TL;DR
This paper argues that human world models are not built through global optimization, but through an incremental, evolutionary process of "blind" local variation and selective retention. By combining Markov Chain Monte Carlo (MCMC) search with "bootstrapping" via Adaptor Grammars, the authors explain how complex conceptual systems can emerge from simple primitives and why human cognition is prone to biases like anchoring and confirmation bias.
What Problem Does This Paper Address?
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 infinite. If you are trying to learn the "program" that governs a new environment, you can't possibly consider every possible line of code. The authors tackle the "monkeys and typewriters" problem: how does the mind generate high-quality candidate models without getting lost in a sea of random, nonsensical possibilities?
Methods
The authors propose a "learning as program induction" framework, drawing a parallel between biological evolution and cognitive growth. They highlight two primary algorithmic mechanisms:
- Stochastic Local Search (MCMC): Instead of starting from scratch, the mind "tinkers" with its current model. They discuss Tree Regrowth (deleting a branch of a conceptual tree and regrowing it) and Tree Surgery (making small, local edits like replacing a single operator). This ensures that new ideas are always "local" to what we already believe.
- Conceptual Bootstrapping (Adaptor Grammars): To prevent the mind from constantly "reinventing the wheel," they use 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.
Datasets
While primarily a theoretical synthesis, the paper relies heavily on results from a specific experimental paradigm:
- The "Magic Eggs" Task (Zhao et al., 2023): A causal induction dataset where participants must learn rules governing how "magic eggs" (with features like spots and stripes) change the length of "sticks."
- The paper also references various established benchmarks in the program induction literature, such as Dreamcoder (Ellis et al., 2021), which uses similar "wake-sleep" library learning on raw pixel data and symbolic tasks.
Key Insights and Ideas
- The "Blind" Innovation: At the "cutting edge" of our knowledge, innovation is necessarily blind. We don't know what the better theory is until we stumble upon it via local mutation.
- Neurath’s Ship: The authors invoke the metaphor of a ship being repaired at sea. We cannot disassemble our entire worldview to find a "global optimum"; we must patch it piece by piece while staying afloat.
- Biases are Features, Not Bugs:
- Anchoring: An inevitable byproduct of local search; we stay close to our "starting point" because jumping too far usually results in a "lethal mutation" (a model with zero fitness).
- Confirmation Bias: In an infinite space, any hypothesis-driven search will look "confirmatory" because we can only test the narrow slice of the world our current model allows us to conceive.
- Path Dependence: The order in which we learn things matters. If we learn the "wrong" sub-concepts early on, they become the building blocks for future errors, making it harder to reach the truth.
Key Results
The paper synthesizes quantitative findings from the "Magic Eggs" experiments to support their theory:
- Order Effects: 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 was 45%, compared to near-chance levels for those shown complex, "noisy" evidence first.
- Model Fit: MCMC-based "Tree Surgery" models provided 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 fixes a current error.
Limitations and Future Work
- Defining Fitness: In these models, "fitness" is usually defined as the Bayesian posterior probability. However, the authors acknowledge that human "fitness" might be driven by utility, emotional resonance, or social conformity, which aren't always aligned with objective truth.
- Computational Cost: Program induction is notoriously expensive. The authors suggest that neurosymbolic hybrids (combining neural networks for fast feature extraction with symbolic search for reasoning) are a necessary next step to explain how this works at scale.
- Offline Thinking: The paper suggests that "thinking" (or dreaming) is essentially the mind running these MCMC chains "offline" to refactor and optimize its internal library, but more empirical work is needed to link these algorithms to specific neural processes like sleep-based consolidation.
Summary updated