📄 Team Aurevia at CLPsych 2026: Local Healthcare NLP for Schema-Constrained Self-State Modeling

Summary

This paper presents Team Aurevia's entry to the CLPsych 2026 Shared Task. The task asks systems to track how self-states shift across social-media timelines using the MIND framework, which uses Affect, Behavior, Cognition and Desire categories (each adaptive or maladaptive) rather than diagnostic labels. The subtasks are:

• post-level ABCD subelement detection and 1–5 presence ratings
• detection of Switches and Escalations
• sequence summarization
• deterioration and improvement signatures

The team tested whether a fully local, privacy-preserving pipeline could compete without sending sensitive text to hosted proprietary APIs. Their system combines TF-IDF retrieval of similar training examples with few-shot, JSON-constrained prompting of a 4-bit Qwen2.5-72B-Instruct model run through MLX. It then applies schema repair, ordinal calibration, rule-based temporal signals and cautious post-processing.

The system did best on coarse presence ratings, summary consistency and deterioration signatures. It was weaker on fine subelements, timeline-level change points and improvement signatures. The work shows that tightly controlled open-weight local LLMs are viable for sensitive mental-health NLP. It also points to within-person temporal modeling as the main unsolved problem.

Key points

• Model: mlx-community/Qwen2.5-72B-Instruct-4bit, with greedy decoding and three retrieved posts as few-shot examples. A non-neural branch uses TF-IDF uni/bigrams (5,000 features) with nearest-neighbor voting. Presence calibration blends model outputs with training/development means at weights 0.30/0.70.
• Task 1.2 presence ranked 3rd of 17 (RMSE 0.981, QWK 0.645, Spearman 0.674). Task 1.1 subelements ranked 5th of 17 (average macro F1 0.381).
• Task 2 change detection ranked 7th of 18 (combined macro F1 0.484). Post-level F1 was 0.556, but timeline-level F1 was only 0.413.
• Task 3.1 summaries ranked 5th of 13 overall. They placed 1st on consistency (0.866) and 2nd on contradiction (0.625), but had weaker recall against reference summaries.
• Task 3.2 deterioration signatures ranked 2nd of 9 (0.676), while improvement signatures ranked 6th of 9 (0.465). The authors ran no ablations and did no clinical validation.