Understanding how humans make maladaptive choices requires identifying disruptions in the neural circuits and cognitive computations that guide adaptive decision-making. My research program develops and tests computational theories of decision-making by combining cognitive modeling with multimodal neuroimaging — from functional MRI to direct neurochemical recordings in humans — linking behavior to neural representations and circuit-level dynamics. My research spans four connected themes.
Leading theories predict that decisions between higher-value rewards should be noisier and more error-prone. Using diffusion decision models and a paradigm built on objective reward magnitudes, I found the opposite: higher value makes options more distinguishable, a robust property across snack foods, abstract art, and learned stimuli. Follow-up modeling showed that attention, not perceptual noise, drives this enhancement. A novel fMRI + eye-tracking paradigm then dissociated brain regions that encode currently sampled evidence (vmPFC, ventral striatum) from those that integrate it over time (pre-SMA), providing direct neural evidence for how attention-weighted value signals are transformed into committed choices.
Key work: Shevlin et al. (2022, PNAS) · Shevlin & Krajbich (2021, J. Math. Psych.) · Shevlin et al. (2025, eLife).
Connecting computational models to the neuromodulatory systems that instantiate them, my postdoctoral work uses machine-learning-enhanced fast-scan cyclic voltammetry to measure dopamine and serotonin at sub-second resolution in patients receiving deep brain stimulation for treatment-resistant depression. Task context shaped these responses — dopamine rose during reversal learning, serotonin during social exchange — and neurochemical responses during the social task uniquely predicted mood improvement and remission at six months. I am extending this framework to Parkinson's disease and, through cross-species collaborations, using a mouse model of autism to bridge genetic and circuit alterations to behavioral computation.
Key work: Shevlin, Fu, et al. (under revision, Nature) · Kato, Shevlin, et al. (in prep) · Jahan, Shevlin, et al. (in prep).
Compulsive behaviors, from binge eating to substance use, offer a window into disrupted valuation and choice. In bulimia nervosa, I showed that negative affect abnormally delays the onset of health-attribute processing, extending the window in which taste drives choice, with the bias scaling with symptom severity. This motivates my hypothesis that compulsive disorders involve imbalanced reward discriminability: heightened sensitivity to disorder-specific rewards but impaired sensitivity to alternatives. Preliminary data show reduced monetary reward discrimination in binge-eating and cannabis-use groups, correlated with symptom severity. These data serve as the foundation of on-going grant submissions. This framework also extends to social decision-making: modeling Ultimatum Game choices with a diffusion decision model, I find that borderline personality disorder involves diminished sensitivity to unfairness and faster, less deliberative responding, a computational signature reflecting altered social valuation.
Key work: Shevlin et al. (2026, eLife) · Shevlin, Le Roux, et al. (in prep).
Rigorous computational science depends on trustworthy data and on methods that extract maximal insight from it. One strand of this work shows that cognitive models can recover meaningful parameters even from sparse, one-shot decisions, widening where these tools can be applied. A second strand confronts a growing threat to behavioral research: large language models now generate synthetic survey responses that evade standard quality checks and can contaminate even clinical mental-health screening. Building on this, I am developing tools to detect exposure to bots and synthetic respondents by combining visual gaze patterns with embedded survey and task checks, aiming to safeguard the integrity of online data collection.
Key work: Shevlin et al. (A little goes a long way, PsyArXiv) · Fernandez, Berner & Shevlin (synthetic respondents, PsyArXiv).