Neil Robert Bramley

The ability to search complex problem spaces efficiently is a hallmark of intelligent behaviour, yet how search varies between people, and whether that variation is systematically associated with sub-clinical psychopathology, remains relatively underexplored. To address this, N = 203 adults completed a generative search task known as 'Zendo', in which the aim is to discover the hidden rules governing visual arrays of cones of varied size and colour by constructing and testing novel arrays. Two-hundred participants also completed the 53-item Brief Symptom Inventory, which measures anxiety, depression, hostility, interpersonal sensitivity, obsessive-compulsive behaviour, paranoid ideation, phobic anxiety, psychoticism, and somatisation. Analyses reveal significant individual differences in generative search behaviour that meaningfully shape learning outcomes. Individuals tailor their search to the structure of the problem space in light of the feedback that they elicit, and those who elicit a balance of positive and negative feedback narrow the problem space most effectively and show a learning advantage at test. Although we report reliable variation in sub-clinical psychopathological traits, converging analyses, including Bayes factor analysis and equivalence tests, provide positive evidence for a true null at the intersection of psychopathological traits and generative search behaviour, which is bounded at most to a small-to-moderate effect for a reliable measure of search efficiency. Overall, people show significant individual differences in their ability to search complex problem spaces efficiently, though this variation is largely orthogonal to sub-clinical psychopathology.

Heuristics have long been viewed as simple decision rules that sacrifice optimality for cognitive efficiency. However, this perspective fails to fully account for their effectiveness in complex, real-world environments. This paper proposes a reinterpretation of heuristics as realizations of likelihood-free approximation to Bayesian inference (LFI), bridging the gap between heuristic and normative approaches to decision-making. We argue that many heuristics can be understood as implementations of Approximate Bayesian Computation (ABC), where observed data are compared to prior predictions via summary statistics (a form of LFI). This view situates heuristics within a probabilistic framework, explaining how they can represent uncertainty and adapt to environmental structure. Our approach addresses three key challenges in the field: It relaxes strong assumptions about cue-target relationships in heuristic models, accounts for implicit uncertainty representation in heuristic decision-making, and invites a more appropriate benchmark for evaluating heuristics' performance. To demonstrate the practical applicability of our framework, we present two case studies --- a classic cue-based decision task and an ecological causal learning task. These show that an ABC model using summary statistics can well approximate the benchmark performance provided by unconstrained Bayesian inference, explain the apparent emergence of different heuristics given environmental changes, and capture human behaviors better than a rational Bayesian benchmark without introducing new mysteries. By unifying heuristics and Bayesian inference through the lens of likelihood-free methods, this work provides a more nuanced understanding of human decision-making, and provides a bridge to the literature on inference-by-sampling.

Human reasoning is often affected by order, such that later judgments depend on earlier ones. For example, order effects have been observed when people answer counterfactual questions. The order in which questions are asked affects whether participants backtrack -- imagining how variables that are upstream of a counterfactual change might have been different. Some counterfactual theories predict backtracking while others do not. We build on this prior work and develop a computational model of counterfactual reasoning that captures order effects. We evaluate different versions of this model on existing empirical results, finding that a model which backtracks and produces systematic order effects best explains human judgments. We conclude by discussing implications of the finding that counterfactual reasoning requires a context-sensitive evidence source for both resource-rational and discourse-coherent explanations of reasoning.

An everyday challenge in social learning is aggregating information from multiple informants, each of whom may have access to a different aspect of a problem. In such setting, simply identifying and following the most knowledgeable informant can weaken complementary evidence and lead to suboptimal decisions. We propose a normative solution based on mentalizing, also known as Theory of Mind: leveraging the complementarity of informants’ partial knowledge to infer the hidden evidence behind judgments, and aggregating the inferred evidence to update beliefs. We investigate whether people perform these sophisticated Theory-of-Mind-based computations when aggregating social information. Across three preregistered experiments, we had participants integrate judgments from agents who they knew to have partially overlapping but distinct perspectives on the same problem to make their own judgment. We designed scenarios to separate normative predictions from several more or less sophisticated heuristic accounts. We formalized three levels of social inference: (i) Level-2 inference, full mentalizing that reconstructs evidence beyond stated judgments; (ii) Level-1 inference, heuristic belief updating by assigning greater weight to the agent (or combinations of agents) expected to be more informative given their access to the problem; (iii) Level-0 inference, unweighted vote tallying. Participants’ response was best captured by more resource-rational Level-1 inference, and only a minority were best fit by Level-2 mentalizing. We conclude that people often fail to leverage the full implications of social evidence but suggest that this is also often a resource-rational behavior given the high cost of mentalizing inferences.

