MIT 刚开源了一个让大模型一口气读完千万 token 的推理库——RLM

来自 MIT CSAIL 的 OASYS lab，背后有 DSPy 和 ColBERT 的作者 Omar Khattab 参与，VentureBeat 都专门报道过，刚开源没多久就冲到了 5k star。

普通模型处理超长文本有个死穴：

一本书、一整个代码库、几百页报告，要么塞不进上下文窗口，要么读到后面忘前面，最后给你一个糊了的答案。

RLM 的思路完全不一样。

它不靠把窗口撑大，而是让模型把整份上下文当成手边的变量，自己写代码去分块、grep、递归调子模型去查，就像程序员在 REPL 里调试，而不是一次把所有代码背进脑子。

论文数据显示，它能处理超出模型原生上下文窗口一两个数量级的输入，就算在短上下文场景，质量也明显优于普通前沿模型。

接入也简单，把代码里的 llm.completion() 换成 rlm.completion()，pip install rlms 一下就行。

上下文窗口的尽头，从来不是模型的极限，而是工具的边界。

alexzhang13/rlm

Source: https://github.com/alexzhang13/rlm

Recursive Language Models (RLMs)

Full Paper • Blogpost • Documentation • RLM Minimal

Overview

Recursive Language Models (RLMs) are a task-agnostic inference paradigm for language models (LMs) to handle near-infinite length contexts by enabling the LM to programmatically examine, decompose, and recursively call itself over its input. RLMs replace the canonical llm.completion(prompt, model) call with a rlm.completion(prompt, model) call, acting as a “language model”. RLMs offload the context as a variable in a REPL environment that the LM can interact with and launch sub-LM calls inside of.

RLMs are a bet on future “language model” design choices. We argue for a CodeAct-style harness (i.e. all language models should have access to a code environment) with sub-(R)LM calls as functions in code, and context / prompts as objects in code. RLMs explicitly defer code execution with sub-calls as functions to the language model itself, which is incredibly flexible and lends itself well to scale if trained correctly. We want to move away from the JSON tool-calling standard for both sub-agents and generic tool calls. The naming comes from the fact that such a system is itself a “language model” (a probabilistic mapping from text to text) that builds around and relies on recursive sub-LLM calls.

This repository provides both an extensible inference engine and training environment for using RLMs around standard API-based and local LLMs. The initial experiments and idea were proposed in a blogpost in 2025, with expanded results in an arXiv preprint.

We now also include a verifiers training environment based on Prime Intellect’s prime-rl in the training/ folder. Train your own RLMs, which directly can be plugged into our inference engine!

This repository contains inference code for RLMs with support for various sandbox environments. Open-source contributions are welcome. This repository is maintained by the authors of the paper from the MIT OASYS lab.

Quick Setup

rlms requires Python 3.11 or later.

You can try out RLMs quickly by installing from PyPi:

pip install rlms

The default RLM client uses a REPL environment that runs on the host process through Python exec calls. It uses the same virtual environment as the host process (i.e. it will have access to the same dependencies), but with some limitations in its available global modules. As an example, we can call RLM completions using GPT-5-nano:

from rlm import RLM

rlm = RLM(
    backend="openai",
    backend_kwargs={"model_name": "gpt-5-nano"},
    verbose=True,  # For printing to console with rich, disabled by default.
)

print(rlm.completion("Print me the first 100 powers of two, each on a newline.").response)

curl -LsSf https://astral.sh/uv/install.sh | sh
uv init && uv venv --python 3.12  # change version as needed
uv pip install -e .

make quickstart

REPL Environments

We support two types of REPL environments – isolated, and non-isolated. Non-isolated environments (default) run code execution on the same machine as the RLM (e.g. through exec), which is pretty reasonable for some local low-risk tasks, like simple benchmarking, but can be problematic if the prompts or tool calls can interact with malicious users. Fully isolated environments use cloud-based sandboxes (e.g. Prime Sandboxes, Modal Sandboxes) to run code generated by the RLM, ensuring complete isolation from the host process. Environments can be added, but we natively support the following: local (default), ipython, docker, modal, prime, daytona, e2b.

rlm = RLM(
    environment="...", # "local", "ipython", "docker", "modal", "prime", "daytona", "e2b"
    environment_kwargs={...},
)

Local Environments

The default local environment LocalREPL runs in the same process as the RLM itself, with specified global and local namespaces for minimal security. Using this REPL is generally safe, but should not be used for production settings. It also shares the same virtual environment (e.g. Conda or uv) as the host process.

