Psychological analysis improved with question-based judgment model
PsyJev: Psychological Analysis via System-1 Judgment
Artificial Intelligence
Summary
Understanding people's feelings and symptoms from what they share online can be complex. The authors made a new model called PsyJev that answers specific psychological questions based on text, using a shared way to judge diverse mental health topics. They trained this model with many different datasets and improved its accuracy using a special technique called Judgment-Anchored RLCD. Larger versions of the model perform better than previous methods, offering more reliable psychological insights from web disclosures.
What this means in practice
- •For mental health software developers: Build tools that analyze user-generated texts to detect symptoms and experiences across different psychological topics using a single, adaptable model.
- •For social media content moderators: Implement better automated systems to flag and understand mental health-related disclosures by interpreting varied psychological signals within texts.
Authors
Fangyuan Liu, Sirui Zhao, Xiaohao Wang, Guo Cheng, Peng Zhang, Enhong Chen
Abstract
Self-disclosures on the Web support diverse psychological analyses, from recognizing expressed experiences to identifying symptom-relevant information. We study a shared, question-conditioned decision interface for this heterogeneous problem in content analysis. We organize PsyJudgeBench, comprising 13 annotated datasets and 82 questions across six psychological areas, while preserving source-specific criteria and masks for unobserved labels. We introduce PsyJev, a System-1 model that directly produces distributions over answers specified in each request. Joint supervised training establishes the shared judgment function, followed by Judgment-Anchored RLCD (JA-RLCD), which adapts logit-space reward optimization to heterogeneous psychological judgments through balanced supervision, selective retention of correct SFT predictions, and question-group reward optimization. The Qwen3.5-based 2B, 4B, and 9B models achieve 64.33, 66.21, and 67.56 average test F1, respectively. The 9B model exceeds the strongest external comparison by 2.54 points. Controlled 2B comparisons examine the training stages and anchored refinement. Further analyses examine question-conditioned judgment, inference cost, and probability quality, complemented by qualitative comparisons across models.