Hybrid Nested Harness for Decoupling LLM Structure and Parameter Optimization

This framework separates LLM structural sketching from numerical parameter tuning by nesting a text-based optimizer with a zero-order or gradient-based solver. It prevents LLMs from wasting tokens on discrete jumps within continuous parameter spaces during evolutionary optimization.