Relate Production Changes to Underlying Grammatical Knowledge

Determine to what extent changes in grammatical production by child-like language models can be interpreted as evidence of changes in their underlying grammatical knowledge.

Background

The study finds that reinforcement-learning models trained with caregiver-like rewards can become more grammatical in free generation, while showing no corresponding improvement on the BLiMP and Zorro minimal-pair benchmarks. The authors note that improved generated output may reflect a shift in production behavior without demonstrating acquisition of new grammatical regularities or generalization beyond the utterance types directly affected by fine-tuning.

The unresolved issue is therefore how to interpret production-based improvements: they may indicate genuine changes in grammatical knowledge, or merely changes in the model’s generation policy. The question is important for evaluating interactive language-learning systems because benchmark performance and free-generation grammaticality appear to capture different aspects of learning.