Infinite-Parameter LLMs generate feed-forward weights from live data instead of freezing them

arXiv·medium signalXBlueskyLinkedInCopy link

Submitted 16 September (arXiv 2609.18842) by Jinli Hu, Ross M. Clarke, Yichuan Zhang and José Miguel Hernández-Lobato, the paper proposes replacing a fixed parameter bank with a compact hypernetwork that converts live session data into weight modulations, borrowing Mixture-of-Experts structure. A Bayesian belief over the generator's latent code updates online, so the effective weight is re-derived through a session rather than fixed after the first pass, keeping storage constant while allowing unbounded weight variation. The claimed advantages over in-context learning are amortized compute, freed context window, and knowledge that persists across turns. The listing describes an evaluation protocol against in-context learning and retrieval but the abstract page carries no specific benchmark numbers, so the empirical case is unverified from the abstract alone.