Tracking how compound words lose meaning over decades

The paper doesn't just measure semantic drift in noun compounds; it formalizes prediction of that drift across decades. The key novelty is the per-decade ratings dataset itself, which enables models to learn temporal patterns rather than simply observe historical change after the fact.

This connects to the RAGU work from the same day, which challenged the assumption that bigger models solve harder linguistic problems. Here, researchers are testing whether language models can reason about how meaning changes over time, a form of linguistic reasoning that may not scale with parameter count alone. Like RAGU's finding that task-specific optimization beats raw model size, compositionality prediction likely requires models tuned for temporal reasoning rather than general language understanding. Both papers suggest NLP benchmarks are shifting from static evaluation toward capability-specific measurement.

If the top-performing models on this benchmark are smaller, domain-adapted variants rather than frontier LLMs, that confirms the pattern RAGU identified. If frontier models dominate instead, it suggests temporal linguistic reasoning is orthogonal to the parameter-scaling dynamics we've seen in other domains.