Overview of the CLPsych 2026 Shared Task: Capturing and Characterizing Mental Health Changes through Social Media Timeline Dynamics

Key Findings

• — Data extends the 40 expert-annotated Reddit timelines of the 2025 task (30 train, 10 test) with 36 added posts to complete windows around Moments of Change, annotated for MIND self-state elements, wellbeing on a rescaled 1-10 GAF scale, presence scores and summaries
• — 39 teams (93 participants) registered and 20 submitted, producing 13 system papers; 17 teams entered the post-level task, 18 the switch and escalation task and 13 the sequence-level task, and every team used large language models
• — Best post-level presence estimates reached RMSE 0.833 for maladaptive and 0.935 for adaptive self-states; the best deterioration-signature score was 0.789
• — The task formalises deterioration detection as a sequence problem (switches versus escalations) rather than a per-post classification, continuing the Moments of Change lineage
• — The test set is ten timelines, and submissions ran through Codabench

Methodology Notes

ACL Anthology 2026.clpsych-1.32, DOI 10.18653/v1/2026.clpsych-1.32, pages 389-421, CLPsych 2026 (San Diego, July 2026; month precision). Organisers at Queen Mary University of London, the Alan Turing Institute, Bar-Ilan University, NIH, George Washington University, University of Ottawa and Johns Hopkins (beat-read PDF title block; curator verified the Anthology page, BibTeX and abstract). The substrate is Reddit timelines, not human-AI dialogue; the shared task measures whether systems can track a person's mental-health trajectory over time from their own writing. Small test set limits the strength of the leaderboard.

Sources