ESPO beats GEPA on prompt length and accuracy by stabilizing selection
On seven NLP benchmarks, ESPO improves average accuracy by 3.76 points over GEPA while producing prompts 47% shorter — and the ablation reveals that adding diversity without a stability gate actively hurts performance by 1.20%.
Why it earned a slot
ESPO (arXiv 2609.04197) was submitted September 3, 2026, and it attacks a problem every practitioner who has tuned prompts knows too well: evolutionary optimizers like GEPA keep appending rules, caveats, and constraints until the prompt is three times longer and no more accurate. The authors trace this bloat to three specific deficiencies — incomplete error observation, limited search diversity, and unreliable selection — and propose a three-phase pipeline. Diagnose clusters all training errors into structural patterns in a single pass. Propose generates candidates through four complementary strategies, each with independent biases. Select applies bootstrap stability selection to filter out noise before committing to a prompt. On seven public benchmarks — Tweet, MMLU, GSM8K, HotpotQA, ScoNe, HoVer, and PUPA — ESPO hits 74.67% average accuracy against GEPA's 70.91%, a 3.76-point gain, while cutting prompt length from 1,878 characters down to 1,004. Here is what I think actually matters: the ablation result is the real contribution. Adding diversity without bootstrap selection actively hurts performance by 1.20%. That is not a minor footnote. It is a direct refutation of the intuition that more exploration is always better. The paper is saying that unguided search is a liability, not an asset, and any prompt optimization pipeline that skips a stability gate is just burning compute on noise. The cross-model results back this up. ESPO yields the best average accuracy on every model tested — Gemma 3 12B, Mistral 14B, Qwen3 32B, and Claude Haiku 4.5 — with the largest gap on Qwen3 GSM8K jumping from 15.00% to 91.40%. That is not a marginal win on one architecture. I do not know whether these prompt-level optimizations survive deployment in production systems where the data distribution shifts daily. The paper's generalization bound lives in the appendix and grounds each phase in a corresponding term of the test-time gap, which is mathematically clean, but I would want to see this tested against real traffic, not held-out benchmark splits. The bigger open question is whether Diagnose-Propose-Select generalizes beyond prompts. If the core insight is that structured error clustering plus a stability filter beats unstructured search, that pattern could apply to code generation, tool calling, or any domain where an optimizer is appending noise to a template. But that is my speculation, not the paper's claim. They tested prompts on NLP benchmarks and that is the scope of what I can defend.