ResMerge: Residual-based Spectral Merging of Large Language Models

Researchers introduced ResMerge, a training-free framework for merging multiple reinforcement-learning language model experts.

It splits task vectors into a strong spectral head and a residual part, then builds a stable backbone with reliability-weighted residual consensus.

ResMerge adds head information back through agreement-gated correction to reduce conflicts between experts.

Across multiple expert groups and capability areas, it preserved learned abilities better than prior task-vector and spectral merging methods.