Mind manipulation in dialogue exploits emotional vulnerabilities via covert tactics and hidden goals. We present a fine-grained framework that (i) detects turn-level manipulation tactics and (ii) infers dialogue-level intent before issuing a final manipulation judgment. To support training and evaluation, we curate a balanced dataset of 776 dialogues (388 manipulative, 388 non-manipulative; 15520 dialogue turns) generated via a three-phase multi-agent simulation with dual verification. On this benchmark, our model achieves 84.01% accuracy for dialogue-level manipulation detection, outperforming the strongest baseline by +6.83%. It further attains 76.25% accuracy on tactic detection and 79.53% BERTScore for intent prediction. A user study on forward-simulatability shows +11.0% accuracy improvement when the tactics and intent detected by our model are provided as rationale for dialogue manipulation detection results. These results indicate that explicit, stepwise reasoning over tactics and intent yields both higher performance and actionable interpretability for proactive monitoring of manipulative conversations.

With the accelerated pace of life and the intensification of social competition, people today experience unprecedented stress. Timely detection of stress could aid people to know their stress at an early stage, hereby taking actions to manage the stress before health deteriorates and bad consequence happens. In this study, we leveraged ubiquitous surveillance videos and deep-learning techniques for group-aware personalized stress detection. We tackled the problem through three subtasks, including pre-processing of surveillance video (subtask 1), learning individual's emotion manifestation (subtask 2) and aggregating group's emotion manifestation (subtask 3). The performance study on two datasets verified the effectiveness of our video-wise emotion-oriented personalized stress detection solution, as well as the proposed emotion-oriented frame encoding with consecutive negative emotions attention and personality learning with group attention. Limitations and further possible improvements are also discussed at the end of the paper.keywords: Surveillance video; personalized stress detection; deep learning; emotion recognition; group attention

Stress detection is critical for mental and physical well-being, yet traditional methods such as self-reports and physiological sensors face limitations in efficiency and scalability. Video-based stress detection, leveraging visual cues learned from an annotated video database, offers a non-invasive, cost-effective alternative. However, most models function as black boxes, lacking transparency in their decision-making process, which hinders their trustworthiness. To address this, we propose an interpretable video-based stress detection model that incorporates Chain-of-Thought (CoT) reasoning of large foundation models. Our model follows a structured reasoning chain “Describes Assess-e-Highlight”, mimicking the decision process of psychology experts. To further enhance model reliability, we integrate a self-refinement mechanism that allows the model to reflect on and improve its predictions using Direct Preference Optimization (DPO) to ensure accuracy and faithfulness. Experimental results on two video-based stress detection datasets demonstrate that our approach outperforms state-of-the-art models in both accuracy and interpretability. We release our code at https://github.com/debby1103/stressdetection.git.

Mental stress assessment is crucial for mental and physical well-being. However, it faces limitations due to domain fragmentation, in which contextual variations in stress triggers and demographics hinder the generalization of assessment models across real-world scenarios. Additionally, mental stress assessment is sensitive and human-centric due to its implications for mental health interventions, emphasizing the need for model transparency and trustworthiness. To address this gap, we propose Retrieval-Augmented Reasoning, a novel framework that bridges domain gaps in mental stress assessment through transparent step-by-step reasoning and dynamic in-context example retrieval. Our framework introduces two key components: (1) a ''detect-then-assess'' reasoning chain decouples stress-relevant facial action units (AUs) from domain-specific noise by first generating textual descriptions as intermediate reasoning step (e.g., ''eyebrow: inner portions raised''). The model then reflects on and learns to refine these descriptions via Direct Preference Optimization (DPO), ensuring faithfulness and helpfulness; (2) a dual-encoder multimodal retriever dynamically selects proper in-context examples from source domain to enhance target-domain assessments, leveraging feedback from the assessment model to optimize retrieval. Experimental results demonstrate that our framework consistently outperforms large multimodal foundation models, stress assessment baselines, and domain generalization methods.

Stress haunts people in modern society, which may cause severe health issues if left unattended. With social media becoming an integral part of daily life, leveraging social media to detect stress has gained increasing attention. While the majority of the work focuses on classifying stress states and stress categories, this study introduce a new task aimed at estimating more specific stressors (like exam, writing paper, etc.) through users' posts on social media. Unfortunately, the diversity of stressors with many different classes but a few examples per class, combined with the consistent arising of new stressors over time, hinders the machine understanding of stressors. To this end, we cast the stressor estimation problem within a practical scenario few-shot learning setting, and propose a novel meta-learning based stressor estimation framework that is enhanced by a meta-knowledge inheritance mechanism. This model can not only learn generic stressor context through meta-learning, but also has a good generalization ability to estimate new stressors with little labeled data. A fundamental breakthrough in our approach lies in the inclusion of the meta-knowledge inheritance mechanism, which equips our model with the ability to prevent catastrophic forgetting when adapting to new stressors. The experimental results show that our model achieves state-of-the-art performance compared with the baselines. Additionally, we construct a social media-based stressor estimation dataset that can help train artificial intelligence models to facilitate human well-being.

