Whereas cognitive models of learning often assume direct experience with both the features of an event and with a true label or outcome, much of everyday learning arises from hearing the opinions of others, without direct access to either the experience or the ground truth outcome. We consider how people can learn which opinions to trust in such scenarios by extending the hedge algorithm: a classic solution for learning from diverse information sources. We first introduce a semi-supervised variant we call the delusional hedge capable of learning from both supervised and unsupervised experiences. In two experiments, we examine the alignment between human judgments and predictions from the standard hedge, the delusional hedge, and a heuristic baseline model. Results indicate that humans effectively incorporate both labeled and unlabeled information in a manner consistent with the delusional hedge algorithm---suggesting that human learners not only gauge the accuracy of information sources but also their consistency with other reliable sources. The findings advance our understanding of human learning from diverse opinions, with implications for the development of algorithms that better capture how people learn to weigh conflicting information sources.

Humans rely on our social networks to make more accurate inferences about the world. Yet it remains unclear how those inferences are shaped by the medium through which information is exchanged and beliefs are shared. In this paper, we report two experiments where participants (N=645) were asked to make inferences about an unknown probability distribution based on limited private observations. They exchanged messages with neighbors in a small social network and were asked to update their beliefs over repeated rounds. We compared three conditions: a unidirectional message medium, a constrained slider medium, and an interactive chat. All groups were able to converge toward more accurate inferences, but their convergence rates varied across conditions in ways not well-captured by common models. We argue that computational models of collective behavior must move beyond the assumption of direct belief transmission to capture the complexities of sharing information through natural language.

Fractions are a critical building block for the development ofhuman mathematical cognition, but the origins of this conceptare not well-understood. Recent work has found that a wholenumber sense is present in deep convolutional neural networks(DCNNs) pre-trained for object recognition and uses them asa model for investigating human numerical cognition. Do DC-NNs also have a fraction sense? If so, is it dependent or in-dependent of whole number processing? We investigated theneural sensitivity of a pretrained DCNN to both whole num-bers and fractions. We replicated and extended previous re-search that the sense of whole number emerges in a differentDCNN architecture. Further, we showed that DCNN is alsosensitive to fraction value, i.e., the ratio of numerosities. Test-ing this model, our results suggest that the fraction sense relieson the whole number sense.

Successful teaching requires an assumption of how the learner learns - how the learner uses experiences from the world to update their internal states. We investigate what expectations people have about a learner with a behavioral experiment: Hu- man teachers were asked to teach a sequential decision-making task to an artificial dog in an online manner using rewards and punishments. The artificial dogs were implemented with either an Action Signaling agent or a Q-learner with different dis- count factors. Our findings are threefold: First, we used ma- chine teaching to prove that the optimal teaching complexity across all the learners is the same, and thus the differences in human performance was solely due to the discrepancy between human teacher’s theory of mind and the actual student model. Second, we found that Q-learners with small discount factors were easier to teach than action signaling agents, challenging the established conclusion from prior work. Third, we showed that the efficiency of teaching was monotonically increasing as the discount factors decreased, suggesting that humans’ theory of mind bias towards myopic learners.

This paper considers how interactions with AI algorithms can boost human creative thought. We employ a psychological task that demonstrates limits on human creativity, namely semantic feature generation: given a concept name, respondents must list as many of its features as possible. Human participants typically produce only a fraction of the features they know before getting “stuck.” In experiments with humans and with a large language model (GPT-4), we contrast behavior in the standard task versus a variant in which participants can ask for algorithmically-generated hints. Algorithm choice is administered by a multi-armed bandit whose reward indicates whether the hint helped generating more features. Humans and the AI show similar benefits from hints, and remarkably, bandits learning from AI responses prefer the same prompting strategy as those learning from human behavior. The results suggest that strategies for boosting human creativity via computer interactions can be learned by bandits run on groups of simulated participants.

Human groups are able to converge to more accurate beliefs through deliberation, even in the presence of polarization and partisan bias --- a phenomenon known as the ``wisdom of partisan crowds.'' Large Language Models (LLMs) are increasingly being used to simulate human collective behavior, yet few benchmarks exist for evaluating their dynamics against the behavior of human groups. In this paper, we examine the extent to which the wisdom of partisan crowds emerges in groups of LLM-based agents that are prompted to role-play as partisan personas (e.g., Democrat or Republican). We find that they not only display human-like partisan biases, but also converge to more accurate beliefs through deliberation, as humans do. We then identify several factors that interfere with convergence, including the use of chain-of-thought prompting and lack of details in personas. Conversely, fine-tuning on human data appears to enhance convergence. These findings show the potential and limitations of LLM-based agents as a model of human collective intelligence.

Accurately simulating human opinion dynamics is crucial for understanding a variety of societal phenomena, including polarization and the spread of misinformation. However, the agent-based models (ABMs) commonly used for such simulations often over-simplify human behavior. We propose a new approach to simulating opinion dynamics based on populations of Large Language Models (LLMs). Our findings reveal a strong inherent bias in LLM agents towards producing accurate information, leading simulated agents to consensus in line with scientific reality. This bias limits their utility for understanding resistance to consensus views on issues like climate change. After inducing confirmation bias through prompt engineering, however, we observed opinion fragmentation in line with existing agent-based modeling and opinion dynamics research. These insights highlight the promise and limitations of LLM agents in this domain and suggest a path forward: refining LLMs with real-world discourse to better simulate the evolution of human beliefs.