Researchers unveil SAPO for smarter AI prompt tuning

A team of researchers from multiple institutions has developed SAPO (Segment-level Automatic Prompt Optimization), a novel approach to prompt engineering that moves beyond monolithic rewrites. SAPO decomposes prompts into four modular segments—role, context, tasks, and output format—and applies targeted improvements based on performance analysis. The method uses a single LLM with static meta-prompts and structured outputs to diagnose weaknesses, extract recommendations, and generate optimized candidates in a two-stage process.

In evaluations across five diverse benchmarks (SQuADv2, TweetEval, XSUM, CommonGen, GSM8K) using GPT-3.5-Turbo and GPT-4o-mini, SAPO achieved the highest average scores compared to Zero-shot baselines and strong APO competitors including APE, OPRO, EvoPrompt, GEPA, and StraGO. The paper was accepted at the IJCAI-ECAI 2026 RobustifAI workshop, highlighting its relevance to reliable and safe generative AI systems.