Why Did Apple Fall: Evaluating Curiosity in Large Language Models

Researchers have developed a comprehensive evaluation framework based on human curiosity scales to assess whether large language models exhibit curiosity-driven learning. The study finds that LLMs demonstrate stronger knowledge-seeking than humans but remain conservative in uncertain situations, with curiosity correlating positively to improved reasoning and active learning capabilities.

This academic research addresses a fundamental question about artificial intelligence cognition: whether large language models can develop curiosity similar to human psychological traits. By adapting the Five-Dimensional Curiosity scale Revised (5DCR)—a validated human assessment tool—researchers created a methodology to measure information-seeking, thrill-seeking, and social curiosity dimensions in LLMs. The work represents an important pivot in AI evaluation, moving beyond traditional benchmarks that focus on accuracy and task completion toward understanding model behavior patterns that mirror human learning psychology.

The findings reveal a nuanced picture. LLMs exceed human baseline performance in pure knowledge acquisition and information seeking, suggesting these systems are optimized for pattern recognition and data absorption. However, they exhibit risk aversion in ambiguous scenarios, preferring predictable outputs over exploratory responses. This conservative tendency indicates current models lack the intrinsic motivation mechanisms that drive human curiosity-driven discovery.

The confirmed relationship between curiosity metrics and reasoning ability has significant implications for AI development. If curiosity behaviors genuinely enhance model performance on complex reasoning tasks, this could justify architectural redesigns that encourage more exploratory behavior during training phases. For developers and researchers, this suggests future LLM improvements may require psychological frameworks alongside traditional machine learning optimization.

The research opens pathways for developing more autonomous, self-directed learning systems that don't require constant external supervision or task specification. However, the current conservative bias in uncertain environments remains a limiting factor for deployment in novel problem-solving scenarios where calculated risk-taking proves necessary.

• →LLMs demonstrate stronger information-seeking capabilities than humans but remain conservative when facing uncertainty.
• →Curiosity-driven behaviors correlate with improved reasoning and active learning performance in language models.
• →The 5DCR framework provides a validated methodology for measuring psychological traits in artificial systems.
• →Current models lack intrinsic motivation mechanisms that naturally drive human exploratory learning.
• →Enhanced curiosity design could lead to more autonomous and self-directed AI learning systems.