CUNY at CLPsych 2026: A Pipeline Approach to Classification and Summarization of Mental Health Changes

Amirmohammad Ziaei Bideh, Shameed Charlomar Job, Ava Yahyapour, Alla Rozovskaya · 2026

Read CUNY at CLPsych 2026: A Pipeline Approach to Classification and Summarization of Mental Health Changes by Amirmohammad Ziaei Bideh, Shameed Charlomar Job, Ava Yahyapour, Alla Rozovskaya (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-24164.txt.

Introduction

Bideh and three co-authors describe the CUNY submission to the CLPsych 2026 shared task on tracking mental-health change through social-media timelines. To infer dominant self-states in posts, they ensemble three open-weight language models by majority vote. Moments of change are predicted by supervised classifiers built on those self-state predictions. Mood-pattern summaries augment example labels from upstream stages, beating zero-shot and plain in-context baselines. The submission placed first on one task, fourth on two more, and third on the last.

Catalog record

CREATORAmirmohammad Ziaei Bideh, Shameed Charlomar Job, Ava Yahyapour, Alla Rozovskaya
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55170
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2605.24164

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

CUNY at CLPsych 2026: A Pipeline Approach to Classification and Summarization of Mental Health Changes Amirmohammad Ziaei Bideh, Shameed Charlomar Job, Ava Yahyapour, Alla Rozovskaya Abstract We describe our submission to the CLPsych 2026 Shared Task on capturing and characterizing mental health changes through social media timeline dynamics. To infer the dominant self-states in posts (Tasks 1.1 and 1.2), we ensemble in-context learning of three open-weight large language models using majority voting. For predicting moments of change in a timeline (Task 2), we train supervised classifiers on features derived from Task 1.1 predictions. To summarize the patterns of mood dynamics and their progression over time within a timeline (Task 3.1), we augment in-context example labels predicted by upstream systems (Tasks 1.1, 1.2, and 2), yielding performance gains over zero-shot and unaugmented in-context learning baselines. Our submission ranked first on Task 1.1, fourth on Task 1.2, fourth on Task 2, and third on Task 3.1. 1 1 1 The source code for the experiments is available at https://github.com/amirzia/clpsych26-cuny . 1 Introduction Mental health conditions affect a significant portion of the global population 19 ; 11 , creating a need for scalable tools to monitor individuals’ psychological states over time. Social media platforms offer longitudinal data for tracking how mental states evolve in response to life events and social interactions 16 , and large language models (LLMs) have shown strong potential in supporting such analysis 22 ; 7 . Recent practice-oriented research further demonstrates how multimodal analysis and AI can uncover the intrapersonal and interpersonal dynamics underlying therapeutic change 3 . The CLPsych 2026 shared task 1 addresses this need by asking participants to track and characterize how users’ mental states evolve across longitudinal Reddit timelines. This paper describes our submission to the CLPsych 2026 shared task 1 . Our best submission for Task 1.1 uses an ensemble of seven LLM predictions across three model backbones with subelement-level In-Context Learning (ICL), where annotated training examples are included directly in the prompt to guide predictions. For Task 1.2, our best submission ensembles five predictions using ICL and Retrieval-Augmented Generation (RAG), which retrieves the most semantically similar training posts as in-context demonstrations. For Task 2, we train supervised classifiers on self-state features derived from upstream task predictions, using a Support Vector Machine (SVM) for Switch detection and a Random Forest (RF) for Escalation detection. For Task 3.1, our best submission uses label-augmented ICL, enriching prompts with predicted ABCD subelement and Moment of Change (MoC) labels from upstream tasks. For Task 3.2, we apply a batch-and-merge pipeline to identify recurrent dynamic signatures of improvement and deterioration across timelines. Our submission ranked first on Task 1.1, fourth on Task 1.2, fourth on Task 2, and third on Task 3.1. The contributions of our work are as follows: 1. We show that LLM ensembling via majority voting substantially outperforms single-model baselines for subelement classification (Task 1.1), and that expanding the ensemble beyond five members further improves subelement classification but slightly hurts presence scoring (Task 1.2), suggesting diminishing returns for larger ensembles on ordinal prediction tasks. 