UT Austin's Tiny 8M Model Generates Testable Child Language Hypotheses
Researchers use small language models trained on child-directed speech as simulated learners to generate testable hypotheses about how children acquire dative verbs.
- Framework uses small LMs as simulated learners to generate novel, lab-testable hypotheses about child language acquisition.
- Case study: dative verb learning (
give X to Yvs.give Y X) using an 8.4M-parameter OPT model trained on AO-CHILDES child-directed speech. - Model replicates three known child findings including asymmetric DO-to-PO generalization and cross-structure training advantage.
- Harmonic Alignment and Animacy Prototypicality (HAAP) scoring beats 127 counterfactual schemes at predicting generalization.
- First postverbal argument has disproportionately strong effect on cross-dative generalization in LM learners.
- Code, stimuli, and five model seeds released on GitHub and Hugging Face.
Kanishka Misra of UT Austin and Najoung Kim of Boston University use a small language model as a simulated learner in their paper. The model searches hundreds of conditions for patterns that researchers can test with children. Their case study examines how learners determine which verbs can appear in two English dative constructions, a problem with too many feature combinations to test exhaustively in a developmental lab.
Why verbs create a learnability trap
English dative sentences describe transfers between an agent, a theme, and a recipient. The verb give supports a prepositional-object construction, or PO, as in Najoung gave a treat to Cookie, and a double-object construction, or DO, as in Najoung gave Cookie a treat. The verb donate ordinarily takes the PO form, making I donated the library the book unacceptable to most speakers.
Baker’s paradox names the learning problem created by such lexical restrictions. Children hear grammatical uses of verbs but receive little direct evidence that an unattested form is prohibited. An absent construction could be ungrammatical, rare, or simply missing from the child’s limited input, yet children eventually avoid many overgeneralizations.
Controlled studies can vary properties such as animacy, definiteness, discourse status, phrase length, and pronoun use. Applying those features to the theme and recipient produces hundreds of plausible exposure conditions. Recruiting enough young children to evaluate the entire grid would require an impractical number of trials and participants.
A five-step filter for simulated learners
Misra and Kim describe language models as “animal models” for acquisition research, using the term for inexpensive, manipulable proxies whose predictions require validation in humans. Their workflow applies five filters:
- Choose a learning problem with measurable human behavior.
- Set minimum competence requirements for the model.
- Verify that it reproduces established experimental findings.
- Run controlled simulations across the larger hypothesis space.
- Convert robust model patterns into human experiments.
The analogy concerns experimental use. Behavioral agreement can justify testing a model-derived prediction, while leaving open whether the model and child rely on similar internal mechanisms.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
