SSR: Can Simulated Patients Learn to Stigmatize Themselves? Modeling Self-Stigma through Internal Monologue

Technical Summary: Simulated Patients Learning to Stigmatize Themselves via Self-Reflection (SSR)

Problem Statement
Current Large Language Model (LLM) simulations for mental health training often fail to capture the nuanced, context-sensitive reality of self-stigma. While existing models can simulate patients, they tend to exhibit either uniformly compliant or uniformly resistant behaviors. They lack the capacity to model self-stigma as a dynamic, internal cognitive process where patients internalize negative stereotypes, leading to situational resistance, avoidance, denial, or self-blame. This gap is exacerbated by three challenges: a lack of integration between psychological models of self-stigma and computational frameworks, the absence of datasets pairing patient utterances with their underlying internal reasoning, and the technical inability of current agents to dynamically modulate stigma expression based on conversational triggers.

Methodology
The authors propose a novel simulation framework grounded in Corrigan's 3A1H psychological model of self-stigmatization (Awareness, Agreement, Application, and Harm). The approach follows a three-stage pipeline:

1. Stigma Utterance Extraction: The authors operationalize self-stigma using five behavioral manifestations derived from clinical literature: Avoidance, Denial, Self-blame, Defensiveness, and Social Concern. Using GPT-5, they automatically annotate dialogue turns from existing mental health corpora (e.g., Alexander Street Transcripts, D4, Client Reaction, ESConv) for these manifestations, ensuring context-awareness by evaluating utterances alongside dialogue history. They also annotate associated life events and symptoms.
2. SSR Dataset Construction: The core innovation is the Stigmatized Self-Reflection (SSR) dataset. For every stigma-labeled utterance, the system generates a corresponding internal monologue that articulates the unspoken, stigma-driven thoughts preceding the utterance. These monologues are structured to reflect the 3A1H progression (e.g., "Everyone says depressed people are weak... I worry they might have a point...") but are expressed as fluent, narrative-style reasoning chains.
3. Training with Chain of Stigma: The authors frame stigma simulation as a sequential conditional generation task. Models are fine-tuned using a standard language modeling objective to produce two outputs in order: first, the internal "Chain of Stigma Thoughts," and second, the external patient response. This forces the model to learn the mapping from context to internal reasoning, and from that reasoning to an appropriate external verbal manifestation.

Key Contributions

• Theoretical Integration: The first computational framework to integrate the psychological 3A1H model to simulate the dynamic, stage-wise progression of self-stigma in patient agents.
• SSR Dataset: A novel resource pairing patient utterances with internal stigma monologues, created by augmenting multiple mental health dialogue corpora.
• Performance Validation: Demonstration that models fine-tuned with the SSR approach significantly outperform specialized baselines (such as UPSD4 and Patient-ψ) in generating authentic, context-sensitive, and expert-rated stigma-aware responses.

Results
Evaluations on held-out test sets (divided into Stigma and Neutral subsets) using Llama-3.1-8B and Qwen3-8B backbones yielded the following findings:

• Context Sensitivity (RQ1 & RQ2): SSR-trained models achieved higher Contextual Stigma F1 (CSF1) scores, particularly in the negative class (restraint in neutral contexts). Unlike baselines that often over-produce stigma regardless of context, SSR models learned to dynamically adjust their expression, showing high sensitivity to triggers while maintaining appropriate behavior in neutral conversations.
• Clinical Authenticity (RQ3):
  • Distribution: SSR models generated a balanced distribution of stigma types (e.g., reducing the over-reliance on "Self-blame" seen in baselines) that aligned more closely with ground truth clinical data.
  • Behavioral Precision: SSR models showed marked improvements in generating challenging behaviors like Denial and Defensiveness, which baselines often failed to produce entirely.
  • Psychometric Validity: When scored against established scales (SSMIS and ISMI), SSR models produced scores indicating meaningful self-stigma levels (shifting from "minimal" to "mild-to-moderate" or "moderate-to-severe"), whereas non-SSR models remained in minimal ranges.
  • Human Evaluation: Expert raters (mental health professionals) rated SSR models as having the highest Authenticity and the clearest shift in perceived stigma levels between neutral and stigma-triggering contexts compared to all baselines.
  • Linguistic Analysis: LIWC analysis confirmed that SSR responses exhibited increased self-focus and reduced social openness, consistent with the linguistic markers of internalized stigma.

Significance and Claims
The paper claims that the SSR framework provides a crucial step toward realistic, theory-informed patient simulations for clinical training and empathetic dialogue systems. By teaching models to generate internal stigma-aware reasoning before producing external responses, the approach yields simulations that are not merely "resistant" but dynamically reflect the complex, situational nature of self-stigma.

The authors explicitly position this work as a training-oriented tool rather than a deployment-ready clinical instrument. They emphasize that the goal is to advance the education of empathetic clinicians and the development of support systems that better understand the impact of self-stigma. The paper maintains a modest stance, acknowledging limitations such as reliance on GPT-5 for annotation and the need for future work to personalize simulations based on individual patient histories. The authors stress that the tool is intended strictly for research and training under expert supervision to avoid risks of misuse, misinterpretation, or the reinforcement of stereotypes.