# AI that learns from every conversation

> Researchers propose language models that update their weights in real time from user interactions instead of staying frozen after training.
- Title: AI that learns from every conversation · arXiv | Shortify
- Summary: Researchers propose language models that update their weights in real time from user interactions instead of staying frozen after training. A hypernetwork…
- Keywords: language models, infinite-parameter llms, hypernetwork, weight generation, live data adaptation, mixture-of-experts, bayesian inference, in-context learning, model scaling, jinli hu, ross clarke, yichuan zhang
- Source: arXiv — https://arxiv.org/abs/2609.18842
- Read time: 1 min
- Topics: language models, infinite-parameter llms, hypernetwork, weight generation, live data adaptation, mixture-of-experts, bayesian inference, in-context learning, model scaling, jinli hu, ross clarke, yichuan zhang
## The frozen model problem
Today's language models stop learning after training. User corrections and facts shared during conversations are lost because weights can't update, forcing the model to re-read context on every request.
## Scaling laws demand more
Language models grow more capable with more parameters and training data. Mixture-of-Experts architectures show this by activating only a fraction of a huge parameter bank per token, but they still rely on static pretraining data.
## Infinite-parameter llms proposed
Researchers introduce a compact hypernetwork that generates feed-forward weights from live data instead of reading them from static memory. This turns user interactions into low-rank modulations of a shared base network.
## Bayesian belief updates weights
The system maintains a Bayesian belief over the generator's latent state and updates it online as conversations progress. Effective weights are re-derived as the belief evolves, without the model's memory footprint growing.
## Three concrete advantages
Storing live data in weights instead of prompts saves compute, frees up context window space, and persists updates across conversation turns. The approach can also generalize better than keeping information only in context.
## Feasible weight space is infinite
Although the model's memory footprint stays constant, the space of weights it can generate becomes effectively infinite. This lets deployed models learn from live interactions without retraining or freezing their architecture.
## Evaluation tests in-context learning
The research applies this methodology to in-context learning and retrieval tasks. Results are preliminary, with the paper submitted to arXiv in September 2026 by Jinli Hu, Ross Clarke, Yichuan Zhang, and José Miguel Hernández-Lobato.
## Key takeaway

A hypernetwork can generate model weights dynamically from live data, letting AI systems learn and adapt during conversations without retraining.