Recursive Language Models — Interactive Demo

input.py

Query

Context (Large Document)

The history of communication technology spans thousands of years. Ancient civilizations used smoke signals and drums. The telegraph, invented by Samuel Morse in 1837, revolutionized long-distance communication. This paved the way for further innovations.

Alexander Graham Bell, a Scottish-born inventor and scientist, is credited with inventing the first practical telephone. On March 7, 1876, Bell was granted U.S. Patent No. 174,465 for his invention, which he called the "electrical speech machine." The first words spoken over the telephone were "Mr. Watson, come here, I want to see you," spoken by Bell to his assistant Thomas Watson.

Meanwhile, in Italy, Antonio Meucci had been working on voice communication devices since 1849. He filed a patent caveat in 1871 but couldn't afford to renew it. The controversy over who truly invented the telephone continues to this day.

Radio technology emerged in the late 1800s, with Guglielmo Marconi demonstrating wireless telegraphy in 1895. The 20th century saw the rise of television, the internet, and mobile phones, each building on the innovations of the past.

Tokens: ~350 Chars: 1,247

rlm_execution.log

🔄

Click "Run RLM Query" to see
recursive decomposition in action

How Recursive Language Models Work

Context as Variable

Instead of flooding the LLM with the entire context, RLMs store it as a Python variable that can be programmatically accessed and manipulated.

REPL Environment

The root LLM operates in a Python REPL, writing code cells to peek at, grep, and partition the context as needed.

Recursive Calls

When analysis is needed, the root LLM spawns child RLM calls with smaller context chunks, avoiding "context rot" degradation.

Final Answer

Results bubble up through the recursion tree. The root LLM aggregates findings and outputs FINAL(answer) when complete.