Research projects
Computational phenotyping of social learning
Precision psychiatry needs reliable person-level measures of latent cognitive processes. Clinical interviews and questionnaires are stable in time but provide little mechanistic knowledge; current computational modeling estimates mechanisms with single-task fitting. Resulting parameters are known to be unstable across measurements, and not comparable across tasks and environments. No validated method currently supports joint cross-task estimation of interpretable latent processes at the person level. This project develops that method, with the intent to discover computational phenotypes (i.e., stable, interpretable cognitive profiles). The project involves model discovery, cross-task testing, and deep learning computational modeling.
Neurocomputational mechanisms of social learning
I lead with Caroline Charpentier the fMRI implementation of the Social Learning Battery, a multi-task project designed to characterize how people learn from direct experience, social information, and changing social partners. The battery combines tasks of risk and ambiguity, trust learning, and observational learning.
My role spans project management, fMRI acquisition, RA supervision, computational infrastructure, and computational modeling - including experimental design, BOLD event simulation, preprocessing and quality control, model specification, and DevOps. I will also lead the trust-learning analyses within this project.
Dynamics-RSG
Co-developed dynamics-rsg, a project comparing predictive coding and RFLO with backpropagation through time (BPTT) in leaky-tanh RNNs trained on the Ready-Set-Go task (Neurogym). We assessed how closely model behavior and neural dynamics aligned with macaque timing behavior and dorsomedial frontal cortex recordings (DANDI:000130). Developed and presented for Neuromatch NeuroAI.
Individual differences in dynamic belief updating during trust learning
There is heterogeneity in social learning. Focusing on how people form and revise trust when repeated interactions involve uncertainty, changing reciprocity, and possible betrayal, I develop a computational approach to studying this heterogeneity.
Different strategies can guide trust behavior, including simple heuristics, reinforcement learning, and belief updating. In a dynamic trust-learning task, I use computational modeling to identify profiles of participants who rely on distinct cognitive strategies. With Caroline Charpentier and Selin Topel, we cluster profiles and link this heterogeneity to psychiatric measures of social anxiety and ASD.
Dynamics of Social Networks - Understanding Belief Formation, Polarization, and Consensus Emergence
Building on my ReceivingNews and SendingNews projects, I developed the framework for studying how beliefs evolve in social networks, with Jean-Claude Dreher, Frédéric Moisan, and Alain Barrat. The project is now being conducted by Gaël Carniel. The design extends my earlier work on information evaluation and information sharing to a collective setting, asking asks how individual belief updating and network-level information propagation shape consensus, dissensus, and the apparition of polarization.
SendingNews - Neurocomputational mechanisms of inferring others' preferences for information
I developed and led SendingNews, a model-based fMRI project testing how people infer whether another person wants clarifying information about ambiguous news. The task separates first-order beliefs about news truthfulness from second-order beliefs about a receiver’s informational preferences.
With Jean-Claude Dreher and Marie Claire Villeval, we show that Senders integrate multiple belief sources through a Bayesian updating rule. They combine their own uncertainty about the news, social-distance cues about the receiver, and feedback about receivers’ general receptiveness to information. Model-based fMRI links these computations to partially separable neural signals, with valuation-related regions tracking beliefs about news truthfulness, social-cognitive regions tracking receiver-related beliefs, and frontopolar cortex tracking an integrated estimate of the receiver’s preference.
ReceivingNews - Metacognition biases information seeking in assessing ambiguous news
I developed and led ReceivingNews, a behavioral project testing how metacognitive confidence shapes information seeking after people evaluate ambiguous news. Participants made incentivized judgments about whether news items were true or false, reported confidence in those judgments, and indicated how much they were willing to pay to receive or avoid additional information.
With Jean-Claude Dreher and Marie Claire Villeval, we show that confidence was not calibrated to accuracy in this setting, but still guided information seeking. Participants with lower confidence were more likely to seek additional information, while those with higher confidence were more willing to avoid it. News imprecision and polarization increased the likelihood of misjudgment, and structural equation modeling showed that the demand for disambiguating information became less effective as it was driven by uncalibrated metacognition.
Testosterone Causes Decoupling of Orbitofrontal Cortex-Amygdala Relationship While Anticipating Primary and Secondary Rewards
I conducted reward-network and functional-connectivity analyses in a pharmacological fMRI study testing how testosterone affects reward anticipation. The task compared primary erotic rewards and secondary monetary rewards, allowing us to ask whether testosterone changes reward-specific valuation or broader prefrontal-limbic coordination.
