Overhypothesis Induction in Small Language Models Trained at Child Scale: A Pre-registered Synthetic Corpus Study

This pre-registered study investigates whether small transformer language models (≤50M parameters) trained on child-scale synthetic corpora (~100k tokens) spontaneously acquire overhypotheses which are abstract, second-order inductive biases that generalise feature regularities across object categories. Overhypothesis induction is a well-documented phenomenon in human conceptual development (Kemp et al., 2007; Smith et al., 2002), but has not previously been tested in neural language models using controlled synthetic stimuli and pre-registered evaluation protocols. Models are trained on eight corpus conditions that systematically vary the presence of feature-category regularities (shape, colour, texture) and labelling structure. A forced-choice wug test battery (1,040 items across 14 item types) evaluates first-order generalisation and second-order overhypothesis formation using log-probability scoring. Study 1 tests four primary hypotheses: (H_main) above-chance second-order generalisation in the regular condition; (H_order) stronger second-order than first-order scores; (H_label) label presence amplifies overhypothesis strength; and (H_swap) feature-domain specificity of acquired priors. Secondary analyses include dose-response curves, Bayesian alignment metrics, and mechanistic interpretability probes. All 155 training runs (125 primary + 30 dose-response) use 5 random seeds per condition. Phase 2 training will not commence before this registration is accepted.