NPO matches complex GEPA in prompt optimization with single-lineage simplicity
omarsar0 · x · 2026-08-31
A new paper introduces Nested Prompt Optimization (NPO), which maintains only a single lineage instead of complex candidate pools or search trees, matching the performance of the more complex GEPA method on a smaller rollout budget.
Core Mechanism:
- At each iteration, NPO runs the student on the current prompt, collecting traces and rewards.
- A sliding window of recent iterations is passed to a teacher model to rewrite the prompt.
Results:
- On two instruction-following benchmarks, NPO uses 3,500 and 6,800 rollouts respectively, performing comparably to GEPA (3,593 and 6,871).
- Performance remains competitive across 22 TextArena games.
- Key Insight: NPO's advantage grows with the strength of the teacher model, suggesting that optimizer-side search complexity has been overemphasized.
More from Research
- Hamel Husain launches updated AI evals course for engineers — HamelHusain · 2026-08-31
- New Paper Proposes "AI 45° Law" for Safe and Capable AGI — CFGeek · 2026-08-31
- Paper by 40 Top Researchers: CoT Monitoring is a Fragile but Promising AI Safety Opportunity — peterjliu · 2026-08-31
- AI can automate robustness checks on papers to detect selective reporting — paulnovosad · 2026-08-31
- Mathathon enforces transparency: logs, compute data, and open review — _sathvikr · 2026-08-31
- Claude Builds Quantum Computer Controller, Fixes Faults in 10 Seconds — imjustnewatai · 2026-08-31