalization.
The key insight is treating competence prediction as a scheduling signal, allowing the system to preserve performance while substantially reducing resource consumption.
Review