Taxonomy of Model Fusion for Large Language Models
September 18, 2026
This survey categorizes LLM model fusion into three hierarchical levels: parameter-level, representation-level, and behavior-level. It provides a systematic framework for integrating capabilities from multiple source models into single target models to improve reuse and capability density.
HOW THIS AFFECTS YOU
●
builderYou can use this taxonomy to select appropriate fusion strategies when merging model capabilities.
●
researcherThis provides a structured roadmap for exploring new integration techniques beyond simple parameter averaging.