The field of mental health AI is rapidly evolving, with a focus on developing more accurate and generalizable models for predicting mental health outcomes. Recent studies have highlighted the importance of addressing criterion contamination in language AI models, which can lead to artificially inflated effect sizes and reduced model generalizability. The development of adaptive question-asking frameworks, such as MAQuA, has shown promise in reducing the number of assessment questions required for score stabilization. Additionally, the use of interpretable domain-adapted language models, such as DepressLLM, has demonstrated improved classification performance and reliable confidence estimates. The integration of large language models with item response theory and factor analysis has also shown potential in multidimensional mental health screening. Noteworthy papers include: MAQuA, which reduces the number of assessment questions required for score stabilization by 50-87%. DepressLLM, which provides interpretable depression predictions and achieves an AUC of 0.789. Decoding Neural Emotion Patterns through Natural Language Processing Embeddings, which proposes a computational framework for mapping textual emotional content to anatomically defined brain regions. A Comprehensive Survey of Datasets for Clinical Mental Health AI Systems, which identifies critical gaps in existing datasets and provides recommendations for future dataset curation.