The article introduces Open-Ended Optimization (OEO), a framework that allows a frontier model optimizer to dynamically compose its improvement process online, rather than relying on prescribed optimization pipelines. The authors compare OEO against two staged approaches, SkillOpt and GEPA, across 14 head-to-head comparisons over eight benchmark-target-model settings.
GPT-5.5-driven OEO achieved 12 wins, one tie, and one narrow loss of 0.21 percentage points. It utilized a median of 34.3 percent of SkillOpt's configured target-interaction token budget. A control test confirmed that gains were not merely due to a single prior-driven rewrite. However, the approach has capability boundaries: SkillOpt outperformed OEO with medium optimizers, and weak optimizers could not operate through the unchanged OEO interface.
The findings recast prescribed pipelines as capability-dependent scaffolding, suggesting that while essential constraints remain external, sufficiently capable optimizers can compose their own route from feedback to improvement.