AI models can now learn new skills without forgetting old ones — and route between them automatically

Researchers built a modular system that lets a single AI model learn tasks in different domains (math, medicine, chat) by stacking small, specialized add-ons that don't interfere with each other. In practice, this means one model can become competent across many domains without the usual tradeoff where learning something new breaks what it already knew.

Until now, training a language model on new domains meant either retraining the whole thing from scratch (expensive, slow) or accepting that the new learning would overwrite old capabilities. This system sidesteps both problems by keeping each domain's learning isolated in frozen layers that stack additively. The unexpected finding is that the model doesn't learn domain-specific knowledge at all — instead, it discovers transferable patterns like reasoning clarity and step-by-step logic that work across domains. That means you could train on chat data and the system would automatically route medical queries to the chat-trained stack because the underlying competence is portable.