The analyses showed that testosterone increased incentive behavior for erotic relative to monetary rewards and altered reward-related neural responses. Most importantly, testosterone also reduced functional coupling between ventromedial prefrontal cortex and amygdala during reward anticipation, suggesting that testosterone modulates prefrontal-limbic coordination during motivated behavior rather than simply increasing reward responsiveness.
Reward proximity and impulsive action
During my research internship in JC Dreher’s Neuroeconomics, Reward and Decision-Making team, I contributed to David O’Connor’s behavioral study on the effects that spatial proximity to rewarding cues has on impulsive behavior. The project used virtual reality to compare responses to food-reward cues presented near or far from the participant. I contributed to data acquisition and behavioral modeling (DDM analyses of response-time and decision processes). The study showed that nearby rewards increased impulsive action, impairing participants’ ability to withhold motor responses, but did not increase impulsive choice in an intertemporal-choice task.
Methods, software, and lab infrastructure
Ten simple rules for achieving computational reproducibility in neuroscience
Gaurav D. Mahajan and I are formalizing the infrastructure we built for the UMD Social Learning and Decisions Lab into a methodological manuscript on computational reproducibility in neuroscience.
The paper argues that reproducibility depends on the continuity of the full computational pathway. We focus on low-cost practices that can be adopted early and scaled, such as project structure, version control, documentation, reusable code, automation, use of AI, lab culture and more.
SLD Lab computational infrastructure
I led the development of lab-level computational infrastructure for behavioral and fMRI studies. Built with implementation work from Gaurav D. Mahajan and with the support of Joshua Berman and Deanna Durben, we automated fMRI and behavioral stratified sampling, experiments logging, data extraction, data preprocessing, and quality control. Employs containers and nightly triggers.
Neurodesign-plus
In decision-making, fMRI experiments must optimize design efficiency while satisfying complex task structure constraints. I led the development of Neurodesign-plus, an open-source Python package for optimizing such fMRI designs. The package extends neurodesign (R. Poldrack’s lab) to support task structures where trial order, timing, or condition assignment can be randomized, probabilistic, and/or path-dependent. Atharv A. Umap implemented much of the core software under my supervision. I defined the required functionality, contributed code, guided the package architecture, and am responsible for tutorials, documentation, and preprint.
FitLINS-based fMRI analysis package
A Python-based fMRI analysis package that wraps FitLINS, Nilearn, and SLURM into simple workflows. The package runs GLM analyses from fMRIPrep outputs, manages model configuration, patches reports, generates statistical maps, builds PDF summaries, and indexes analyses. Supports array, parallel and interactive execution.
fMRIPrep QC guide and logging system
fMRI quality control (QC) is hardly automatable (see MRIQC and MRIQCEPTION on no-reference IQMs). To standardize fMRIPrep QC across lab members, I designed a human QC outputs evaluation process. We found fMRI quality control more teachable, auditable, and consistent across lab members (as evidenced by rating spread between RAs).
Cookiecutter Neuro Research Project
In experimental cognitive sciences, research projects involve multiple contributors, software environments, and preprocessing stages. This creates vulnerabilities that impede computational reproducibility. This Cookiecutter template creates structured, FAIR-compliant repositories for psyc/neuro.
SPM-based fMRI analysis templates
Standardized SPM-based templates. They support fMRI project setup, preprocessing, first- and second-level Informed Basis Set definition and contrasting, and MRIcroGL figures, with tutorials on fMRI design and preprocessing.
Personal and research tools, AI experiments, and technical projects
SUPLEX-agentic-workflow
The more complex the task to perform, the more LLMs tend to make assumptions and to drift away from initial task specifications. To reduce these risks, I developed SUPLEX, a workflow/control system for LLM-assisted work. The system consists of .md documentation and logging that layers on top of pre-existing agentic scaffolds. It structures AI-assisted coding around a) clear goals and bounded tasks, b) explicit task handoffs, c) execution reports, and d) a memory system. It preserves human supervision and reviewability.
AI-research-feedback
Claude and Codex skills to support AI-assisted review of manuscripts, grants, pre-analysis plans, and paper-code packages. Adapted from Claes Backman, economist; rewritten and extended for psychology and neuroscience research projects.
March Madness 2026 Forecasting Engine
To celebrate my first-ever March Madness, I built a forecasting engine for the 2026 NCAA Men/Women tournament. The system fetched tournament data, tested predictive models, generated bracket portfolios, and diagnosed prediction errors after each round. The result is still live on dedicated webpages. I proudly ranked 3/11 on both Men and Women friends bracket tournaments, with +EV on my bets.