Suicide presents a global health challenge, prompting the development of diverse prevention strategies. Among them, timely identification of individuals at risk of suicide remains challenging. Although social media offers potential for tracking users’ mental status, harnessing collaboration between AI and human experts for real-time prediction of suicide risk is inadequately explored. This study presents a human-in-the-loop framework for real-time suicide risk prediction based on social media. Once a user made a new post on social media, the AI model assesses user’s suicide risk within the next month with explanation based on the historic and new posts plus domain knowledge. Human experts on the other side look into the explanation to confirm/clarify uncertain information as feedback, enabling consistent evolution of the model. Experiments on the constructed dataset, containing 66 suicidal users and 66 non-suicidal users, show that our method achieved 82.58% prediction accuracy, outperforming competitive baselines by 6.57%. Leveraging human feedback improved prediction accuracy by 4.12%. Consultation with 18 experts (including 6 medical staff and 12 psychologists) was conducted to examine the validity of our method. Ethics considerations, as well as potential and limitations of large language models in mental condition prediction, are also discussed at the end of the paper.

Leveraging social media for stress detection has been growing attention in recent years. Most relevant studies so far concentrated on training a stress detection model on the entire data in a closed environment, and did not continuously incorporate new information into the already established models but instead regularly reconstruct a new model from scratch. In this study, we formulate a social media based continuous stress detection task with two particular questions to be addressed: (1) when to adapt a learned stress detection model? and (2) how to adapt a learned stress detection model? We design a protocol to quantify the conditions that trigger model's adaptation, and develop a layer-inheritance based knowledge distillation method to continually adapt the learned stress detection model to incoming data, while retaining the knowledge gained previously. The experimental results on a constructed dataset containing 69 users on Tencent Weibo validate the effectiveness of the proposed adaptive layer-inheritance based knowledge distillation method, achieving 86.32% and 91.56% of accuracy in 3-label and 2-label continuous stress detection. Implications and further possible improvements are also discussed at the end of the article.

Psychological stress has become a wider-spread and serious health issue in modern society. Detecting stressors that cause the stress could enable people to take effective actions to manage the stress. Previous work relied on the stressor dictionary built upon words from the stressor-related categories in the LIWC (Linguistic Inquiry and Word Count), and focused on stress categories that appear frequently on social media. In this paper, we build a meta-learning based stress category detection framework, which can learn how to distinguish a new stress category with very little data through learning on frequently appeared categories without relying on any lexicon. It is comprised of three modules, i.e., encoder module, induction module, and relation module. The encoder module focuses on learning category-relevant representation of each tweet with Dependency Graph Convolutional Network and tweet attention. The induction module deploys Mixture of Experts mechanism to integrate and summarize a representation for each category. The relation module is adopted to measure the correlation between each pair of query tweets and categories. Through the three modules and the meta-training process, we can then obtain a model which learns to learn how to identify stress categories and can directly be employed to a new category with little labelled data. Our experimental results show that the proposed framework can achieve 75.3 accuracy with 3 labeled data for the rarely appeared stress categories. We also build a stress category dataset consisting of 12 stress categories with 1,553 manually labeled stressful microblogs which can help train AI models to assist psychological stress diagnosis.

Automated deception detection has been found as a vital and concerned task, capable of assisting human users to assess truthfulness and detect deceptive behaviors in several situations (e.g., medical, legal, as well as occupational domains). As contact-free video cameras and data analysis techniques are leaping forward, leveraging one’s visual, acoustic and textual information captured in a video can be cost-effective for deception detection. In this study, we aim at a fine-grained question-level deception detection task, focusing on detecting whether a subject lies or not in answering each question rather than providing an overall “lie or not” judgement. A Graph-based Cross-modal Fusion Model (GCFM) is presented to learn the inherent associations among the subject’s reactions to different questions, plus a novel cross-modal attention mechanism to enhance the model’s learning capability. As revealed by the experimental results on the two datasets, the proposed graph-based GCFM outperformed eight other methods, and its two alternatives (clustering K-means based and sequential learning LSTM based methods), achieving accuracy 88.14% and 86.91% on the two datasets, respectively. Besides, through association learning, GCFM could increase the accuracy by 1.87% and 4.33% on the two datasets, respectively. Furthermore, its cross-modal attention mechanism led to the improvement of accuracy by 2.44% and 2.95% on the two datasets, respectively.