ESPO cuts prompt bloat in prompt optimization

Why prompts keep inflating

Evolutionary prompt optimization methods like GEPA repeatedly append rules and exceptions to the prompt. Even after three or more rounds, prompt length can grow 3x while accuracy barely improves. The authors identify three structural causes: incomplete error observation, limited search diversity, and unreliable selection. In each round, GEPA samples only 3–8 random errors, so systematic failure patterns get missed. It then mutates prompts with a single operator, which tends to stack redundant rules instead of fixing root causes. Finally, it picks survivors by point estimation on a small validation set, where noise can win out over true quality.

ESPO’s three-phase framework

ESPO stands for Error-Structured Prompt Optimization, and it recasts optimization as structured statistical estimation. Phase 1, Diagnose, collects all training errors under the current prompt and clusters them into 3–7 structural patterns in one round, so the optimizer gets complete error coverage instead of a sparse random snapshot. Phase 2, Propose, generates candidates through four complementary strategies: diagnostic revision, consolidation, ablation, and factual injection. Each strategy carries a different inductive bias, so the candidate pool explores distinct remedy types rather than one mutation profile. Phase 3, Select, applies bootstrap stability selection with B=20 resamples. The candidate that wins the most resamples wins, with ties broken in favor of shorter prompts.

Quantitative comparison with GEPA

The authors benchmark on Tweet, MMLU, GSM8K, HotpotQA, ScoNe, HoVer, and PUPA using a deliberately weak default prompt and Claude Sonnet 4.5 as the student model. ESPO averages 74.67% accuracy versus GEPA’s 70.91%, a +3.76 pp gain. At the same time, the average prompt length is 1,004 characters, 47% shorter than GEPA’s 1,878 characters. On HotpotQA, GEPA balloons to 2,845 characters while ESPO keeps 1,008 characters and preserves a usable accuracy edge. The shorter prompt also translates into lower inference latency.

Generalization across models

To test whether the optimization transfers, the authors rerun on Gemma 3 12B, Mistral 14B, Qwen3 32B, and Claude Haiku 4.5. ESPO achieves the best average accuracy on every tested model. The largest jump appears on Qwen3 GSM8K, where the default prompt scores 15.00% and ESPO climbs to 91.40%, a +56.00 pp improvement over GEPA. Across the four models, ESPO prompts are 30–60% shorter while tracking toward higher accuracy.

Why this structure matters

Length-only constrained GEPA lifts accuracy from 70.91% to 71.00%, essentially flat. Simply capping length does not guarantee accuracy gains because the underlying selection problem remains noisy. ESPO improves accuracy and compresses prompts simultaneously because its Diagnose phase targets the right patterns, Propose builds focused candidates, and Select stabilizes choice under validation noise. The three components move together.

Practical use and limitations

ESPO’s bootstrap selection is well suited to small validation sets, but the experiments rely on a single strong LLM—Claude Sonnet 4.5—for both reflection and selection, using a 70/30/500 split. That setup keeps the design tight, though external validation on other reflection models is still limited. The framework offers a practical route to shorter, more accurate prompts, but users should still evaluate optimized prompts on bias-relevant slices before deployment, especially for sensitive applications.