How Past Experiences Shape Analogical Reasoning: A Bayesian Framework

When we encounter something new, we don't start from scratch—we draw on a lifetime of experiences to make sense of it quickly. A team of researchers has investigated exactly how this process works, developing both behavioral experiments and a computational model to explain analogical reasoning across multiple learned environments.

The Experiment: Learning Through Virtual Games#

The team designed an online behavioral study in which participants played a series of virtual games, each governed by its own underlying relational structure. By observing participants' performance across these games, the researchers tracked how exposure to one relational pattern influenced expectations about subsequent tasks. The findings revealed a striking effect: when people encountered a particular relational structure repeatedly, they became increasingly biased toward expecting that same structure in new situations.

How Experience Accumulates#

The strength of this bias scaled predictably with repetition. The more times participants had seen a given relational pattern, the stronger their tendency to apply it when facing an unfamiliar test game. This demonstrates that our generalizations from past experiences are not all-or-nothing but rather accumulate and compound as evidence builds.

A New Probabilistic Model#

To explain these observations, the researchers propose a nonparametric Bayesian framework that treats analogical reasoning as a form of inference. Rather than relying on a single source analog, the model integrates predictions from all previously encountered environments, weighting each one according to its analogical relevance to the current problem. This approach captures how people leverage multiple diverse experiences rather than defaulting to just the most similar past case.

Why This Matters#

This framework provides fresh insight into long-standing questions about how analogy supports cognition and learning. By modeling analogy as Bayesian inference over relational structures, the work bridges the gap between classical cognitive theories and modern probabilistic approaches—and offers a foundation for understanding how learners continuously build and deploy knowledge throughout their lives.

The research is available on arXiv.