2. We show that escalation is easier to predict than switch, and traditional supervised classifiers trained on LLM-derived presence scores are competitive for MoC detection (Task 2). 3. We find that propagating predicted self-state labels from upstream tasks into downstream prompts with larger models yields consistent gains in summary quality over zero-shot prompting with smaller models and standard ICL baselines (Task 3.1). 2 Shared Task Description The shared task is grounded in the Multimodal Intrapersonal and Interpersonal Dynamics (MIND) framework 2 . The framework includes widely used psychotherapy constructs on the patient mental state referred to as self-state. A self-state is a dominant mode of experience and consists of a combination of Affect, Behavior, Cognition, and Desire (ABCD) components, their adaptivity level and their more fine-grained subcategorization into subelements 2 . Task 1.1 aims to identify which predefined ABCD subelements are expressed in a post and how they combine into adaptive and maladaptive self-states, referred to as the two valences of a self-state. Task 1.2 requires quantifying the degree to which each identified self-state is present in the post on a 1-5 scale (referred to as the presence score). Task 2 involves detecting clinically meaningful MoC within a user timeline, 2 2 2 A timeline is a chronologically ordered collection of posts authored by a single user. identifying Switches (sudden change in well-being) and Escalations (gradual intensification of mood) 16 . Task 3.1 involves generating a structured summary describing the progression of self-state dynamics within a sequence of posts surrounding an identified change event. Task 3.2 aims to identify recurrent dynamic signatures of improvement and deterioration that recur across multiple sequences and individuals. Dataset. The training set contains Reddit timelines from 30 users, annotated with self-states according to the MIND framework. Figure 3 shows an anonymized excerpt of a sample timeline. The training set also provides gold summaries for sequences, describing the patterns of self-state dynamics. A sequence is a chronologically ordered list of posts within a timeline that culminates in an MoC. We randomly hold out 10 training timelines as our validation set for selecting the best models for submission. The 20 timelines are used for training and providing in-context examples. 3 Related Work Previous editions of the CLPsych shared task have attracted a wide range of system submissions. In CLPsych 2022 16 , team BLUE 6 experimented with several text representation methods, and their best system consisted of an ensemble of machine learning (ML) classifiers. Team WResearch 5 adopted a pipeline approach in which pre-trained BERT 8 was used to compute emotion and sentiment scores that were passed as input features to downstream ML models. Team UoS 4 achieved competitive results using a bidirectional long short-term memory (Bi-LSTM) network for mood change prediction and suicide risk level assessment. The following year’s edition, CLPsych 2025 17 , introduced more challenging subtasks such as evidence span detection and summarization. Team uOttawa 7 explored various prompt engineering strategies on top of a 70B-parameter LLM and obtained the best score on self-state identification. Team BULUSI 14 achieved strong results by combining an ensemble with an optimization step on top of LLM predictions. Finally, team BLUE 15 obtained competitive performance on summarization through zero-shot prompting of open-weight LLMs. 4 Methodology This section presents our approaches, which integrate findings from the CLPsych 2025 shared task, as we enhance and refine our strategies. 4.1 Tasks 1.1 and 1.2 We adopt a joint prediction setup in which an LLM is tasked with predicting both the subelements and their presence scores in a single pass. The default system prompt (Figure 4 ) contains a detailed description of the MIND framework, the definitions of self-states, the characteristics of each subelement, and the criteria for assigning presence scores. We compare several prompting techniques. Zero-shot prompting. We use the default system prompt and pass the content of the post to be labeled in the user message. Post-level In-Context Learning (ICL). Here, k full training posts selected uniformly at random, and their gold subelement annotations are included in the prompt. This exposes the LLM to the surrounding context of each subelement and to the relationships between subelements within a post. A drawback is

Read the complete text