Tracing Morphology, Semantics, and Divergence in Language Acquisition

Key points

• In this work, we leverage intermediate checkpoints from open-source models (e.g., Pythia and OLMo) under teacher-forcing evaluation to systematically investigate language acquisition across training steps and model scales.1
• We focus on four core characteristics: morphological structure acquisition, embedding drift, repetition dominance, and co-occurrence effects.1
• Our findings reveal that models acquire morphological structures well before semantic embeddings stabilize.1
• Furthermore, we show that training dataset size has a positive impact on divergence recognition, while co-occurrence during pre-training and repetition during inference have a negative impact on divergence recognition and associative link generation.1

Sources (1)

• [1]Tracing Morphology, Semantics, and Divergence in Language Acquisition
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 06:29 PM

  Understanding the dynamics of language acquisition in Large Language Models (LLMs) is crucial for uncovering how statistical representations evolve into structured linguistic capabilities.

  In this work, we leverage intermediate checkpoints from open-source models (e.g., Pythia and OLMo) under teacher-forcing evaluation to systematically investigate language acquisition across training steps and model scales.

Extractive summary: sentences quoted from the sources.