The medial prefrontal cortex is central to learning flexible internal models across diverse domains, yet the functional specialization enabling this remains unknown. We tested whether medial prefrontal specialization is representational (encoding domain-specific features) or computational (implementing domain-general computations). During fMRI, participants learned probabilistic features of virtual environments representing spatial, social, and sequential domain knowledge. Although each domain used different features, they shared the same feature-to-latent state mapping, matching their computational demands. The medial PFC showed no domain-specific feature representations. Instead, its neural patterns revealed a triad of specialized yet domain-invariant computations. Ventromedial PFC patterns reflected probabilistic inference, abstracting hidden probability distributions from observations and tracking trial-wise posterior task state changes within a low-dimensional latent space. Anteromedial PFC organized task states along orthogonal axes, tracking directional shifts within states and switches between different states, suggesting a global task coordinate system. Dorsomedial PFC patterns represented task dynamics, using predictive surprise to monitor validity of the current internal model and switch task policies. These results suggest a principled architecture in medial PFC where three general-purpose computations jointly enable learning world-models across diverse environments.

A longstanding focus in the causal learning literature has been on inferring causal relations from contingencies, where these abstract away from time by collating independent instances or by aggregating over regularly demarcated trials. In contrast, individual causal learners encounter events in their daily lives that occur in a continuous temporal flow with no such demarcation. Consequently, the process of learning causal relationships in naturalistic environments is comparatively less understood. In this article, we lay out a rational framework that foregrounds the role of time in causal learning. We work within the Bayesian rational analysis tradition, starting by considering how causal relations induce dependence between events in continuous time and how this can be modeled by stochastic processes from the Poisson-Gamma distribution family. We derive the qualitative signatures of causal influence and the general computations needed to infer structure from temporal patterns. We show that this rational account can parsimoniously explain the human preference for causal models that invoke shorter, more reliable, and more predictable causal influences. Furthermore, we show this provides a unifying explanation for human judgments across a wide variety of tasks in the reanalysis of seven experimental data sets. We anticipate the framework will help researchers better understand the many manifestations of continuous-time causal learning across human cognition and the tasks that probe it, from explicit causal structure induction settings to implicit associative or reinforcement learning settings. (PsycInfo Database Record (c) 2026 APA, all rights reserved).

Reliance on internal predictive models of the world is central to many theories of human cognition. Yet it is unknown whether humans acquire multiple separate internal models, each evolved for a specific domain, or maintain a globally unified representation. Using fMRI during naturalistic experiences (movie watching and narrative listening), we show that three topographically distinct midline prefrontal cortical regions perform distinct predictive operations. The ventromedial PFC updates contextual predictions (States), the anteromedial PFC governs reference frame shifts for social predictions (Agents), and the dorsomedial PFC predicts transitions across the abstract state spaces (Actions). Prediction-error-driven neural transitions in these regions, indicative of model updates, coincided with subjective belief changes in a domain-specific manner. We find these parallel top-down predictions are unified and selectively integrated with visual sensory streams in the Precuneus, shaping participants’ ongoing experience. Results generalized across sensory modalities and content, suggesting humans recruit abstract, modular predictive models for both vision and language. Our results highlight a key feature of human world modeling: fragmenting information into abstract domains before global integration.

Many real-world problems are defined by complex systems of interlocking constraints. How people are able to solve these problems with such limited working memory capacity remains poorly understood.We propose a formal model of human problem-solving that uses metacognitive knowledge of its own memory limits and imperfect reasoning to guide subproblem choice. We compare our model to human gameplay in two experiments using a variant of the classic game Minesweeper. In Experiment 1, we find that participants' accuracy was influenced both by the order of subproblems and their ability to externalize intermediate results, indicative of a memory bottleneck in reasoning. In Experiment 2, we used a mouse-tracking paradigm to assess participants' subproblem choice and time allocation. The model captures key patterns of subproblem ordering, error, and time allocation. Our results point toward memory limits and strategies for navigating those limits --- including the careful choice of subproblems and memory-offloading --- as central elements of human problem-solving.

People often make judgments about uncertain facts and events, for example 'Germany will win the world cup'. Judgment under uncertainty is often studied with reference to a normative ideal according to which people should make guesses that have a high probability of being correct. According to this normative ideal, you should say that Germany will win the world cup if you think that Germany is in fact likely to win. We argue that in many cases, judgment under uncertainty is instead best conceived of as an act of lossy compression, where the goal is to efficiently encode a probability distribution, rather than express the probability of a single outcome. We test formal computational models derived from our theory, showing in four experiments that they accurately predict how people make and interpret guesses. Our account naturally explains why people dislike vacuously-correct guesses (like 'Some country will win the world cup'), and sheds light on apparently sub-optimal patterns of judgment such as the conjunction fallacy.

Social learning can be a powerful tool, allowing us to acquire knowledge and adaptive behaviours while bypassing many of the costs of learning through direct experience. However, not everyone’s behaviour is equally valuable to learn from, as other people’s goals or preferences may differ dramatically from our own. In this paper, we consider the problem of selectively learning from others on the basis of direct and indirect inferences about their task-relevant preferences. Specifically, we focus on the setting where a social learner must generalise preference judgements across individuals using shared features and other cues, and so develop a formal account that can reconcile a seemingly disparate empirical picture of group-based selective social learning. Across three behavioural experiments, we demonstrate that people are sensitive to the contextual significance of group identity cues when choosing who to learn from in partially observed environments. We show that this behaviour cannot be accounted for by a range of simpler heuristic strategies.