LittleLearner shows pretraining filters set absolute knowledge boundaries

Training 0.6B to 5B models on the 88B-token LittleCurriculum—a corpus strictly limited to U.S. K-5 standards—demonstrates that post-training techniques like GRPO cannot elicit knowledge outside the pretraining distribution. The study confirms that scaling and instruction tuning amplify curriculum-specific skills but fail to provide out-of-scope knowledge acquisition.