Agent rewrites optimizer source code mid-training to build a reactive optimizer
giffmana · x · 2026-10-22
Researchers let an agent watch training dynamics and rewrite the optimizer's source code as training proceeds, turning optimization from a static procedure into a reactive one.
Reported results:
- Across every language and image setting tried without a tuned recipe, it matches or beats the best one-shot human recipe
- It self-guards against loss spikes and overfitting
- The authors argue optimization decisions may soon belong to agents, not humans
Paper and code are available to try on your own tasks.
More from coding & agent
- Richard Socher's Recursive launches Flint, a local coding model hitting 125+ tok/s on MacBook — aryaman2020 · 2026-10-22
- Claude Duet launches in research preview with Link wallet and Stripe MCP for agent payments — emilygsands · 2026-10-22
- Anthropic launches Claude Duet always-on agent; reviewer shifts to 50/50 from ChatGPT — every · 2026-10-22
- Datadog team has been generating TLA+ via skills, eyeing a meta-skill for full suites — tianyin_xu · 2026-10-22
- Agent planned a whole trip solo: perfect logistics, zero vibe — tech__unicorn · 2026-10-22
- "AI swarms that patch the world's vulnerable software for free" — who's building this? — voooooogel · 2026-10-22