A Recursive Language Model (RLM) is an inference scheme in which a long prompt is held as a variable in a program execution environment rather than placed entirely in the model's context; the language model writes code to slice and read the parts it needs and calls itself recursively on sub-problems. It was proposed by MIT researchers in 2025.
It is studied as a way to handle inputs far longer than the context window, and because the model decomposes the input programmatically, even small models can process long inputs.