Recursive Language Models

Summary

Replaces llm.completion with rlm.completion: the prompt lives in a CodeAct-style REPL that the model can inspect, decompose, and recurse into via sub-(R)LM calls. Paper arXiv 2512.24601: process inputs up to two orders of magnitude beyond the base context window. pip install rlms. Sandboxes: local, ipython, docker, modal, prime, daytona, e2b. Training via verifiers + prime-rl (OOLONG example). Maintained by the paper authors at MIT OASYS. MIT license.

Keywords

rlm · recursive-language-models · long-context · codeact · mit · oasys · repl

Topics

long-context inference, recursive LM calls, agent harnesses

Research notes

• Primary: GitHub README + API (MIT, Python, 5467 stars / 885 forks at check). Paper arXiv 2512.24601 (Zhang/Kraska/Khattab). Docs https://alexzhang13.github.io/rlm/. Discord posted the repo. Inference/training library, not a new hosted corpus, so no datasets_local row.