Boltzbit and Cambridge Propose 'Infinite Parameters' LLM Method
Boltzbit and Cambridge University researchers proposed a new method for LLMs that writes knowledge directly into model weights instead of the prompt context window. A lightweight hypernetwork dynamically compiles user-provided real-time data into low-rank LoRA weights, which evolve via a Bayesian mechanism during conversation. This 'infinite parameters' approach outperforms traditional context-based methods in long documents, multi-distractors, and multi-turn conversations.