Do large language models learn like humans: Interleaved and spaced practice in morphological learning.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computational Linguistics

Background:

• Understanding human learning mechanisms is crucial for developing advanced AI.
• Large Language Models (LLMs) show promise in language acquisition but their learning processes are not fully understood.
• Investigating how LLMs learn artificial languages can reveal fundamental differences compared to human cognition.

Purpose of the Study:

• To compare the acquisition of morphological patterns in an artificial language between humans and LLMs.
• To examine how input structure (blocked vs. interleaved, juxtaposed vs. spaced) affects learning in both systems.
• To identify divergences in learning mechanisms and cognitive strategies.

Main Methods:

• An artificial language learning paradigm was employed.
• Participants included humans and three LLMs (GPT4mini, DeepSeek_V3, Llama3.1).
• Verb classification and inflection tasks were used with varied input sequencing and presentation.

Main Results:

• LLMs consistently outperformed humans, demonstrating superior few-shot learning.
• Human learning was influenced by interleaving and spacing effects, consistent with cognitive theories.
• LLMs showed model-dependent responses to input structure, diverging from human learning patterns.
• Humans exhibited metacognitive illusions regarding learning preferences, unlike most LLMs.

Conclusions:

• Human learning mechanisms like interleaving and spacing do not directly translate to LLMs.
• LLM performance is sensitive to input structure, but in ways distinct from human cognition.
• Different LLM architectures exhibit varying degrees of sensitivity to input sequencing and presentation.