Professional mental‑health services are often scarce, prompting many users to turn to platforms such as Reddit for peer support grounded in lived experience. A large share of help‑seeking posts, however, remain unanswered, opening an opportunity for large language models (LLMs) to step in. While LLMs have shown strong results on clinical benchmarks, their capacity to generate community‑aligned, experience‑informed support has been under‑examined. To address this, we introduce the COmmunity‑centered Peer Engaged Support (COPES) dataset and a three‑axis evaluation framework that measures alignment on strategy, emotion, and tone for mental‑health queries. We evaluate zero‑shot models against models fine‑tuned with supervised fine‑tuning (SFT) and direct preference optimization (DPO). Post‑training on COPES raises strategy alignment by roughly 50% for general‑purpose models and improves emotion and tone matching. The gains are heterogeneous: alignment improvements vary widely across subreddits and requested coping strategies. Moreover, fine‑tuning induces a distributional shift, heavily favoring problem‑focused recommendations while suppressing emotion‑focused ones. In sum, community‑driven data can enhance LLM alignment, yet performance remains uneven across sub‑communities and specific mental‑health needs.
Review