IPython (requires pip install 'rlms[ipython]')

IPythonREPL runs cells inside a real IPython session — either in-process (default) or in a separate ipykernel subprocess. Subprocess mode adds hard cell_timeout enforcement and full namespace isolation from the RLM host. See the IPythonREPL docs for details.

Docker (requires Docker installed)

We also support a Docker-based environment called DockerREPL that launches the REPL environment as a Docker image. By default, we use the python:3.11-slim image, but the user can specify custom images as well.

Isolated Environments

We support several different REPL environments that run on separate, cloud-based machines. Whenever a recursive sub-call is made in these instances, it is requested from the host process.

Modal Sandboxes

To use Modal Sandboxes as the REPL environment, you need to install and authenticate your Modal account.

uv add modal  # add modal library
modal setup   # authenticate account

Prime Intellect Sandboxes

Prime Intellect Sandboxes are currently a beta feature. See the documentation for more information. We noticed slow runtimes when using these sandboxes, which is currently an open issue.

To use Prime Sandboxes, install the SDK and set your API key:

uv pip install -e ".[prime]"
export PRIME_API_KEY=...

Model Providers

We currently support most major clients (OpenAI, Anthropic), as well as the router platforms (OpenRouter, Portkey). For local models, we recommend using vLLM (which interfaces with the OpenAI client). To view or add support for more clients, start by looking at rlm/clients/.

Training

We provide a simple RL training harness for training RLMs used in this repo (specifically the local REPL). The implementation uses no sandboxes for simplicity and slots easily your use case, but an ideal setup would use sandboxes for safety. Training logic is isolated to the training/ folder, which exposes rlm.RLM as a verifiers Environment and plugs straight into prime-rl. See the training README for the launch command. The harness uses subprocess-isolated local REPL execution (no cloud sandboxes), matching the local environment above.

A worked example with an example .toml lives in training/environments/oolong/ (OOLONG long-context QA). New training environments can be added the same way — author a verifiers env that wraps your task (see the verifiers docs), then reference it from a config.

Relevant Reading

• [Dec ’25] Recursive Language Models arXiv
• [Oct ’25] Recursive Language Models Blogpost

If you use this code or repository in your research, please cite:

@misc{zhang2026recursivelanguagemodels,
      title={Recursive Language Models},
      author={Alex L. Zhang and Tim Kraska and Omar Khattab},
      year={2026},
      eprint={2512.24601},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2512.24601},
}

RLMs in the Wild

There are many amazing demos and production-ready use cases of RLMs. We provide a list of notable examples that explicitly use RLMs as a central piece of their design.

• DSPy.RLM
• Ax
• context-labs/HALO: RLM-based Automatic Agent Optimization Loop
• viplismism/rlm-cli: CLI for Recursive Language Models
• alphaXiv Official Blog. Reinforcing Recursive Language Models
• Daytona. Building Deep Recursive Language Models
• Symbolica. SotA ARC-AGI-2 Results with REPL Agents
• Google Cloud Community Articles. RLMs in ADK
• Prime Intellect Blog. Recursive Language Models: the paradigm of 2026

Optional: Trajectory metadata, logging, and debugging

RLMChatCompletion has an optional metadata field (default None) that holds the full trajectory (run config + all iterations and sub-calls) so you can reconstruct the run. Pass an RLMLogger to capture it:

• In-memory only (trajectory on completion.metadata): logger=RLMLogger() (no log_dir).
• Also save to disk (JSONL for the visualizer): logger=RLMLogger(log_dir="./logs").

Visualizing logs. We also provide a simple visualizer to inspect code, sub-LM, and root-LM calls. Use RLMLogger(log_dir="./logs") so each completion writes a .jsonl file:

from rlm.logger import RLMLogger
from rlm import RLM

logger = RLMLogger(log_dir="./logs")
rlm = RLM(..., logger=logger)

To run the visualizer locally, we use Node.js and shadcn/ui:

cd visualizer/
npm run dev        # default localhost:3001