Postdoctoral Fellow, Psychiatric University Hospital Zurich and Neuroscience Center Zurich (UZH & ETH Zurich)
I study how humans and artificial systems construct models of the world, how these models shape cognition and behavior, and how they differ across systems or become disrupted in psychiatric disorders.
My research combines behavioral experiments, computational modeling, neuroimaging, and analyses of large language models to investigate human cognition and artificial intelligence. I am currently a postdoctoral fellow in the Homan Lab. I received my PhD in Cognitive Neuroscience from the IDG/McGovern Institute for Brain Research at Beijing Normal University, with research experience at the Zurich Center for Neuroeconomics during my PhD.
People now ask chatbots why someone acted as they did. I compare large language models with human data and read their internal states with probing and steering. Models turn out to see less malice and more virtue than people do.
Papers →How does the brain build internal maps of knowledge, and what changes in psychosis? I study how relational knowledge is built from images and from language, and how it shows up in natural speech.
Papers →Why do people cooperate, punish and help? I use computational models and neuroimaging to separate the motives behind altruism, trust and cooperation, from childhood to adulthood.
Papers →† Co-first authors. * Co-last authors. Full list on Google Scholar.
Language models judge a whistleblower's motives more charitably than people do. Reading the model's internal states shows two separate sources: the signal for malevolent motives is weaker inside the model, and benevolent motives are amplified when the model reports its answer.
Large language models reproduce human moral evaluations of whistleblowers while attributing systematically different motives to them, so agreement in average judgments can conceal a different attributional structure.
Schizophrenia-spectrum cognitive disorganization selectively disrupts the ability to construct and use internal relational maps from visual input, but not from language, supporting the shallow cognitive map hypothesis.
Explaining knowledge to others externalizes spatially structured representations of conceptual memory in naturalistic speech. Cognitive disorganization selectively weakens this effect after image-based, but not language-based, learning.
Third-party punishment and helping are not driven by a single moral impulse. Behavioral experiments and computational modeling identify seven socioeconomic motives, the "motive cocktail", whose mix differs between people and contexts.
Children's third-party altruism is also a cocktail of dissociable social motives, which shape punishment and helping in response to inequality across contexts, gender and individuals.
Adolescents cooperate less than adults not because they understand others less well, but because reciprocity is less rewarding to them, reflecting a stronger self-interest motive.
In the Prisoner's Dilemma, cooperation is best explained by a model that combines the intrinsic reward of reciprocity with learning about what others expect.
Acute stress shifts third parties from punishing offenders to helping victims, driven by changes in connectivity between the amygdala and the dorsolateral and ventromedial prefrontal cortex.
Virtual reality makes it possible to study heroism safely. In simulated crime scenes, both the situation and personal attributes such as gender and empathy shape whether people step in.
Monetary and social incentives both increase cooperation, but people with stronger primary psychopathic traits respond less to social incentives because they discount what others expect of them.