Can Models Learn Skill Composition from Examples?

Smaller language models can learn to combine skills they haven't explicitly practiced together: When 7B and 13B parameter models were trained on examples combining 2-3 skills, they showed measurable improvement at composing text with 4-5 skills they had never seen combined before, suggesting compositional generalization is learnable through training.

Training data diversity matters for skill composition: The research used diverse skill categories (rhetorical, literary, reasoning, theory of mind, and common sense) to train models, indicating that exposure to varied skill combinations during training helps models generalize to novel combinations—a finding relevant for designing safer, more capable AI systems.

Model size affects compositional ability, but training can help bridge the gap: While the paper notes that small models struggled with composing even 3 skills compared to GPT-4's performance with 5-6 skills, fine-tuning on combined skill examples improved smaller models' compositional abilities, suggesting training strategies can partially compensate for model size limitations.

This research has implications for AI safety and alignment: The paper explicitly connects compositional generalization to AI safety concerns, as the ability to combine learned skills in novel ways relates to how models might behave in unforeseen situations—an important consideration for developing more predictable and controllable AI systems.