We introduce a novel task exploring how people make causal generalizations over the abstract features of the objectsinvolved in a causal interaction. Specifically, we investigate how people generalize from a single observation of two sim-ple objects in which one (the agent, or cause) interacts with another (the recipient, or effect) resulting in some featurechange(s). In line with recent demonstrations of human strength in few-shot concept learning, we find strong and sys-tematic patterns of generalizations that are well explained by a Bayesian inference model favoring simpler causal rules.However, we also identify a clear order effect depending on what order generalizations are made. To capture the observedpatterns, we develop a causal hypothesis generation model that takes peoples natural generalization tendency and the ordereffect into consideration, and outperforms plain Bayesian inference both in computational efficiency and in match to thebehavioral data.

Counterfactual models perturb a situation model to measure the explanatory power of different causes. We modify one to explore how people explain others' behaviour in two online experiments. First, 90 UK-based adults rated the likelihood of various scenarios combining short biographies with characters' trajectories through a gridworld. Then 49 others saw each scenario and outcome, and verbally gave their best explanations for why the character went the way they did. Participants generated a range of explanations for even the most incongruous behaviour, citing factors outside the given information. Using the innovative features of crowdsourced parameters and free text responses, we show people anchor to real world dependency c.75% of the time, and stochastically sample the rest. We present an expanded version of the counterfactual effect size model (Quillien & Lucas 2022) which handle surprising behaviours and performs better than the existing model it is based on.

Optimal decision making depends on people’s ability to generate and test hypotheses of their environments. To understand how hypotheses are tested, researchers often focus on uncovering and documenting the typical testing strategies that people spontaneously use (e.g., a positive or confirmatory testing strategy). Under this approach, it is difficult to account for people’s capacity to overcome the limitations of their existing hypothesis testing strategies and/or to invent completely new strategies. Here we sketch a model for how hypothesis testing strategies can themselves be generated and evaluated and discuss its implications for existing models of hypothesis testing.

We explore whether people can recognise the epistemic goal of active learners. In a novel online experiment, 110 adults watched screen recordings of other adults (``players'') manipulating objects in a 2D simulated physical microworld. Players had the goal of either identifying the magnet-like force connecting two of the objects, or their relative masses. Observers were asked to identify the learning goal of the player. By drawing from a previously collected dataset of active physical learning interactions and an ideal observer analysis, we manipulated how informative the players' actions are about the target property, and observers' level of access to the players' micro-control actions. We found observers were better at identifying the goals of successful players and of players trying to identify force, while the micro-dynamic evidence improved accuracy on identifying the mass goal. We use mixed methods to explore what cues observers used to make these judgments.

Many of the real world phenomena that cognizers must grapple with are continuous, not only in the values they can take, but also in how these values change over time. The mind must somehow abstract from these inputs to extract useful discrete concepts such as objects, events and causal relationships. We investigate several factors that affect basic inferences about causal relationships between continuous variables based on observations in continuous time. In a novel experiment, we explore the ways in which causal judgments are sensitive to factors that relate to causal inductive biases (e.g. causal lags, the direction of variation) and causal perception (e.g. the range and rapidity of variation). We argue standard statistical time-series models have limited utility in accounting for human sensitivity to these factors. We suggest further work is needed to fully understand the cognitive processes that underlie causal induction from time-series information.

A number of recent studies have used ideal observer models to capture human physical reasoning as based on approximatemental simulation driven through a realistic inner physics engine. While these approaches can match human competencein specific tasks, they are still relatively far from cognitive plausibility and are limited in their ability to capture patternsof human biases and errors. In this work, we train a recurrent neural network (RNN) extensively on a physical reasoningtask – conceptually mimicking the lifetime of experience that human adults have to build physical competence. We thenexamine its behavior alongside that of adults in the same test set of problems. We find that the RNN matches humanpatterns of judgments and errors much better than the idealised simulation account. We highlight specific situations whereboth RNN and humans erred and discuss the ramifications for current debates about the prevalence of physical simulationin cognition.

How do people use temporal information to make causal judg-ments? A number of studies have investigated the role of timein inferring generative causal structure, while few have exam-ined prevention. Here, we focus on a challenging task in whichparticipants learn the structure of several causal “devices” bywatching the devices’ patterns of activation over time. Eachdevice potentially includes both generative (producing an acti-vation of its effect) and preventative (blocking any effect acti-vations within a short time window) causal relationships. Weexamine judgment patterns through the lens of a normativemodel which incorporates actual causation with considerationsof prevention. We contrast this with a more computationallytractable feature-based approximation. Participants’ perfor-mance was substantially above chance in all conditions. Themajority of participants’ causal judgments were best fit by thefeature-based approximation based on delay and count heuris-tic cues.

Human learning and generalization benefit from bootstrapping: we arrive at complex concepts by starting small and building upon past successes. In this paper, we examine a computational account of causal conceptual bootstrapping, and describe a novel experiment in which the sequence of training data results in a dramatic order effect: participants succeed in identifying a compound concept only after experiencing training data in a “helpful” order. Our computational model represents causal relations as reusable, modular programs, which can themselves be “chunked” and flexibly reused to tackle more complex tasks. Our specific approach is based in combinatory logic and adaptor grammars, building on previous theories that posit a “language of thought” for concept representation, but making the learning process more sensitive to a learner’s experiences than any particular choice of conceptual primitives. Crucially, we demonstrate that a caching mechanism like that used in adaptor grammars is key to explain human-like bootstrapping patterns in causal generalization.

Activities such as watching a sports match and reading a novel often provoke suspense and surprise (S&S). Computation-ally, we hypothesize that these feelings derive from the dynamics of our beliefs. In our experiment, participants watch realvideotaped volleyball games or play a card game, where their belief dynamics (e.g. chance of winning) can be affected byboth the stimuli and background information (e.g. game rules and prior beliefs about the teams / the card deck). FollowingEly et al (2015) we formalize instantaneous suspense as a function of expected variance in future belief, and surprise asrelated to the magnitude of belief changes. Through probabilistic model we generate point-by-point predictions of S&S.We find that ratings of S&S for the same games depend on experimentally manipulated in qualitative agreement with ourmodel, but we also identify several situations where the model fails.