SelfMorph Adapts LLM Computational Topology During Inference

SelfMorph enables frozen LLMs to dynamically change their internal computation shape—via serial reuse, branching, or merging—based on input complexity. A lightweight selector chooses the optimal morphology by analyzing the backbone's depth-wise hidden state trajectory at test-time.