Researchers Test What an LLM Learns When Capped at a Fifth-Grade Curriculum
A new project called Little Learner explores what happens when a language model is trained exclusively on material that wouldn't go beyond a fifth-grade reading and knowledge level. Instead of the usual approach of feeding models massive scrapes of the internet, books, and advanced technical text, this experiment restricts the training set to simpler, curated content roughly matching what a ten-year-old might encounter.
The idea is to probe how much of an LLM's apparent intelligence comes from exposure to advanced, specialized text versus more basic building blocks of language and reasoning. It's a small, independent effort rather than a paper from a major lab, but it taps into a broader curiosity about data curation, curriculum learning, and how model capability scales with the complexity of training material.
The project's site walks through the setup and early observations, though it's still an exploratory, low-scale experiment rather than a benchmark-driven study.