Berkeley team packages LLM optimizer methods into GEPA, then beats them with omni
mrdrozdov · x · 2026-07-24
Researchers at Berkeley’s STAR Lab are packaging multiple LLM-based optimization algorithms into a single API via GEPA, and a new meta-optimizer, optimizeanything omni, combines them to exploit complementary strengths.
- The post says AI-based research and engineering is now a major topic in the STAR Lab and increasingly used in industry.
- GEPA bundles several “autoresearch” algorithms so users can mix and match them for tasks like prompt writing, agent design, and code optimization.
- In the quoted thread, the authors compare GEPA, AutoResearch, and Meta-Harness and note that each wins on different tasks.
- Their new optimizer, omni, leverages the best parts of each method.
- At a matched budget, omni beats every standalone optimizer in their tests.
Related event: Berkeley Team Releases 'omni', a Meta-Optimizer for LLMs(3 posts)→
More from coding & agent
- Using AI Subscription Limits to Tackle Unsolved Math Problems — nptacek · 2026-07-24
- Claude Code users discuss “skills as layers” with a practical reliability playbook — samgoodwin89 · 2026-07-24
- Claude Code workaround gist resurfaces after a messy 19-tool run — PawelHuryn · 2026-07-24
- OpenComputer launches a durable agent sessions API and asks for feedback — zeeg · 2026-07-24
- Lovable users are told to save the MVP plan as a root-level .md file — PrajwalTomar_ · 2026-07-24
- Claude Code Meme Recasts “Allow Once” as the Whole Job — WhatTheLJW · 2026-07-24