We think weโve found a better path to continual post-training: merging models in input space.
๐น A 256-token soft prompt (~0.01% of params) matches full-parameter RL
๐น Plug-and-play composition: prompts trained separately just stack, with no joint training, and the combo beats the experts. No more MOPD!
๐น Better OOD generalization, almost no forgetting
๐น ~0 inference overhead, even with a different prompt per request. Built for personalization.