Aligning with Lived Experience: Heterogeneous Benefits of Fine Tuning in Mental Health Support Generation
Introduces COPES (Community-centered Peer Engaged Support), a dataset of mental-health support-seeking Reddit queries with community-endorsed responses (5,536 posts after filtering, across five community types), and a three-axis framework for judging whether language-model replies align with community perspectives. Post-training on COPES (supervised fine-tuning and DPO) raised strategy alignment by more than 50% for general-purpose models and improved emotion and tone alignment, but the gains were uneven across subreddits and coping strategies, and post-training shifted outputs toward problem-focused recommendations while suppressing emotion-focused ones.
Publisher
arXiv (Georgia Institute of Technology)
Published
17 Sept 2026
Added
3 weeks ago
DOI
—
Key Findings
- 5,536 filtered support-seeking posts with endorsed peer responses across five community types; three-axis alignment evaluation
- Post-training on COPES improved strategy alignment by more than 50% for general-purpose models and improved emotion and tone alignment
- Improvements vary significantly across subreddits and requested coping strategies
- Post-training induces a distributional shift toward problem-focused recommendations and away from emotion-focused strategies
Methodology Notes
Reddit-derived dataset; zero-shot versus fine-tuned and DPO-trained models; alignment judged against community-endorsed responses rather than clinical outcomes. Date is the arXiv v1 submission date (17 September 2026); announced 21 September 2026; 25 pages, 17 tables; CC BY. One author is affiliated with a company (Magic Hour).
Sources
arXiv preprint(opens in a new tab) (primary)
Archived snapshot (Wayback Machine)(opens in a new tab) — preserved against link rot
Authors
Chandra, Mohit, Kim, Nabin, Min, Eli, Sawant, Aamogh, Sutar, Tanmay, De Choudhury, Munmun
Tags
Cite This
APA
Chandra, Mohit et al. (2026). Aligning with Lived Experience: Heterogeneous Benefits of Fine Tuning in Mental Health Support Generation. arXiv (Georgia Institute of Technology). https://arxiv.org/abs/2609.21075
Related Insights
Like a Therapist, But Not: Reddit Narratives of AI in Mental Health Contexts
Association for Computational Linguistics (Findings of ACL 2026) · 1 Jul 2026
Informational Help-Seeking on Reddit Did Not Decline After ChatGPT
arXiv (Computer Science, New York University Abu Dhabi) · 11 Sept 2026