XNews

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

Boltzbit and Cambridge University researchers propose a method to write knowledge into model weights dynamically rather than stuffing it into the prompt. A lightweight hypernetwork compiles user interaction data into low-rank (LoRA) weights, and a Bayesian mechanism evolves these weights during conversation. This 'infinite parameters' approach outperforms traditional methods in long documents and multi-turn dialogues.

Source posts