Humans are remarkably adaptive instructors who can adjust advice on complex tasks based on a learner's current knowledge and goals. However, paradigms in cognitive psychology generally examine pedagogy in constrained and discrete tasks, like categorization or feature learning. We examine teaching in continuous domains, where there are theoretically infinite hypotheses, and model how teachers can formulate a computationally tractable Bayesian inference using Gaussian process regression. Taking inspiration from function learning tasks, we investigated how one teaches visual underlying functions by giving pedagogically-informed point examples. Preliminary evidence suggests teachers are sensitive to learners' priors about continuous functions. For instance, when learners expect a diverse range of function types (linear, quadratic, periodic, etc.) then teachers tend to select examples that help distinguish between those types. Conversely, teachers relaxed this constraint if learners had not seen multiple function types. Our results provide insight into mechanisms of pedagogical guidance in complex, continuous task domains.

We often rely on others' testimony when learning about new topics, such as health benefits of a novel food. However, the sources are not always knowledgeable, helpful, or unbiased, necessitating an assessment of their credibility. Here, we present a Bayesian model of source credibility, where a listener simultaneously infers the expertise and intention of the source while trying to discern the truth. A key prediction is that rational inference of credibility requires anchoring it on some kernel of shared knowledge. We consider a scenario where both parties have noisy access to the ground truth of familiar topics (e.g., is broccoli healthy?), which serves as a basis for reasoning about a source's credibility on novel topics (e.g., is avocado healthy?). This approach provides a computational framework for understanding how people respond to information in domains like science communication and media consumption.

Value alignment is essential for building AI systems that can safely and reliably interact with people. However, what a person values---and is even capable of valuing---depends on the concepts that they are currently using to understand and evaluate what happens in the world. The dependence of values on concepts means that concept alignment is a prerequisite for value alignment---agents need to align their representation of a situation with that of humans in order to successfully align their values. Here, we formally analyze the concept alignment problem in the inverse reinforcement learning setting, show how neglecting concept alignment can lead to systematic value mis-alignment, and describe an approach that helps minimize such failure modes by jointly reasoning about a person's concepts and values. Additionally, we report experimental results with human participants showing that humans reason about the concepts used by an agent when acting intentionally, in line with our joint reasoning model.

Speakers communicate to influence their partner's beliefs and shape their actions. Belief- and action-based objectives have been explored independently in recent computational models, but it has been challenging to explicitly compare or integrate them. Indeed, we find that they are conflated in standard referential communication tasks. To distinguish these accounts, we introduce a new paradigm called signaling bandits, generalizing classic Lewis signaling games to a multi-armed bandit setting where all targets in the context have some relative value. We develop three speaker models: a belief-oriented speaker with a purely informative objective; an action-oriented speaker with an instrumental objective; and a combined speaker which integrates the two by inducing listener beliefs that generally lead to desirable actions. We then present a series of simulations demonstrating that grounding production choices in future listener actions results in relevance effects and flexible uses of nonliteral language. More broadly, our findings suggest that language games based on richer decision problems are a promising avenue for insight into rational communication.

Humans can flexibly bind familiar functional roles to novel entities in their environment. For example, children who have the concept of ``goal posts'' can bind this abstract role to two hats placed on the street. In doing so, they can port over existing expectations of ``goal posts" for the duration of the game. In this paper, we seek to explore artificial agents' ability to perform flexible role binding and rebinding. To this end, we designed a Gridworld navigation game and tested a popular CNN-based agent which has had success in other tasks involving visual and spatial state spaces (e.g. Atari or Minigrid). To our surprise, we found that while this architecture was capable of overfitting to the training set, it was not able to learn flexible role binding without intervention. We ultimately show that with carefully engineered data augmentation techniques, our artificial agent is able to learn the task. This suggests that the diversity of the training dataset was a limiting factor.

