ESPO: Error-Structured Prompt Optimization via Diagnose, Diversify, and Stabilize
ESPO (Error-Structured Prompt Optimization) is proposed to address prompt bloat in evolutionary prompt optimizers like GEPA. It decomposes prompt optimization into three phases: Diagnose clusters training errors into structural patterns; Propose generates candidates via four complementary strategies; Select applies bootstrap stability selection. On seven public NLP benchmarks (Tweet, MMLU, GSM8K, HotpotQA, ScoNe, HoVer, PUPA), ESPO improves average accuracy by +3.76 percentage points over GEPA (74.67% vs 70.91%), matching or exceeding GEPA on every dataset while producing prompts 47% shorter (1,004 vs 1,878 characters) and faster at inference. Cross-model experiments on four additional student models (Gemma 3 12B, Mistral 14B, Qwen3 32B, Claude Haiku 4.5) show ESPO yields the best average accuracy on every model tested, with the largest gap on Qwen3 GSM8K (15.00% to 91.40%).
ESPO is a new prompt optimization method that improves accuracy and reduces prompt length compared to existing evolutionary optimizers. It uses error clustering, diverse candidate generation, and stable selection to produce shorter, more effective prompts across multiple NLP benchmarks and student models.
ESPO's three-phase design directly targets the root causes of prompt bloat: incomplete error observation, limited search diversity, and unreliable selection. By clustering all training errors in one round and using four complementary generation strategies with independent biases, it achieves better coverage of error patterns. Bootstrap stability selection likely reduces overfitting to noisy evaluation signals, leading to more robust prompts. The reported 47% reduction in prompt length suggests improved token efficiency, which could lower inference costs and latency.
Prompt optimization remains a practical lever for improving LLM performance without retraining. ESPO's ability to produce shorter prompts with higher accuracy could reduce API costs and improve latency for applications relying on prompt engineering. The cross-model results indicate broad applicability across different model families, making it relevant for enterprises and developers seeking model-agnostic optimization.
ESPO offers potential cost savings and performance gains for organizations using LLMs with hand-crafted or optimized prompts. Shorter prompts reduce token usage, lowering per-request costs, while higher accuracy can improve task success rates. The method's model-agnostic nature makes it valuable for multi-model deployments.
Next observable signals include: publication of the full paper with generalization bounds; release of code or implementation details; independent replication on additional benchmarks; adoption in prompt engineering tools or LLM orchestration frameworks; and comparisons with other prompt optimization methods like OPRO or APE.