Boltzbit and Cambridge Propose 'Infinite Parameter' LLM with Dynamic LoRA Weights
Boltzbit and Cambridge University researchers developed a method to write knowledge directly into model weights via a lightweight hypernetwork that dynamically compiles user interaction data into low-rank LoRA weights, instead of requiring LLMs to re-read long contexts. This approach outperforms traditional prompt-based reading in long-document and multi-turn dialogue scenarios while keeping per-turn computational cost constant.