Intel / article

Lineage Is Memory: Parent Inheritance and Feedback Disclosure in Short-Budget LLM-Guided Optimization

Source: arxiv.org
Author: Shanglin Yang, Zhan Shi
Date: Published Nov 2, 2026

Why it matters

When building prompt or program optimizers, preserving the surviving parent matters more than showing the proposer a broader summary of failures, so spend effort on lineage first.

Key takeaways · AI-distilled

The authors argue that earlier tests of richer reflection (full-set error clustering, targeted minibatches, larger proposers) all ran inside lineage-preserving search, so the benefit of keeping a parent was confounded with how much failure feedback the proposer saw.
Full-set error clustering and diverse failure selection do improve single rewrites from the seed and speed up the opening rounds, but the paper reports sequential inheritance closes that gap because solved error modes stay in the candidate and drop out of later failure lists.
Breaking down GEPA's spend, the authors find proposer reflection is only a small fraction of total cost, while validation and minibatch gating consume nearly the entire budget.
The study spans nine prompt and program synthesis benchmarks under matched budgets, and the authors release code, pre-registrations and logs for all 1,750 runs.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source arxiv.org