Tiny transformers can't compose concepts on the fly — and a new study pinpoints exactly why, and when, the format of their inputs matters.
Researchers trained transformers with 6,000 to 10,000 parameters on exhaustively enumerated "factored worlds" — small, fully mapped environments where every possible input can be tested without sampling error. They tested three input types: symbolic tokens, clean "oracle" codes, and entangled perceptual vectors. None of them produced reliable zero-shot compositional binding; every input format landed at or below chance despite a theoretical ceiling of 1.0. For few-shot learning, though, two factors explained most of the variation: how much an input pathway shares parameters across the network, and how readable the underlying code is. Notably, the cleanest oracle input was not the most sample-efficient, which runs counter to a common assumption.
The findings complicate a tidy story that researchers and engineers often tell: that cleaner, more structured inputs automatically produce better generalization. Instead, the study suggests that the inductive biases baked into a model's architecture — not just data quality — shape what it can and cannot learn efficiently. That has practical implications for how teams design tokenization schemes and input representations for small, specialized models.
The study is limited to very small transformers on finite synthetic tasks, so extrapolating to production-scale models takes more than a leap of faith — it takes another paper.