# RLMs

> **TL;DR.** RLMs (Recursive Language Models) are an inference strategy where an LLM treats its prompt as an object inside a Python REPL, then recursively calls sub-LLMs over chunks of it instead of stuffing everything into one forward pass.

- **Category:** AI / Inference / Context Management
- **Stage:** established
- **Age:** 244 days
- **Origin date:** 2025-10-15
- **First detected:** 2026-04-21
- **Canonical URL:** https://earlyterms.com/term/rlms
- **Sources:** 7 primary URLs

## Definition

RLMs (Recursive Language Models) are an inference strategy where an LLM treats its prompt as an object inside a Python REPL, then recursively calls sub-LLMs over chunks of it instead of stuffing everything into one forward pass. The root model sees only the query and decides how to decompose the context.

The term was coined by MIT's [Alex L. Zhang](https://alexzhang13.github.io/blog/2025/rlm/) in an October 2025 blog post, formalized in the [December 2025 arXiv paper](https://arxiv.org/abs/2512.24601) with Tim Kraska and Omar Khattab, and declared "the paradigm of 2026" by Prime Intellect on January 1. The tagline "RLMs are the new reasoning models" — framing the 2026 shift the way 2025 shifted from LLMs to reasoning models — drove the April surge.

## Example

An RLM run over a 10M-token codebase doesn't try to fit it in context. It spawns a Python REPL where the full prompt lives as a variable, then writes grep and partition calls, launches sub-LLMs over each chunk, and only returns the distilled answer. RLM(GPT-5-mini) beats vanilla GPT-5 by 33% on OOLONG at 132k tokens for the same API cost.

## Analogy

If a reasoning model is a student writing out longhand, an RLM is the same student who first opens a filing cabinet and indexes it.

## Why it's emerging now

On April 20, 2026 raw.works published "RLMs are the new reasoning models," compressing Alex Zhang's six-month-old MIT thesis into a shareable 2026 tagline. Combined with a 3.5k-star reference library and Drew Breunig's "context rot becomes a coding problem" framing, the term is in the window between technical credibility and SEO crowding.

## Related terms

- *related:* context-rot
- *related:* context-engineering
- *related:* context-window
- *related:* agent-harness
- *related:* agent-loop
- *related:* grpo
- *competitor:* retrieval augmented generation
- *parent:* reasoning models
- *related:* chain-of-thought
- *related:* claude-opus-4-7

## Sources

1. [Alex Zhang — Recursive Language Models (origin post)](https://alexzhang13.github.io/blog/2025/rlm/)
2. [arXiv 2512.24601 — Recursive Language Models (Zhang, Kraska, Khattab)](https://arxiv.org/abs/2512.24601)
3. [Prime Intellect — the paradigm of 2026](https://www.primeintellect.ai/blog/rlm)
4. [Drew Breunig — The Potential of RLMs](https://www.dbreunig.com/2026/02/09/the-potential-of-rlms.html)
5. [raw.works — RLMs are the new reasoning models](https://raw.works/rlms-are-the-new-reasoning-models/)
6. [alexzhang13/rlm — reference inference library](https://github.com/alexzhang13/rlm)
7. [Hacker News — Recursive Language Models discussion](https://news.ycombinator.com/item?id=44984018)

---
_Generated by EarlyTerms · https://earlyterms.com/term/rlms_