Turns out the order you teach an AI model matters more than we thought.
A new paper builds a framework for measuring compositional reasoning in language models, the ability to chain simple skills into harder multi-step tasks. Using data-structure problems, which have clean right-or-wrong answers, the researchers ran reinforcement learning post-training and tested transfer in both directions. They found a consistent asymmetry: training a model on the full composed task also makes it better at the simpler sub-skills, but training it on the sub-skills first does not reliably make it better at the composed task. The pattern held up under length extrapolation, structural changes, and transfer to unseen skills, and a small pilot on tool-calling benchmarks suggested it shows up outside the lab too.
This cuts against a common assumption in how AI labs build training curricula, that breaking a hard task into easier pieces and mastering each one is the safe path to competence. If the composed-to-decomposed direction is the one that actually transfers, teams optimizing reasoning models might get more out of throwing harder, end-to-end tasks at RL post-training rather than scaffolding up from basics.
The caveat is that data-structure puzzles are a tidy testbed, not a messy production agent, and the tool-calling result is explicitly a pilot. Worth watching for a larger-scale replication before anyone rewrites their training pipeline.