Tag
This paper proposes Activation-Prune-Merge (APM), a training-free framework for cross-scale fusion that improves smaller language models using larger donors without semantic alignment, achieving performance gains on multiple benchmarks.