Original title: ESPO: Error-Structured Prompt Optimization via Diagnose, Diversify, and Stabilize
Lihao Liu, Peng Tang, Kunwar Yashraj Singh, Shabnam Ghadar · September 2026
Writing instructions, or prompts, for artificial intelligence models can be tricky. Automatic prompt-writing tools often create prompt bloatWhen prompts become longer with added rules but do not improve accuracy.. With every round of editing, they tack on extra rules and warnings. This makes instructions up to 3× longer without improving accuracy. Researchers created a new system called ESPO to fix this across three steps. First, it diagnoses errors by grouping all training mistakes into structural patterns in one round. Second, it proposes candidate prompts using four different strategies. Third, it selects the best prompt using a technique called bootstrap stability selectionA technique that repeatedly tests candidates to choose the most consistent ones.. The researchers tested ESPO on seven public language benchmarks. ESPO produced prompts that were 47% shorter, faster to run, and more accurate than previous methods. It matched or beat the leading tool on every tested dataset.
Imagine cooking a complex dish from a recipe that has grown with sticky notes and corrections over time. The recipe becomes cluttered, hard to follow, and some notes might contradict each other, leading to inconsistent results. In this paper, the recipe is like a prompt, and the sticky notes are added rules that cause bloat. ESPO cleans up the recipe by first grouping common mistakes, then suggesting new simplified versions, and finally testing which versions work best.
For developers and users of AI systems, shorter prompts reduce computational costs and speed up responses. Higher accuracy means more reliable outputs in tasks like question answering and reasoning, making AI tools more trustworthy.
The abstract does not explicitly state limitations, but the work is tested on seven specific NLP benchmarks and four additional models, so it may not generalize to all domains or model types. The improvements are measured on average, and performance might vary across different tasks.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
ESPO: Error-Structured Prompt Optimization via Diagnose, Diversify, and Stabilize — Lihao Liu, Peng Tang, Kunwar Yashraj Singh et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.