Researchers pinpoint why larger language models pick up skills that small ones miss

In three linesA study comparing models from 4M to 4B parameters reveals small models fail at rare tasks because frequent ones constantly overwrite learned skills. A practical solution: increase target task frequency in training data rather than scaling up the model.

Summary generated by Claude — human-verified