SEPO edits typed pieces, not whole prompt strings

SEPO edits typed pieces of a prompt instead of rewriting the whole string, and it outperforms GEPA across a 14-task held-out suite while spending less to get there. For teams running optimization loops in production, the tokens spent tuning a prompt are pure overhead. A cheaper optimizer changes whether that overhead is worth paying.

SEPO splits a prompt into stable typed units inside a two-layer schema. Each edit links to the examples it newly fixed or broke, carrying that record forward. A team can revert one unit without rerunning the whole search.

On the held-out suite, the margin was 3.1 points on Llama-3.1-8B-Instruct and 2.2 on Qwen3-8B. SEPO also spent 2.9 million optimization tokens against GEPA's 4.1 million, and its prompts came out about five times shorter.

The paper doesn't say how SEPO handles a prompt that resists decomposition, like one dense instruction block with no natural seams. That gap will decide whether the approach generalizes beyond benchmark tasks or stays limited to prompts with clean structure.