A Transdiagnostic Space of Disorder-Like Phenotypes in Reinforcement Learning Agents

The paper presents a dose-controllable framework for inducing seven disorder-like phenotypes in appraisal-guided PPO agents: anxiety, mania, obsessive-compulsive checking, depression, impulsivity, addiction, and PTSD. Each phenotype is controlled by one appraisal-signal knob and evaluated with a preregistered assay. Across more than 1,000 runs using 10 seeds and four controls, the authors report graded monotone dose-response relationships that controls do not reproduce. They also describe an emergent two-dimensional affective space, selective remission after knob removal, exposure-based recovery for avoidance phenotypes, and nonadditive interactions between simultaneous knobs. Depression and addiction effects were further tested in MiniWorld with a standard convolutional agent.