SAPO: Segment-level Automatic Prompt Optimization Achieves State-of-the-Art Results

A novel technique for automatic prompt optimization, referred to as SAPO (Segment-level Automatic Prompt Optimization), has been detailed in a paper on arXiv. Unlike conventional methods that modify prompts as a whole—often enhancing one aspect while harming another—SAPO breaks prompts into four distinct segments: role, context, tasks, and output format. It implements focused enhancements derived from the top-5 and bottom-5 examples in the training dataset. Utilizing a single large language model (LLM), the optimization loop features static meta-prompts and organized outputs for segmentation, analysis of weaknesses, and generation of candidates. The study outlines a train/validation methodology and a two-phase generation process. SAPO was assessed using five benchmark datasets—SQuADv2, TweetEval, XSUM, CommonGen, and GSM8K—with GPT-3.5-Turbo and GPT-4o-mini as the foundational models. It outperformed zero-shot baselines and strong APO benchmarks, including APE, OPRO, EvoPrompt, GEPA, and StraGO. This paper can be found on arXiv with the identifier 2608.11219 in the Computer Science > Artificial Intelligence category.