LongCounsel-8 is a new benchmark suite comprising three datasets with 7,749 five‑session counseling dialogues, each grounded in real client profiles, depression trajectories, symptom compositions, and counseling patterns. The benchmark addresses challenges of longitudinal consistency, empirical grounding of symptom progression, and natural expression of controlled depression states without exposing labels. Experiments show that lower single‑session error does not ensure accurate trend detection, methods perform worse on worsening trajectories, and adding more session history can reduce trend prediction accuracy.
By Jiayi Li, Zhaomin Wu, Bingsheng He
EviDial is a new benchmark for depression dialogue that separates label agreement from evidence grounding. It contains 382 dialogues and 2,744 participant-symptom cells, each annotated as absent, partial, conflicting, or sufficient evidence. The study finds that most predictions that match benchmark labels lack sufficient textual evidence, highlighting a gap between label accuracy and evidence-based assessment.
By Junjie Xu, Jin Zhu, Yuzhe Yang, Luhan Wang, Xingjiao Wu, Daoguo Dong, Jing Yang, Wenxin Wu, Liang He
arXiv:2605.22286v2 Announce Type: replace-cross
Abstract: Text-based counseling provides a valuable source of information for assessing depression severity. We study prediction of the total score on...
By Zhaomin Wu, Jiayi Li, Bingsheng He
The study investigates how a large language model, Gemma-3-27B-PT, internally represents depressive symptoms. By applying mechanistic interpretability methods to the model’s residual stream, researchers found that symptom groups are geometrically distinct at layer 21, and that projected symptom vectors align with clinician-annotated rankings across mood, somatic, and suicidality dimensions. Additionally, a single depression vector at this layer can differentiate depressive from non-depressive text with an AUC of 0.789, suggesting a potential emotional valence gate for symptom projection.
By Fangyi Zhu, Ajay Subramanian, Allison Constant, Camille Wang, Ravish Gupta, Corey J. Keller
BALMS is a benchmark for evaluating large language model (LLM) agents that analyze longitudinal wearable data to predict mental‑health wellbeing scores and generate evidence‑grounded rationales. It covers three real‑world datasets, two task families (score prediction and rationale generation), and tests five LLM backbones across open‑ and closed‑source paradigms. The study finds that zero‑shot agents rarely beat a simple mean baseline, and while chain‑of‑thought prompting helps reasoning, it does not ensure temporal grounding or numerical accuracy.
By Yu Yvonne Wu, Arvind Pillai, Yuliang Chen, Yuwei Zhang, Sudarshan Regmi, Tess Z. Griffin, Michael V. Heinz, Lisa A. Marsch, Nicholas C. Jacobson, Andrew Campbell
MERID is a framework that uses recursive self‑improvement agents to autonomously develop multimodal pipelines for detecting major depressive disorder. It aligns multimodal records with depression targets, jointly modifies representations, fusion, and predictors, and guides revisions through evidence‑guided evolution to validate improvements before inheritance. Experiments on depression benchmarks show MERID outperforms existing multimodal and agent‑based baselines, especially highlighting the importance of acoustic and linguistic cues.
By Lei Liu, Zhaokang Liang, Qingcheng Zeng, Chenda Duan, Lu Mi, Zhen Tan, Tianyu Liu