About

Yael Niv is a Professor in the Department of Psychology at Princeton University and serves as the Director of Graduate Studies. She is affiliated with the Princeton Neuroscience Institute. Her research focuses on the neural and computational processes underlying reinforcement learning and decision-making. Her lab studies how animals and humans learn from trial and error to predict future events and act to maximize reward and minimize punishment, emphasizing the interaction of attention and memory processes with reinforcement learning to create representations that facilitate efficient learning of new tasks. Niv's work employs model-based experimentation using computational models to formulate precise hypotheses, design experiments, and analyze results. She aims to provide normative explanations of behavior through models that offer a principled understanding of the computational algorithms used by brain mechanisms and their potential optimality. A recent focus of her lab is computational cognitive neuropsychiatry, where her team uses computational tools to better diagnose, understand, and treat psychiatric illnesses such as depression, OCD, schizophrenia, and addiction. Her research is conducted under the auspices of the Rutgers-Princeton Center for Computational Cognitive Neuropsychiatry.

Research topics

• Computer Science
• Cognitive psychology
• Psychology
• Neuroscience
• Cognitive science
• Machine Learning
• Artificial Intelligence
• Biology
• Knowledge management
• Social psychology

Selected publications

• Do Large Language Models Mentalize When They Teach?

  ArXiv.org · 2026-04-02

  articleOpen access

  How do LLMs decide what to teach next: by reasoning about a learner's knowledge, or by using simpler rules of thumb? We test this in a controlled task previously used to study human teaching strategies. On each trial, a teacher LLM sees a hypothetical learner's trajectory through a reward-annotated directed graph and must reveal a single edge so the learner would choose a better path if they replanned. We run a range of LLMs as simulated teachers and fit their trial-by-trial choices with the same cognitive models used for humans: a Bayes-Optimal teacher that infers which transitions the learner is missing (inverse planning), weaker Bayesian variants, heuristic baselines (e.g., reward based), and non-mentalizing utility models. In a baseline experiment matched to the stimuli presented to human subjects, most LLMs perform well, show little change in strategy over trials, and their graph-by-graph performance is similar to that of humans. Model comparison (BIC) shows that Bayes-Optimal teaching best explains most models' choices. When given a scaffolding intervention, models follow auxiliary inference- or reward-focused prompts, but these scaffolds do not reliably improve later teaching on heuristic-incongruent test graphs and can sometimes reduce performance. Overall, cognitive model fits provide insight into LLM tutoring policies and show that prompt compliance does not guarantee better teaching decisions.

  PublisherOA PDF

• Do Large Language Models Mentalize When They Teach?

  arXiv (Cornell University) · 2026-04-02

  preprintOpen access

  How do LLMs decide what to teach next: by reasoning about a learner's knowledge, or by using simpler rules of thumb? We test this in a controlled task previously used to study human teaching strategies. On each trial, a teacher LLM sees a hypothetical learner's trajectory through a reward-annotated directed graph and must reveal a single edge so the learner would choose a better path if they replanned. We run a range of LLMs as simulated teachers and fit their trial-by-trial choices with the same cognitive models used for humans: a Bayes-Optimal teacher that infers which transitions the learner is missing (inverse planning), weaker Bayesian variants, heuristic baselines (e.g., reward based), and non-mentalizing utility models. In a baseline experiment matched to the stimuli presented to human subjects, most LLMs perform well, show little change in strategy over trials, and their graph-by-graph performance is similar to that of humans. Model comparison (BIC) shows that Bayes-Optimal teaching best explains most models' choices. When given a scaffolding intervention, models follow auxiliary inference- or reward-focused prompts, but these scaffolds do not reliably improve later teaching on heuristic-incongruent test graphs and can sometimes reduce performance. Overall, cognitive model fits provide insight into LLM tutoring policies and show that prompt compliance does not guarantee better teaching decisions.

  PublisherDOI

• Protocol for a randomized trial to predict the efficacy of cognitive and behavioral interventions for symptoms of depression

