Prompt optimizer GEPA lifts Meta Muse Spark 1.1 success from 22.2% to 100% while cutting queries to 0.3%
iamrobotbear · x · 2026-09-09
dvnxmvlhdf5 reports that GEPA, the prompt optimizer from the DSPy team, improved Meta's Muse Spark 1.1 task success rate from 22.2% to 100% while query share fell from 51.1% to 0.3%. The learned prompts encoded predictive semantics: public utility ranks predicted hidden hazards, replacing direct acquisition.
More from coding & agent
- OpenAI's Codex lead praises Ant's new coding agent for matching Codex background computer use — soumitrashukla9 · 2026-09-09
- Ora launches ax: see how agents experience any product with one npx command — EdenEmarco177 · 2026-09-09
- Teknium clarifies: Tinyfish search and scrape are free, cloud browser is not — Teknium · 2026-09-09
- Hermes Desktop OAuth Refresh Bug Forces Reauth, User Urges Fix — Teknium · 2026-09-09
- AWS gives students at 132 universities in 18 countries a free year of Kiro coding agent — pdamodaran · 2026-09-09
- Astra's weird hack: turns unresolved corner cases into 'known limitations' and never fixes them — banteg · 2026-09-09