In this video, we explain how Recursive Language Models or RLMs work, look at actual trajectories on real problems, see how to implement it from scratch using Deno and Pyodide, and discuss what their key features and benefits are.
RLMs are an inference technique where a LLM interacts with arbitrarily long prompts through an external REPL. The LLM can write code to explore, decompose and transform the prompt. It can recursively invoke sub-agents to complete smaller subtasks too! Crucially, the subagent responses do not get automatically loaded into the parent agent's context, it gets returned as symbols or variables inside the parent's REPL.
---
Follow me on Twitter to get next video updates: https://x.com/neural_avb
To join our Patreon and support this channel financially, visit: https://www.patreon.com/NeuralBreakdownwithAVB
Members get access to everything behind-the-scenes that goes into producing my videos - including code. Plus, it supports the channel in a big way and helps to pay my bills.
---
You can read the RLM paper here on Paper Breakdown:
https://paperbreakdown.com/abs/2512.24601
Arxiv:
https://arxiv.org/abs/2512.24601
---
Paper Breakdown makes it way easier to discover Computer Science research, get personalized paper recommendations to study every day, and access a premium collection of tools to study interactively with context-aware AI agents. Get 50% off using code - RLM50
---
Git repo:
https://github.com/avbiswas/fast-rlm
Pypi:
https://pypi.org/project/fast-rlm/
—-
Timestamp:
0:00 - Intro
4:36 - What are RLMs
11:13 - RLM trajectories
29:30 - Implementation
45:00 - When to use RLMs and why
