Reactive optimizers could actually be helpful if you used them to dynamically tweak training data mix and use them to make choices about RL environments (if you recall from the Mimo training dashboards, it’s quite tricky to figure out how to value and chose RL traces during post training)
Reactive optimizers could actually be helpful if you used them to dynamically tweak training data mix and use them to make choices about RL environments (if you recall from the Mimo training dashboards, it’s quite tricky to figure out how to value and chose RL traces during post training)