Reactive Optimizers Let Agents Rewrite Training

This is a compelling step toward treating optimization as a closed-loop programming problem, especially for new architectures and datasets where established recipes are unreliable.

• –OPAL can modify per-parameter-group updates, learning-rate schedules, momentum, precision, clipping, and other optimizer behavior while training continues. [GitHub](https://github.com/xingjian-bai/opal)
• –The strongest use case is under-tuned training: results improve on canonical baselines, but a tuned static sweep still beats at least one OPAL run, so this is an adaptive fallback rather than a universal replacement. [Research details](https://xingjianbai.com/reactive-optimizers/)
• –Asynchronous operation limits reported wall-clock overhead to roughly 1.01×–1.14×, while blocking decisions can reach 1.30×–4.39×. [Evaluation summary](https://www.alphaxiv.org/abs/2610.33826)
• –Practical adoption is constrained by limited multi-node support, missing FSDP and DeepSpeed support, and reproducibility risks from agent-driven changes. [Repository](https://github.com/xingjian-bai/opal)