Meta-learning ecological priors from large language models explains human learning and decision making

Researchers introduce Ecologically Rational Meta-learned Inference (ERMI), a computational framework combining large language models with meta-learning to model human cognition as adaptive optimization to real-world environments. The approach successfully predicts human behavior across 15 experiments in function learning, category learning, and decision-making, suggesting human cognition reflects principled adaptation to ecological statistical structures.

This research bridges cognitive science and machine learning by proposing that human learning emerges from implicit adaptation to environmental regularities rather than hard-wired cognitive mechanisms. The team leverages LLMs to generate realistic task distributions, then uses meta-learning to derive models optimized for these ecological contexts. ERMI achieves superior trial-by-trial prediction accuracy compared to established cognitive models, demonstrating practical explanatory power.

The work addresses a longstanding debate in cognitive science: whether human behavior reflects rational principles adapted to specific environments or follows domain-specific heuristics. Traditional rational analysis assumes abstract optimality, while this framework grounds rationality in actual task statistics humans encounter. By using LLMs to capture real-world statistical structure at scale, the researchers sidestep the need for hand-crafted cognitive tasks, dramatically expanding the scope of testable predictions.

For AI development, this research illuminates how neural networks trained on natural data distributions develop human-like learning algorithms. The approach suggests meta-learning on ecologically valid tasks produces models whose inductive biases align with human cognition. This has implications for interpretability and alignment—understanding why AI systems behave like humans requires understanding the environments they optimize for.

The findings open new research directions: testing ERMI across clinical populations, exploring whether ecological structure explains cognitive biases, and leveraging meta-learned models to improve human-AI collaboration. Future work should examine whether this framework extends to social cognition and high-level reasoning, where ecological validity becomes harder to define.

• →ERMI uses LLMs to generate ecologically valid task distributions for meta-learning human-aligned cognitive models.
• →The framework outperforms established cognitive models at predicting human behavior across function learning, category learning, and decision-making tasks.
• →Results support the hypothesis that human cognition reflects adaptive optimization to environmental statistical structures rather than fixed heuristics.
• →This approach bridges interpretability gaps by showing how neural networks develop human-like algorithms when trained on realistic task distributions.
• →The research suggests meta-learning on ecological priors could improve human-AI alignment and collaborative systems.