Language Acquisition Device in Large Language Models

In three linesResearchers propose LAD-inspired PPT, pre-pretraining on MP-STRUCT, a formal language encoding hierarchical composition and long-distance displacement. After 500 steps, this approach matches formal-language baselines in token efficiency while imparting LLMs human-like resistance to structurally implausible languages.

Summary generated by Claude — human-verified