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

This study investigates fundamental differences in the acquisition of morphological patterns by humans and large language models (LLMs) within an artificial language learning paradigm. Specifically, it compares how each system responds to variations in input structure-blocked versus interleaved sequences and juxtaposed versus spaced presentation-across verb classification and inflection tasks. While LLMs (GPT4mini, DeepSeek_V3, Llama3.1) consistently outperformed humans-demonstrating superior few-shot learning ability-their learning mechanisms diverged sharply from human cognition in three key aspects. First, humans benefited more from interleaved compared to blocked sequences under the juxtaposition condition, consistent with the discriminative contrast hypothesis, whereas LLMs showed model-dependent performance (interleaving effect independent of juxtaposition in GPT4mini / Llama3.1 vs. less sensitivity to input sequencing in DeepSeek_V3). Second, spacing effects revealed a divergence: while human learners benefited from spaced exemplars, LLMs performed better with juxtaposed input. Third, humans exhibited metacognitive illusions-preferring blocked study despite superior interleaved performance-whereas LLMs returned either consistent interleaving option (Llama3.1/DeepSeek_V3) or statistical tendency resembling human misjudgments (GPT4mini). These findings reveal that human learning mechanisms such as interleaving and spacing do not straightforwardly apply to LLMs, and different model architectures mediate input-structure sensitivity.