LLMs struggle to balance persona-specific support with generic empathy for drug users

Researchers have developed a persona-aware approach for Large Language Models (LLMs) to provide support to individuals who use drugs, addressing the issue of self-stigma. Through Latent Profile Analysis on Reddit data, they identified four distinct personas related to self-stigma expression. Sequential Bayesian and recurrent neural classifiers were able to identify these personas from user posting histories, outperforming standard LLM baselines. However, clinical expert evaluations showed a divergence: while persona-matched responses achieved targeted behavioral shifts, raters preferred the generic empathy of persona-neutral LLMs, indicating a potential conflict between clinically aligned design and holistic empathy judgments. AI

IMPACT This research highlights the challenges in designing LLMs for sensitive support roles, suggesting a need for more nuanced evaluation metrics beyond generic empathy.

RANK_REASON Academic paper detailing a novel methodology for LLM-based support. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Hugging Face
• Large Language Models
• Latent Profile Analysis
• recurrent neural classifiers
• Reddit
• Sequential Bayesian learning for State Space Models

• paper
• safety
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →