XNews

Boltzbit and Cambridge Develop 'Infinite Parameters' LLM with Dynamic Weight Updates

Researchers from Boltzbit and Cambridge University introduced a method to dynamically generate LoRA weights from user interactions using a hypernetwork, allowing the model to 'write knowledge into weights' instead of relying solely on context. The approach outperforms traditional methods in long documents and multi-turn conversations, with constant compute cost per turn.

Source posts