Steering a Language Model's Latent Mental-Health Representation Modulates Affect and Working Memory in Human Conversation Partners
Year: 2026
Type: preprint
Abstract: Prior work in this program identified a linear direction in an open-weight language model's residual stream, derived from probes trained on psychiatric constructs, that causally controls the model's outputs and behaviours. This preregistration specifies a confirmatory test of whether steering this latent mental-health representation during a human-AI text conversation transfers to the human conversation partner, shifting the participant's own subjective affect and modulating performance on an interleaved working-memory task.
Authors: Sankalp Garud, L. S. Morris
Language: English
Topic: Mental Health via Writing
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7171947093