Training a small language model on compositional opaque arithmetic did not reliably improve its ability to compose a new ordered operation—even when every primitive fact in the target domain was available.

Question

Correct answers to multistep problems leave an important ambiguity: did the model learn reusable structure, or did it benefit from exposure, labels, frequency, or a shortcut? We tested whether structural lessons in one domain transfer to a semantically disjoint target domain.

Design

We generated 16 independent semantic worlds. Each contains a fresh four-value operation that is deliberately noncommutative and nonassociative. The target test has 80 frozen nested expressions per world. The source is five-value opaque modular arithmetic with randomized digit and operator names.

Every critical curriculum contained 139 training examples: 16 target primitive facts, 75 source primitive facts, and 48 matched lessons. The structural arm received nested opaque-arithmetic compositions. Its control received an exact atomic placebo: a visible depth-one fact with the same answer, length, vocabulary, character multiset, label distribution, example count, exposure, batch schedule, and optimizer budget—but no valid composed operator application.

Primary result

Source structure312 / 1,28024.375%
Exact atomic placebo342 / 1,28026.719%
Paired-world difference−2.344 pp95% CI: −5.0 to +0.234 pp

All 16 paired worlds passed the prespecified full-training and primitive-mastery gates; none were excluded. Source structure won in 7 worlds, the placebo won in 8, and 1 tied. The one-sided exact paired-world sign test was 22,819 / 32,768 = 0.6964.

Interpretation

The preregistered conclusion is narrow: for this causal decoder, opaque modular source, ordered finite target population, exact atomic placebo, and fixed training budget, we found no reliable positive cross-domain transfer. This is a negative result, not an impossibility theorem about all language models or all forms of reasoning transfer.

What was held fixed

Each fit used 2,048 optimizer updates. The primary model was a small scratch-trained, character-tokenized, decoder-only Transformer. Eligibility was fail-closed: both compared models had to master every training example and every target primitive before their held-out target predictions counted.