XNews

Boltzbit and Cambridge Introduce 'Infinite Parameter' LLM with Dynamic LoRA Weights

Boltzbit and Cambridge University researchers proposed a new method that writes knowledge into model weights instead of prompts, using a lightweight hypernetwork to dynamically compile user interaction data into low-rank (LoRA) weights. A Bayesian mechanism allows these weights to evolve during conversation, achieving higher accuracy in long-document and multi-turn scenarios while keeping computational cost constant per turn. The approach outperforms traditional methods in long-context tasks but not in short-text scenarios.

Source posts