Overview

RLM is a novel inference library from MIT's OASYS Lab that reframes how language models handle long contexts. Instead of stuffing enormous inputs into a single context window, RLM allows a language model to programmatically examine, decompose, and recursively call itself over its input — treating context as a variable in a REPL-style execution environment. The result is effective handling of near-infinite length inputs without the degradation typical of long-context approaches.

With 3,489 stars and 630 forks since its late 2025 release, RLM is gaining traction among researchers and practitioners exploring alternatives to brute-force context extension.

Key Features

• Recursive Self-Calls: The model can launch sub-calls over decomposed portions of its input, assembling results bottom-up — analogous to divide-and-conquer for language modeling
• Multiple Sandbox Backends: Supports local Python exec, Docker containers, and cloud-isolated environments including Modal, Prime Intellect, Daytona, and E2B for safe code execution
• Broad Model Provider Support: Works with OpenAI, Anthropic, OpenRouter, and Portkey APIs, plus local models via vLLM, making it provider-agnostic
• Trajectory Logging & Visualization: Built-in interactive dashboard for inspecting execution paths, sub-call traces, and intermediate results — critical for debugging recursive reasoning chains
• Simple PyPI Installation: Available as pip install rlms with minimal dependencies

Use Cases

RLM is particularly well-suited for processing large codebases where the model must understand cross-file dependencies, analyzing lengthy research papers or legal documents beyond any single context window, complex multi-step reasoning tasks that benefit from hierarchical decomposition, and research into advanced LM inference patterns such as chain-of-recursive-thought.

Technical Details

RLM implements the Recursive Language Model paradigm described in the accompanying arXiv paper (arXiv:2512.24601). The library abstracts away sandbox management, provider routing, and result assembly behind a clean Python API. Recursive depth and termination conditions are configurable. The REPL environment ensures that sub-calls are isolated, preventing context pollution between recursive levels.

Getting Started

pip install rlms

from rlm import RLM

rlm = RLM(
    backend="openai",
    backend_kwargs={"model_name": "gpt-4o"},
)
result = rlm.completion("Analyze this 100k-token codebase...").response
print(result)