  Frontiers in Psychiatry · 2026-04-16

  articleOpen access

  Introduction: Cognitive behavioral therapy (CBT) is one of the most common interventions for depression and has two key components: Cognitive Restructuring (CR) and Behavioral Activation (BA). However, no evidence-based guidelines exist to help clients and clinicians decide whether CBT would be a good first-line treatment for a given individual based on their personal characteristics, and which CBT intervention would benefit them more. We propose that specific capacities to learn from new information and experiences are prerequisites for response to CBT and that BA and CR require different learning capacities. In this study, we aim to develop predictive models of symptom change based on computationally-derived variables from behavioral tasks, in addition to clinical and demographic self-report data, to identify parameters and variables that can determine which individuals with depressive symptoms would benefit from CBT-based interventions and, ideally, which specific interventions they would benefit from more. Methods and analysis: We plan to recruit at least 1,500 adult participants who report having symptoms of depression and reside in U.S. After completing a series of questionnaires and behavioral tasks to assess their learning propensities, participants will be randomly assigned to a BA or a CR group. Using an online self-help tool, participants will then engage with designated modules according to their assigned group for five weeks. We will assess symptoms 1 week post-intervention (main end point of study) and follow up at 6, 18, and 42 weeks post-intervention. Upon enrolling and consenting into the main study, participants will be randomly assigned to either the training dataset or the held-out test dataset at a ratio of 2:1. This enables a clean separation of training and test datasets and prevent data leakage. We plan to build cross-validated predictive algorithms on the training dataset, and preregister our analysis plan before we validate our models and hypotheses in the held-out, unseen, test dataset. Enrollment of the study started 23rd January, 2024. Study protocol registration: ClinicalTrials.gov, identifier (NCT06631183). The protocol follows the Standard Protocol Items: Recommendations for Interventional Trials (SPIRIT) guidelines. Numbers in brackets follow subsection numbers in the guidelines.

  PublisherOA PDFDOI

• Testing the longitudinal and causal effect of social media rewards on mood and mental health in young adults

  2026-03-09

  articleOpen accessSenior author

  As social media platforms have increasingly pervaded our lives, societal concern about their potential detrimental effects on users’ mental health has surged. However, mechanistic evidence for this impact is still lacking. Recent work has identified a link between depression and heightened sensitivity to social media rewards. These rewards (i.e. likes, comments, followers, direct messages) are known to impact mood and affective states in the lab, yet little is known about their effect in daily-life situations. Moreover, it remains unclear how these rewards impact different aspects of social media behavior, such as logging onto the app vs. posting. Here, we take advantage of detailed behavioral data – the timing of various behaviors and social rewards on Instagram – to study the effects of social rewards on behavior as well as mood and mental health. Instagram posters (target N = 400 for this initial study) will complete five weekly surveys about their mood and mental health, as well as share their Instagram data for 3 months. Using both direct statistical analyses and computational modeling, we will quantify the reinforcing effects of social rewards on posting as well as the ‘approach’ motivation effect of these rewards on drawing users onto the app, enhanced by notifications. We will test whether these different forms of social-reward sensitivity 1) are associated with symptoms of depression and anxiety and 2) moderate the impact of social rewards on mood and symptoms. In a randomized controlled design, we will ask a random subset of the participants to turn off their notifications, and test 3) whether turning off notifications leads to fewer log-ins, less time spent on Instagram, and better mood and mental health. Our study will comprehensively test the effect of social rewards on Instagram behavior and users’ mental health using both observational and experimental approaches. We will also develop a novel model of social media behavior, with a temporal resolution that has not been feasible in prior studies. Based on our results, future studies will be able to investigate other user populations (e.g., adolescents) to determine the specificity and variability of the effects of social media rewards on Instagram users.

  PublisherDOI

• Improving the reliability of the Pavlovian go/no-go task for computational psychiatry research

  2025-11-06

  preprintOpen accessSenior author

  Background: The Pavlovian go/no-go task is commonly used to measure individual differences in Pavlovian biases and their interaction with instrumental learning. However, prior research has found suboptimal reliability for computational model-based performance measures for this task, limiting its usefulness in individual-differences research. These studies did not make use of several strategies previously shown to enhance task-measure reliability (e.g., task gamification, hierarchical Bayesian modeling for model estimation). Here we investigated if such approaches could improve the task’s reliability. Methods: Across two experiments, we recruited two independent samples of adult participants (N=103, N=110) to complete a novel, gamified version of the Pavlovian go/no-go task multiple times over several weeks. We used hierarchical Bayesian modeling to derive reinforcement learning model-based indices of participants' task performance, and additionally to estimate the reliability of these measures. Results: In Experiment 1, we observed considerable and unexpected practice effects, with most participants reaching near-ceiling levels of performance with repeat testing. Consequently, the test-retest reliability of some model parameters was unacceptable (range: 0.379–0.973). In Experiment 2, participants completed a modified version of the task designed to lessen these practice effects. We observed greatly reduced practice effects and improved estimates of the test-retest reliability (range: 0.696–0.989). Conclusion: The results demonstrate that model-based measures of performance on the Pavlovian go/no-go task can reach levels of reliability sufficient for use in individual- differences research. However, additional investigation is necessary to validate the modified version of the task in other populations and settings.

