LLM‑guided evolutionary search can discover complex programs, yet most existing approaches only retain candidate programs and fitness scores, discarding which component edits caused the fitness changes. Consequently, the mutator LLM must infer the effect of prior edits from cluttered histories, making the search slow and unstable—especially for locally servable LLMs that need to evolve multi‑component systems.
We introduce component‑aware feedback: for each evaluated program we compare it with its parent, identify the components that changed, and log those components together with the associated metric differences into an attribution memory that subsequent mutations read. The memory stores each change in two reference frames—local against the parent it originated from and global against the seed program—showing both the immediate impact of a change and the cumulative progress since the seed.
We evaluate the approach on LLM reranking, a multi‑objective optimization problem where a multi‑stage pipeline must balance quality against serving cost. Across twelve \textsc{Bright} datasets, our method reaches the strongest baseline’s final quality after only one‑third of the search budget and achieves $7.2\%$ higher held‑out nDCG@10. Under a cost‑aware objective, it discovers pipelines that are on average more accurate while using $11\%$ fewer tokens per query. These results demonstrate that component‑aware feedback is a promising direction for more efficient self‑evolving systems.
Review