This paper introduces ASRouter, a framework that addresses the issue of prompt defects in Large Language Models by routing prompts to specialized processing modules based on detected defect types. The method uses a learned gating network for defect identification and dynamic assignment to expert modules. The routing problem is formalized as a constrained optimization with load balancing, and a curriculum learning strategy is proposed for stable expert specialization. Experiments show a 23% improvement in output quality over baselines.
Key findings
ASRouter improves LLM output quality by 23% over uniform baselines.
Learned routing outperforms rule-based dispatching, especially for prompts with multiple defect types.
A curriculum learning strategy promotes stable expert specialization.
Limitations & open questions
The framework's performance may vary across different types of LLMs.
The complexity of the routing mechanism might introduce additional computational overhead.