Abstract
An open question regarding how people develop their models of the world is how new candidates are generated for consideration out of infinitely many possibilities. We discuss the role that evolutionary mechanisms play in this process. Specifically, we argue that when it comes to developing a global world model, innovation is necessarily incremental, involving the generation and selection among random local mutations and recombinations of (parts of) one's current model. We argue that, by narrowing and guiding exploration, this feature of cognitive search is what allows human learners to discover better theories, without ever grappling directly with the problem of finding a "global optimum," or best possible world model. We suggest this aspect of cognitive processing works analogously to how blind variation and selection mechanisms drive biological evolution. We propose algorithms developed for program synthesis provide candidate mechanisms for how human minds might achieve this. We discuss objections and implications of this perspective, finally suggesting that a better process-level understanding of how humans incrementally explore compositional theory spaces can shed light on how we think, and provide explanatory traction on fundamental cognitive biases including anchoring, probability matching and confirmation bias.
Introduction: The Problem of Infinite Theory Spaces
A central mystery in cognitive science is how humans develop "world models"—the complex, hierarchical systems of concepts we use to navigate reality. While Bayesian frameworks describe how we select the best hypothesis from a set, they often struggle to explain where those hypotheses come from in the first place.
The space of possible theories is effectively infinite. If a learner is trying to induce the "program" that governs a new environment, they cannot possibly consider every possible line of code. This presents a "monkeys and typewriters" problem: how does the mind generate high-quality candidate models without getting lost in a sea of random, nonsensical possibilities? We argue that human world models are not built through global optimization, but through an incremental, evolutionary process of "blind" local variation and selective retention.
Algorithmic Mechanisms: Tinkering and Bootstrapping
To explain how the mind navigates this vast space, we propose a "learning as program induction" framework, drawing a parallel between biological evolution and cognitive growth. Two primary algorithmic mechanisms facilitate this process:
1. Stochastic Local Search (MCMC)
Instead of starting from scratch when faced with new data, the mind "tinkers" with its current model using Markov Chain Monte Carlo (MCMC) search. This involves two primary operations:
- Tree Regrowth: Deleting a branch of a conceptual tree and regrowing it.
- Tree Surgery: Making small, local edits, such as replacing a single operator.
By restricting search to these local mutations, the mind ensures that new ideas are always "local" to what is already believed, preventing the system from collapsing into total incoherence.
2. Conceptual Bootstrapping (Adaptor Grammars)
To prevent the mind from constantly "reinventing the wheel," we utilize Adaptor Grammars. This mechanism allows the learner to "cache" successful sub-programs—useful conceptual fragments—and treat them as new primitives. This creates a "ratchet effect," where complexity builds upon itself over time. Once a concept is cached, it can be used as a single block in future mutations, allowing for the evolution of increasingly sophisticated models.
Empirical Evidence: The "Magic Eggs" Task
This theoretical synthesis is supported by results from causal induction experiments, specifically the "Magic Eggs" task (Zhao et al., 2023). In this paradigm, participants must learn rules governing how features of "magic eggs" (such as spots or stripes) change the length of "sticks."
Order Effects and Path Dependence
The research highlights the importance of the order in which evidence is presented. Participants shown evidence in an order that allowed them to "cache" a simple sub-concept first (e.g., eggs with no spots) were significantly more successful. Their generalization accuracy reached 45%, compared to near-chance levels for those shown complex, "noisy" evidence first. This suggests that the order of learning matters deeply; if we learn the "wrong" sub-concepts early on, they become the building blocks for future errors, making it harder to reach the truth.
Model Fit
Quantitative analysis shows that MCMC-based "Tree Surgery" models provide a better fit for human conceptual revision patterns than normative Bayesian models or simpler "Tree Regrowth" algorithms. Humans tend to make the smallest possible change to a theory that resolves a current discrepancy, rather than re-evaluating their entire worldview.
Cognitive Biases as Evolutionary Features
This evolutionary perspective suggests that many common cognitive biases are not "bugs," but rather inevitable byproducts of a local search strategy:
- Anchoring: This is a byproduct of local search; we stay close to our starting point because jumping too far into the theory space usually results in a "lethal mutation"—a model with zero fitness or explanatory power.
- Confirmation Bias: In an infinite space, any hypothesis-driven search will appear "confirmatory." We can only test the narrow slice of the world our current model allows us to conceive.
- Neurath’s Ship: We are like sailors repairing a ship at sea. We cannot disassemble the entire hull to find a "global optimum"; we must patch it piece by piece while staying afloat.
Conclusion and Future Directions
Innovation at the "cutting edge" of knowledge is necessarily blind. We do not know what a better theory looks like until we stumble upon it via local mutation. While program induction is computationally expensive, the human mind manages this through efficient local search and library learning.
Future work should explore neurosymbolic hybrids—combining neural networks for fast feature extraction with symbolic search for reasoning—to explain how this works at scale. Additionally, the role of "offline thinking" (such as dreaming) may be essential for refactoring and optimizing the internal library of concepts. By understanding the mind as a "blind watchmaker" of its own world models, we gain a clearer picture of both human genius and human fallibility.
Summary updated