  PublisherDOI

• Cognitive modeling of real-world behavior for understanding mental health

  Trends in Cognitive Sciences · 2025-09-27

  reviewSenior author
  PublisherDOI

• Schemas, reinforcement learning, and the medial prefrontal cortex

  2025-07-06 · 1 citations

  preprintOpen accessSenior author

  Schemas are rich and complex knowledge structures about the typical unfolding of events in a context. For example, a schema of a dinner at a restaurant. In this Perspective, we suggest that reinforcement learning (RL), a computational theory of learning the structure of the world and relevant goal-oriented behavior, underlies schema learning. We synthesize literature about schemas and RL to offer that three RL principles might govern the learning of schemas: learning via prediction errors, constructing hierarchical knowledge using hierarchical RL and dimensionality reduction through learning a simplified and abstract representation of the world. We then suggest that the orbito-medial prefrontal cortex is involved in both schemas and RL due to its involvement in dimensionality reduction and in guiding memory reactivation through interactions with posterior brain regions. Last, we hypothesize that the amount of dimensionality reduction might underlie gradients of involvement along the ventral-dorsal and posterior-anterior axes of the orbito-medial prefrontal cortex. More specific and detailed representations might engage the ventral and posterior parts, whereas abstraction might shift representations toward the dorsal and anterior parts of the medial prefrontal cortex.

  PublisherDOI

• Cognitive Modeling of Real-World Behavior for Understanding Mental Health

  2025-03-17 · 1 citations

  preprintOpen accessSenior author

  A core strength of computational psychiatry is its focus on theory-driven research, in which cognitive processes are precisely quantified using computational models that formalize specific theoretical mechanisms. However, the data used in these studies often come from traditional lab-based cognitive tasks, which have unclear ecological validity. Here, we argue that the same theoretical frameworks and computational models can be applied to real-world data such as experience sampling, passive and/or digital-behavior data (e.g. online activity such as on social media). In turn, modeling real-world data can benefit from a theory-driven computational approach to move from purely predictive to explanatory power. We illustrate these points using emerging studies and discuss challenges and opportunities of using real-world data in computational psychiatry.

  PublisherOA PDFDOI

• Anxiety modulates event segmentation

  2025-11-20 · 1 citations

  articleSenior author

  Anxiety is one of the most prevalent mental health concerns. Current theories suggest that anxiety may arise due to deficits in segmentation of continuous experience into discrete context representations (‘event segmentation’), which leads to overgeneralization of fear across contexts, or conversely, overly rigid segmentation that prevents safety learning. Here, in two segmentation tasks (N=1109), we found novel and direct evidence that anxiety is associated with changes in event segmentation. Individuals with higher anxiety symptoms responded more slowly to transitions between events (event boundaries). They also segmented movies into discrete events more typically and more hierarchically, two hallmarks of precise segmentation. This precision was linked to self-reporting fewer context changes in daily life, suggesting individuals with anxiety prefer stable and predictable environments. These findings challenge overgeneralization theories of anxiety, revealing instead that individuals with anxiety exhibit precise, and potentially overly rigid segmentation. Such segmentation could maintain fear by preventing generalization from safe to fearful contexts, which has important implications for interventions like exposure therapy.

  PublisherDOI

• Understanding goal selection through human limitations

  2025-10-26

  articleOpen accessSenior author

  Human goal selection is simultaneously flexible and structured. Existing accounts explain thisdichotomy in terms of a biological reward function that may have proved adaptive on evolutionarytimescales. However, such explanations struggle to capture the human ability to quickly and flexiblyreevaluate goals and stimuli in response to changing contexts. In this article, we consider howhuman goal selection may be shaped by cognitive limitations. We structure this perspective aroundthree computational-style problems that people face when selecting their own goals: consideration,evaluation, and feasibility. Overall, we offer a brief survey of relevant literature alongside a resource-rational account of structured goal selection without appealing to intrinsic rewards.

  PublisherDOI

Recent grants

• Quantifying the latent-cause inference process in humans

  NIH · $444k · 2019–2022

• Orbitofrontal cortex as a cognitive map of task states

  NIH · $1.8M · 2016–2022

• Neural and computational mechanisms of selective attention in decision making

  NIH · $1.1M · 2012–2017

• fMRI investigations of how we learn what is relevant for a decision

  NIH · $242k · 2011–2013

Frequent coauthors

• Daniel Bennett

  Monash University

  43 shared

• Peter Dayan

  Max Planck Institute for Biological Cybernetics

  39 shared

• Geoffrey Schoenbaum

  National Institute on Drug Abuse

  39 shared

• Nathaniel D. Daw

  Princeton University

  29 shared

• Samuel J. Gershman

  Harvard University

  26 shared

• Nicolas W. Schuck

  Max Planck Institute for Human Development

  26 shared

• Angela Radulescu

  University of Minnesota Medical Center

  24 shared

• Zohar Levi

  Tel Aviv University

  24 shared

Labs

• Niv LabPI

Awards & honors

• Graduate Mentoring Award