Sharing cutting-edge literature in cognitive neuroscience.
Basic Information
Title: The explicit-Bayes hypothesis for cognition
Publication Date: 2026-04-29
Journal: Nature Reviews Psychology
Impact Factor: 21.8
Obtain the Original Paper:
Click "Read Original" at the bottom of the page to jump to the paper's webpage.
Background and Pain Points
"Cognition is Bayesian" has been a common refrain in cognitive science, largely because Bayesian inference can indeed provide a unified language for rational analysis across different tasks. However, this phrase often conflates the computational level and the algorithmic level within Marr's three-level analysis. The fact that behavioral data can be well-fitted by a Bayesian model does not mean the cognitive algorithm is explicitly executing Bayes' rule. The authors argue that the persistent nature of this debate stems from this confusion of levels, which makes the proposition too broad and difficult to falsify.
Core Framework and Integrative Logic
This paper proposes the explicit-Bayes hypothesis, which advocates for testing whether cognitive algorithms explicitly apply Bayes' rule in isolation. By doing so, Bayesianism at the computational level can still explain "why behavior is this way," while Bayesianism at the algorithmic level becomes a mechanistic hypothesis that can be compared and refuted.
The authors emphasize that the criterion is not "whether it represents uncertainty" or "whether it is rational," but rather whether it explicitly computes the prior, likelihood, and posterior, and applies Bayes' rule. By this standard, some sampling, amortized inference, and predictive coding models, while potentially viewed as Bayesian at the computational level, might be classified here as non-explicit-Bayes. To test this hypothesis, tasks need to be simple enough to allow control over the model space, yet also rich enough in conditions to enhance model identifiability.
Key Insights and Future Directions
Insight 1: The authors tighten the claim at the algorithmic level, not dismiss the Bayesian framework at the computational level
This commentary serves to decompose the question "Is cognition Bayesian?" into two separate inquiries: whether a Bayesian explanation is suitable at the computational level, and whether Bayes' rule is explicitly used at the algorithmic level. If the latter is repeatedly disproven, it constrains the interpretive level of Bayesian models, but does not diminish its normative status nor its role as an abstract organizational framework.
Insight 2: The value of the explicit-Bayes hypothesis lies in drawing a clearer empirical boundary for "Bayesian"
The authors use "whether Bayes' rule is explicitly applied" as a classification standard because "uncertainty representation" or "computational rationality" are often insufficient to cleanly separate Bayesian from non-Bayesian algorithms. For a specific task, within a given model space, if the best-fitting model explicitly uses Bayes' rule, it supports the hypothesis; conversely, it constitutes a refutation.
Insight 3: Confidence judgments in perceptual decision-making serve as a workable testing ground, but existing support remains preliminary
The authors believe that the perceptual domain is relatively low-dimensional, and confidence judgments often require participants to report probabilities, either explicitly or implicitly, making it a particularly suitable area for testing this hypothesis. The main text reviews that the Bayesian confidence hypothesis has been tested multiple times, and current model comparisons lean more toward a non-explicit-Bayes explanation based on sensory evidence strength. However, this judgment is still preliminary and cannot be directly extrapolated to higher-level domains like memory and reasoning.
TL;DR Summary
This commentary narrows "Is cognition Bayesian?" from a broad slogan into a narrower, more falsifiable algorithmic-level proposition: Do cognitive algorithms explicitly use Bayes' rule? It preserves the explanatory value of Bayesian models at the computational level while demanding that algorithmic-level claims face stricter model comparisons. Currently, confidence research in perceptual decision-making provides a testable paradigm, but conclusions remain largely confined to this specific domain for now.