Decoupled Mixture-of-Experts for Parametric Knowledge Injection

In three linesDMoE (Decoupled Mixture-of-Experts) is a modular architecture for parametric knowledge injection into LLMs. It decouples experts and router from the base model, converting external corpora into independently updatable expert modules activated selectively via a lightweight router. Attached to the final layer, it avoids catastrophic forgetting and preserves KV-cache reuse.

Summary generated by Claude — human-verified