Symbolic-MoE introduces a fine-grained expert routing system that breaks complex problems into specific skills, selecting the optimal LLM for each reasoning component.

• Creates a symbolic representation of tasks to identify required skills
• Dynamically routes queries to specialized LLM experts based on the identified skills
• Achieves improved performance without requiring gradient-based training
• Enables adaptive instance-level mixing of pre-trained models for heterogeneous reasoning tasks

Medical Impact: Enhances diagnostic reasoning by efficiently routing complex medical queries to specialized experts, improving performance on medical question answering tasks like MedMCQA and handling biomedical reasoning with greater accuracy.

Symbolic Mixture-of-Experts: Adaptive Skill-based Routing for Heterogeneous Reasoning