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

Problem Statement
Human cognition is deeply influenced by the environments in which it operates, yet it remains an open question whether learning and decision-making can be explained as a principled adaptation to the statistical structure of real-world tasks. Two dominant frameworks attempt to address this: rational analysis, which derives optimal strategies from formal environmental models but struggles with complex, undefined environments, and ecological rationality, which relies on hand-crafted heuristics tuned to specific task structures but lacks a principled method for deriving these heuristics in novel domains. The core challenge is to develop a computational framework that unifies the normative foundations of rational analysis with the ecological grounding of real-world tasks, without requiring researchers to manually specify environmental distributions or design specific heuristics.

Methodology: Ecologically Rational Analysis and ERMI
The authors introduce Ecologically Rational Analysis, a framework that leverages Large Language Models (LLMs) and meta-learning to derive computational models of human cognition. The approach consists of three main stages:

1. Scalable Task Generation via LLMs: The authors use LLMs (specifically CLAUDE-V2) as generative engines to create ecologically valid cognitive tasks at scale. This is a two-stage process:

   • Feature Synthesis: The LLM generates plausible task features and targets (e.g., mapping "sodium, fat, protein" to "healthy/unhealthy" categories).
   • Data Generation: The LLM generates corresponding input-target pairs consistent with these features, creating datasets that approximate the statistical regularities of naturalistic problem spaces.
   • Validation: The generated tasks are statistically compared to real-world datasets (e.g., OpenML-CC18, real-world regression tasks) to verify ecological validity. Metrics include classification accuracy, feature correlation, sparsity (Gini coefficient), and linearity.

2. Meta-Learning for In-Context Inference: The authors train neural networks (specifically transformer-based architectures) on these LLM-generated tasks using meta-learning. The goal is to derive Ecologically Rational Meta-learned Inference (ERMI) models. These models learn to perform in-context learning: they adapt to the structure of a new problem by modifying internal activations based on a sequence of input-output examples, without updating their parameters. This process allows the models to internalize the "ecological priors" (statistical regularities) of the training distribution, approximating Bayes-optimal inference conditioned on those statistics.

3. Evaluation Against Human Behavior: The performance of ERMI is evaluated across 15 experiments spanning three cognitive domains: Function Learning, Category Learning, and Decision Making. The models are tested on their ability to replicate hallmark behavioral signatures (e.g., learning curves, generalization patterns, strategy shifts) and are compared against established cognitive models (e.g., Rational Model of Categorization, Generalized Context Model, Rule-based models) and meta-learned inference models with hand-crafted priors (MI).

Key Contributions

• A New Framework: The proposal of "Ecologically Rational Analysis," which synthesizes rational analysis and ecological rationality by automating the derivation of adaptive strategies from the statistical structure of natural environments.
• ERMI Models: The development of a new class of learning algorithms (ERMI) that internalize ecological priors through meta-learning on LLM-generated tasks, eliminating the need for hand-crafted heuristics or explicit parameter updates during inference.
• Empirical Validation: Demonstrating that ERMI captures human behavior across diverse domains, outperforming several established cognitive models in trial-by-trial prediction.
• Methodological Innovation: Utilizing LLMs not as direct behavioral models, but as tools to generate the ecological priors necessary for training adaptive computational models.

Results

• Ecological Validity: LLM-generated tasks successfully mirrored the statistical properties of real-world data, including distributions of function types (linear, exponential, etc.), feature correlations, and sparsity.
• Function Learning: ERMI replicated five hallmark human findings: (i) better learning of positive linear functions than negative ones; (ii) faster learning of linear vs. nonlinear functions; (iii) superior learning of monotonic vs. non-monotonic functions; (iv) easier learning of non-cyclic vs. cyclic functions; and (v) better performance in interpolation than extrapolation. ERMI also exhibited the human-like bias of underestimating values during extrapolation (bias toward zero offset) and provided a better fit to human data than meta-learned inference with hand-crafted priors.
• Category Learning: ERMI captured the difficulty ordering of Shepard's six category types (TYPE 1–6) and outperformed competing models in explaining human choices (highest posterior model frequency). It also successfully modeled the shift in human strategy from prototype-based to exemplar-based learning over time and replicated human generalization patterns to unseen inputs.
• Decision Making: In paired comparison tasks, ERMI adapted its decision strategy based on the statistical structure of the environment (ranking, direction, or unknown conditions), implementing one-reason, equal-weighting, or weighted-additive strategies respectively, matching human behavior. ERMI accounted for human responses more frequently than baseline models (Single-cue, Equal-weighting, Neural Networks, and MI).

Significance and Claims
The paper claims that adaptive alignment with the statistical structure of ecologically valid environments is sufficient to account for a broad spectrum of human learning and decision-making behavior.

The authors argue that human cognition may not require complex, domain-specific heuristics designed by researchers, but rather emerges from an attunement to the statistical regularities of the world. By using LLMs to distill these regularities into training data, ERMI demonstrates that general-purpose systems can mirror how people learn, categorize, and decide. The work suggests that much of human cognition reflects an "ideal statistical inference" under the constraints of natural task structures.

The authors remain modest regarding the scope of their claims, noting that while ERMI reduces the burden of hand-crafting priors, it does not eliminate the need for scientific judgment in feature selection. They acknowledge that future work must determine if additional components (e.g., specific cognitive constraints like working memory limits, or participant-specific factors) are required to fully account for human behavior, particularly in cases where humans deviate from ideal inference. The paper posits that this framework offers a scalable path to understanding the ecological roots of cognition, potentially extending to embodied, high-dimensional environments in the future.