A new research paper introduces a method for large language models to update their internal parameters during live use, addressing a longstanding limitation in how these systems operate. Current models rely on static weights trained on fixed datasets, preventing them from incorporating information encountered after deployment. The proposed approach, called Infinite-Parameter LLM, enables adaptation by generating dynamic weights from real-time user interactions rather than storing them permanently.

The architecture uses a compact hypernetwork to convert incoming data into low-rank adjustments of a shared base model. Unlike traditional systems that discard contextual information after processing, this method maintains a Bayesian belief state over latent parameters, continuously refining the effective weights throughout a conversation. This allows the model to retain learned behaviors across multiple interactions without expanding its context window or depending solely on prompt engineering.

The technique draws inspiration from Mixture-of-Experts architectures but diverges by focusing on weight modulation rather than token routing. According to the authors, this could improve efficiency and generalization compared to retrieving information through prompts or external databases. Evaluation metrics in the study compare the approach directly to in-context learning and retrieval-based methods, measuring performance across adaptive tasks.

The paper, authored by Jinli Hu, was submitted to arXiv in September 2026 and categorized under artificial intelligence and machine learning. It presents the framework as a technical exploration rather than a deployed solution, emphasizing its potential for future research directions in adaptive AI systems.