一句话：RIDER 是一个进化式提示优化框架，通过自适应算子选择与错误导向进化，显著提升 LLM 辅助软件工程任务的提示质量。
LLM 辅助的软件工程工具高度依赖提示质量，但提示通常依靠手工反复试错构造。本文提出 RIDER，一个元启发式框架：它维护一个提示种群，使用基于 LLM 的算子进行进化，并以验证集反馈驱动搜索。其核心贡献包括三点：（1）将自适应算子选择建模为多臂老虎机问题，在九种算子上采用窗口化 Thompson 采样；（2）错误导向进化，将误分类样例反馈回生成过程；（3）evaluate-first 的 (μ+λ) 保留策略，并以多样性打破平局。作者在覆盖分类、问答、生成与代码智能的六个基准上评估 RIDER，在约 3500 次 API 调用的可比预算下，RIDER 在全部六个基准上均排名第一，平均超过最佳基线 14.5%。
LLM-assisted SE tools depend critically on prompt quality, yet prompts are typically hand-crafted via trial and error. We present RIDER, a metaheuristic framework that evolves a population of prompts using LLM-based operators and validation-set feedback. RIDER contributes (1) adaptive operator selection as a multi-armed bandit with Windowed Thompson Sampling over nine operators, (2) error-directed evolution that feeds misclassified examples back into generation, and (3) evaluate-first (μ+λ) retention with diversity tiebreaking. On six benchmarks spanning classification, QA, generation, and code intelligence, RIDER ranks first on all six, outperforming the best baseline by 14.5% on average at a comparable ~3,500 API-call budget.
@inproceedings{DragomirovKMMSM26,
title = {RIDER: Evolutionary Prompt Optimization with Adaptive Operator Selection for Software Engineering},
author = {Daglar Dragomirov and Nikita Kulin and Sergey Muravyov and Ilya Makarov and Daniil Sukhorukov and Mikhail Mozikov},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3807393}
}