This paper investigates the surprising effectiveness of easy training data in enabling models to perform well on hard tasks. The authors find that current pretrained language models (LMs) often generalize well from easy to hard data, often performing as well as oracle models fine-tuned on hard data. They demonstrate this using seven different measures of datapoint hardness, including six human-based measures and one model-based measure. The study shows that even if one cares most about model performance on hard data, it can be better to collect easy data rather than hard data for fine-tuning, since hard data is generally noisier and costlier to collect. The experiments use open models up to 70b in size and four publicly available question-answering datasets with questions ranging in difficulty from 3rd grade science questions to college level STEM questions and general-knowledge trivia. The results show that easy-to-hard generalization in LMs is surprisingly strong for the tasks studied. The paper also explores the cost-benefit tradeoffs of collecting easy vs. hard training data and finds that easy data can outperform hard data when more easy data can be collected within a budget or when easy data is less noisy. The study concludes that current LMs generalize relatively well to test data across human difficulty levels even when finetuned on data that is measurably easier than the test data. The authors hypothesize that this occurs because easy data elicits latent knowledge and skills from pretrained models in a hardness-invariant way. The results are robust across model family and scale, six different human hardness measures and a model-based measure, four datasets/tasks, and several fine-tuning methods. The findings suggest that the scalable oversight problem may be easier than previously thought.This paper investigates the surprising effectiveness of easy training data in enabling models to perform well on hard tasks. The authors find that current pretrained language models (LMs) often generalize well from easy to hard data, often performing as well as oracle models fine-tuned on hard data. They demonstrate this using seven different measures of datapoint hardness, including six human-based measures and one model-based measure. The study shows that even if one cares most about model performance on hard data, it can be better to collect easy data rather than hard data for fine-tuning, since hard data is generally noisier and costlier to collect. The experiments use open models up to 70b in size and four publicly available question-answering datasets with questions ranging in difficulty from 3rd grade science questions to college level STEM questions and general-knowledge trivia. The results show that easy-to-hard generalization in LMs is surprisingly strong for the tasks studied. The paper also explores the cost-benefit tradeoffs of collecting easy vs. hard training data and finds that easy data can outperform hard data when more easy data can be collected within a budget or when easy data is less noisy. The study concludes that current LMs generalize relatively well to test data across human difficulty levels even when finetuned on data that is measurably easier than the test data. The authors hypothesize that this occurs because easy data elicits latent knowledge and skills from pretrained models in a hardness-invariant way. The results are robust across model family and scale, six different human hardness measures and a model-based measure, four datasets/tasks, and several fine-tuning methods. The findings suggest that the scalable oversight problem may be easier than previously thought.