Recursive Language Models (RLM)

Status	Active
Path	AI/ML
Prereqs	Instructor
Repo	pysprings/rlm

Overview

An implementation of Recursive Language Models in 77 lines of Python, based on the 2025 MIT paper by Zhang, Kraska, and Khattab. The key insight: instead of feeding entire documents into an LLM's context window (where quality degrades with length), store the document as a Python variable and let the LLM write code to explore it.

How It Works

User Query → Main LLM → Writes Python Code → Executes in REPL
                              ↓
                      Calls Sub-LLM on Chunks
                              ↓
                      Reads Print Output
                              ↓
                      Iterates Until Done

Architecture (8 Chapters)

The project uses literate programming — the presentation slides are the source of truth, and the Python code is automatically extracted from them using Entangled.

1. Title & Problem — Context rot: LLMs degrade on long documents
2. Build Steps 1–6 — REPL namespace, sub-LLM, conversation loop, FINAL signaling, print capture, full assembly
3. Implementation — The complete 77-line function
4. Chatbot — Interactive variant
5. Demo — Live examples with Project Gutenberg books

Key Concepts

• Literate programming (markdown → tangled Python)
• Python exec() and namespace manipulation
• LLM orchestration (main model + sub-model)
• Exception-based control flow (FINAL() signal)
• REPL environment simulation

Try It

1. Clone: git clone https://github.com/pysprings/rlm
2. cd rlm && pip install -e .
3. Browse the slides: mkdocs serve → http://localhost:8000
4. Run the demo: python code/demo.py

Where to Go Next

• marketing-bot — Another architecture for LLM-driven systems
• Entangled — Deep dive into the literate programming tool used here