Already 9 months old when the scout caught it.
Recursive language models
What it is
Recursive Language Models (RLMs) are an inference-time strategy for handling prompts far longer than a model's context window. Instead of stuffing a long document into the context, the system loads it as a variable in a code-execution environment (a REPL) and has the model write code to inspect, slice, and search that variable — calling itself, or a fresh child instance of itself, on whatever sub-pieces it decides it needs, then assembling the results. No special training is required; it's a scaffold around an existing model.
Why it's emerging now
Context windows keep growing on paper, but quality quietly degrades on long, information-dense inputs well before the stated limit is reached. At the same time, giving models a code-execution loop (the same capability behind tool use and computer use) has become routine enough that treating the prompt itself as data the model can programmatically query, rather than tokens it must passively absorb, became a practical option. MIT's original paper (Zhang, Kraska, Khattab) showed RLMs handling inputs two orders of magnitude beyond normal context limits and beating long-context scaffolds on dense multi-document tasks; since then the pattern has been picked up independently in production tooling, including Google's Agent Development Kit, as a way to scale agents past hard context ceilings without retraining.
How it relates to what you know
It's a natural extension of tool use / function calling and agent memory: where harness-managed virtual memory pages context in and out of the window for the model, RLMs let the model itself decide what to pull in, via recursive self-calls over an externalized context store. It's also a cheaper, untrained alternative to the kind of latent reasoning loops seen in opaque recurrence — recursion happens at the scaffold level, in visible code, rather than inside the model's own forward pass.