It’s Tokens All The Way Down: How RLMs are Different — Kevin Madura, AlixPartners

A recursive language model can work through a large input by writing code to inspect it and asking additional model calls to analyze selected portions. The input lives in an external programming environment, so the main model does not have to receive the entire document in every step. A research paper on the approach describes recursive language models as treating long prompts as part of an external environment that the model examines, decomposes and recursively calls itself over.

The talk applies that idea to tasks such as analyzing tables and consolidating invoices. Code can perform exact filtering or arithmetic, while model calls handle interpretation. The useful distinction is where the work happens: the agent manipulates stored input and intermediate results, bringing selected findings back into its reasoning.