TLDR: A new computational framework, Ecologically Rational Meta-learned Inference (ERMI), uses large language models (LLMs) to generate realistic cognitive tasks and meta-learning to derive models that explain human learning and decision-making. ERMI unifies rational analysis with ecological grounding, demonstrating that human cognition reflects adaptation to environmental statistics. It successfully captures human behavior across function learning, category learning, and decision-making experiments, outperforming traditional cognitive models by internalizing ecological priors without hand-crafted heuristics.
Human intelligence is remarkably adaptive, allowing us to make sound decisions and learn effectively in complex, unpredictable environments. But how do these abilities arise, and how can we accurately model them? A new computational framework, Ecologically Rational Meta-learned Inference (ERMI), offers a compelling answer by suggesting that much of human cognition can be explained as a principled adaptation to the statistical structure of real-world tasks.
Traditionally, two main frameworks have attempted to explain human learning and decision-making: rational analysis and ecological rationality. Rational analysis seeks optimal strategies within formal models of the environment but is limited to simpler scenarios due to the need for explicit environmental models. Ecological rationality, on the other hand, focuses on simple heuristics tuned to real-world tasks, but it requires researchers to hand-design these strategies, making it challenging to apply to new domains.
ERMI bridges the gap between these two approaches. It introduces a novel method that unifies the normative foundations of rational analysis with ecological grounding. The core idea is to leverage the power of large language models (LLMs) to generate a vast array of ecologically valid cognitive tasks – problems that mirror the kinds of structures humans encounter daily. These tasks are then used to train neural networks through a process called meta-learning, allowing the models to internalize the statistical regularities of naturalistic problem spaces.
How ERMI Works
The ERMI framework operates in two main stages. First, an LLM (such as CLAUDE-V2) is used to generate realistic cognitive tasks. This involves a two-step process: the LLM first synthesizes plausible task features (e.g., predicting whether a food item is healthy based on its sodium, fat, and protein content) and then generates corresponding input-target pairs consistent with these features. The ecological validity of these LLM-generated tasks is rigorously checked by comparing their statistical properties (like classification accuracy, feature correlation, sparsity, and linearity) against real-world datasets.
In the second stage, a transformer-based neural network is trained on these LLM-generated tasks using meta-learning. This training process enables the network to develop a new class of learning algorithms that can adapt flexibly to novel situations without requiring hand-crafted heuristics or explicit parameter updates. This ability, known as in-context learning, allows the model to modify its internal activations based on observed examples, effectively approximating Bayes-optimal inference conditioned on the statistics of its training environment.
Explaining Human Behavior Across Domains
The researchers demonstrated ERMI’s effectiveness across 15 experiments spanning three fundamental cognitive domains:
- Function Learning: ERMI successfully captured hallmark human behaviors in learning relationships between inputs and outputs. It replicated findings such as people interpolating more accurately than extrapolating, and exhibiting biases towards linear functions with positive slopes. ERMI also provided superior trial-by-trial predictions compared to established cognitive models.
- Category Learning: In this domain, ERMI mirrored human learning difficulties across different category structures, showed similar shifts in categorization strategy (from prototype-based to exemplar-based with experience), and replicated human generalization patterns to unseen inputs.
- Decision Making: ERMI demonstrated that it could adopt the same decision-making heuristics as humans, such as one-reason decision-making, equal weighting, or weighted-additive strategies, depending on the statistical structure of the paired comparison tasks. It consistently outperformed other baseline models in predicting human choices.
Also Read:
- AI Agents Master Complex Tasks by Integrating Linguistic Guidance and Direct Experience
- Agentic Reinforcement Learning: Empowering LLMs as Autonomous Decision-Makers
The Broader Impact
The findings suggest that human learning and decision-making largely reflect an adaptive alignment to the ecological structure of the problems encountered in everyday life. ERMI offers a powerful framework for understanding this by automatically deriving computational models that implement approximately optimal strategies adapted to natural environments. This approach significantly reduces the need for extensive hand-engineering of priors or heuristics, a common limitation in previous cognitive modeling efforts.
The work highlights LLMs not as direct models of human behavior, but as “cultural technologies” that distill the collective knowledge of societies, capturing statistical regularities on an unprecedented scale. By using LLMs as generative sources for ecological tasks, ERMI allows researchers to test specific hypotheses about human cognition and reverse-engineer the environmental ingredients that shape our minds.
Future research aims to integrate participant-specific information, incorporate cognitive constraints like attention and working memory, and use ERMI as a foundation for fine-tuning on human choice data to further refine our understanding of human cognition. For more details, you can read the full research paper here.
