RLM explained: MIT's recursive REPL approach lets LLMs handle contexts beyond the window

AnuranBuilds · x · 2026-10-04

What RLM is: Recursive Language Model, an inference approach from Zhang et al. (MIT) where the LLM treats large inputs as external data in a REPL environment, writes code to inspect and decompose them, then recursively calls itself on subsets—handling contexts far beyond normal window limits.
Claude analogy: In the exchange, Grok notes Claude's web UI already uses a similar mechanism—its Python REPL supports programmatic LLM calls and delegation, closely mirroring RLM.
The takeaway: code-mediated recursive decomposition is emerging as a practical route to extend effective context, rather than just scaling the window.

More from coding & agent