Speech-acts can have literal meaning as well as pragmaticmeaning, but these both involve consequences typically in-tended by a speaker. Speech-acts can also have unintentionalmeaning, in which what is conveyed goes above and beyondwhat was intended. Here, we present a Bayesian analysis ofhow, to a listener, the meaning of an utterance can significantlydiffer from a speaker’s intended meaning. Our model em-phasizes how comprehending the intentional and unintentionalmeaning of speech-acts requires listeners to engage in sophisti-cated model-based perspective-taking and reasoning about thehistory of the state of the world, each other’s actions, and eachother’s observations. To test our model, we have human partic-ipants make judgments about vignettes where speakers makeutterances that could be interpreted as intentional insults or un-intentional faux pas. In elucidating the mechanics of speech-acts with unintentional meanings, our account provides insightinto how communication both functions and malfunctions.

When faced with a complex problem, people naturally break it up into several simpler problems. This hierarchical decom-position of an ultimate goal into sub-goals facilitates planning by reducing the number of factors that must be consideredat one time. However, it can also lead to suboptimal decision-making, obscuring opportunities to make progress towardsmultiple subgoals with a single action. Is it possible to take advantage of the hierarchical structure of problems withoutsacrificing opportunities to kill two birds with one stone? We propose that people are able to do this by representing andpursuing multiple subgoals at once. We present a formal model of planning with compositional goals, and show that itexplains human behavior better than the standard ”one-at-a-time” subgoal model as well as non-hierarchical limited-depthsearch models. Our results suggest that people are capable of representing and pursuing multiple subgoals at once; how-ever, there are limitations on how many subgoals one can pursue concurrently. We find that these limitations vary byindividual.

Successful teaching entails a complex interaction between ateacher and a learner. The teacher must select and convey in-formation based on what they think the learner perceives andbelieves. Teaching always involves misaligned beliefs, butstudies of pedagogy often focus on situations where teachersand learners share perceptions. Nonetheless, a teacher andlearner may not always experience or attend to the same as-pects of the environment. Here, we study how misaligned per-ceptions influence communication. We hypothesize that theefficacy of different forms of communication depends on theshared perceptual state between teacher and learner. We de-velop a cooperative teaching game to test whether concretemediums (demonstrations, or “showing”) are more robust thanabstract ones (language, or “telling”) when the teacher andlearner are not perceptually aligned. We find evidence that (1)language-based teaching is more affected by perceptual mis-alignment, but (2) demonstration-based teaching is less likelyto convey nuanced information. We discuss implications forhuman pedagogy and machine learning.

Humans are adept at using punishments to influence and modify the behavior of others. Current approaches model pun-ishment as a direct, immediate imposition of cost. In contrast, our research suggests that people interpret punishment as acommunicative act. We show that people expect costless, yet communicative, punishments to be as effective as cost impos-ing punishment (Experiment 1). Under some situations, people display a systematic preference for costless punishmentsover more canonical, cost imposing punishments (Experiment 2). People readily seek out and infer the communicativemessage inherent in a punishment (Experiment 3). And, people expect that learning from punishment depends on theease with which its communicative intent can be inferred (Experiment 4). Taken together, these findings demonstrate thatpeople expect punishment to be generated and interpreted as a communicative act.

Teaching plays a crucial role in human learning, from formal educational environments to mentorship scenarios, yet its cognitive underpinnings remain underexplored. We focus on the distinction between teaching by reasoning using Theory of Mind (i.e., explicitly inferring what a learner knows) and teaching using heuristics (i.e., relying on a simple rule). We use a graph-navigation task where a learner agent with limited knowledge attempts to navigate through the most rewarding trajectory, with guidance from a human teacher. Our findings reveal that teachers utilize a blend of learner-specific strategies and general heuristics. We model learner-specific strategies using Bayesian Theory of Mind (Baker, Saxe, & Tenenbaum, 2009) and demonstrate that the most effective teachers incorporate this strategy. Intriguingly, we show that teaching strategies can be altered without explicit feedback. This suggests that subtle changes in the environment may significantly alter teaching approaches, highlighting the importance of understanding the cognitive processes behind teaching.