LLM Chatbot Uses Game Theory to Pick the Right Therapy Move for Stressed Students

Researchers have built a mental health chatbot that treats every conversation turn as a strategic game. The system, called DBT-wellMind-GameTree, combines large language models with game-theoretic modeling and the principles of Dialectical Behavior Therapy (DBT) to decide, before each response, which therapeutic approach will best serve a student in that moment. In a study published in BMC Artificial Intelligence, the framework scored higher than ChatGPT and Gemini baselines on measures of goal achievement, satisfaction, and coping strategies among 31 students facing academic stress.

The motivation behind the work lies in a gap the authors identified across existing AI well-being chatbots. Most prior systems, they report, have centered on cognitive behavioral therapy, focusing on cognitive restructuring and behavioral activation, while leaving DBT components unexplored. DBT is a specialized form of cognitive behavior therapy built around four skill modules: Mindfulness, Distress Tolerance, Emotion Regulation, and Interpersonal Effectiveness. According to the researchers, no earlier study they could find had systematically integrated all four components into a single system with strategic decision-making to select the optimal component for a user’s psychological needs.

The framework’s core mechanism works like this: when a student sends a message, a dedicated reward generator agent, itself powered by a large language model, analyzes the utterance, the conversation history, and the previous tree state. Using chain-of-thought prompting structured as a multi-level thinking process, the agent generates possible mental states for each of the four DBT components, along with a reward score and a probability for each state. The reward reflects how effectively the following DBT component would address the detected mental states, with a value of 1.0 representing the optimal therapeutic effect achievable, 0.8 to 0.9 indicating significant improvement, and 0.6 to 0.7 indicating moderate improvement.

These scores then feed a dynamic game tree in which nodes represent DBT components and edges represent possible transitions based on the therapeutic context. The system computes an expected utility for each component by summing the product of reward and probability values across all identified mental states, then selects the component with the highest utility as the next action. The selected component’s associated skills are activated to craft the response, and the previously chosen action becomes the root of the tree for the next turn. This iterative cycle repeats before every response until the conversation ends, allowing the model to perform lookahead planning, track context across turns, and adapt as the student’s emotional state evolves.

To generate the actual therapeutic responses, the team employed two state-of-the-art models, GPT-4o and Claude 3.7, configured with a token size of 60 per response and a temperature of 0.4, a setting chosen to balance coherent, focused answers with moderate variation. The response generation uses a multi-shot learning approach supported by DBT-skill-based conversation datasets that the researchers crafted specifically for the study. The framework also includes personalization and risk detection features that repeatedly analyze the user’s context, behavioral cues, and language tone, including screening for any indication of suicidal activity.

The evaluation involved 31 participants, 12 females and 19 males, all currently enrolled students ranging from 12th grade through doctoral study. To ensure cultural diversity, the team invited students from Japan, France, India, Bangladesh, and China, recruited by direct invitation from their research laboratory. Participants chatted with the system on their own devices and then rated their experience on questionnaires using a 7-point Likert scale. The study obtained consent from all participants, and the system collects no personal identification data; chat logs are stored as anonymized sessions with SSL encryption on the server.

The headline results show a consistent advantage for the game-tree approach. The framework running on GPT-4o achieved a mean score of 5.16 (SD 0.61) on goal achievement, 5.03 (SD 0.53) on satisfaction, and 5.19 (SD 0.68) on coping strategies. With Claude 3.7, the scores rose to 5.65 (SD 1.05), 5.58 (SD 0.86), and 5.61 (SD 0.94), respectively. By comparison, the ChatGPT baseline yielded means of 4.50, 4.63, and 4.40, while Gemini scored 4.11, 4.12, and 4.08 on the same metrics. Using paired one-tailed t-tests, appropriate because the same participants assessed both systems, the differences against the ChatGPT baseline were statistically significant for the GPT-4o version: p = 0.02 for goal achievement, p = 0.01 for satisfaction, and p = 0.004 for coping strategies. For the Claude 3.7 version, the p-values were 0.001, 0.0003, and 0.0003, respectively.

Usability and reliability assessments reinforced the picture. On positive usability questions, the Claude 3.7 version peaked at 5.73 for providing useful information responses and 5.68 for explaining its purpose, while the GPT-4o version scored highest on ease of use at 5.32. Negative items revealed room for improvement: the GPT-4o version’s highest negative scores were 1.93 for feeling too robotic and 1.6 for unfriendliness, with similar patterns for Claude 3.7. On reliability, the Claude 3.7 configuration achieved the highest mean scores across all five questions, reaching 4.97 for belief in the chatbot’s knowledge, though notably higher standard deviations, up to 1.34, indicated less stable user perceptions compared with the tighter 0.51 to 0.81 spread seen for the other configurations. The framework was also compared against eleven conversational quality metrics from prior state-of-the-art chatbot studies on a 5-point scale, where the Claude 3.7 version scored 4.27 for helpfulness, 4.32 for satisfaction, and 4.2 for naturalness, against 3.45, 3.3, and 1.64 for the existing method.

The authors are careful to frame what the study does and does not establish. The reward generation process is an LLM-based intrinsic approach that does not rely on criteria-based rewards, and the long-term effectiveness of the system was not measured using standardized instruments such as PHQ-9 or GAD-7 scores. The participant pool, while diverse in nationality and academic level, remains modest in size, and the authors state that stronger evaluation with more participants is needed. They also acknowledge that the robotic tone and perceived unfriendliness flagged in the usability data require improvement through more effective prompt instructions, and that the system cannot detect severe activities reliably.

Those caveats aside, the study offers a concrete demonstration of how structured therapeutic design can be layered onto large language models. Rather than letting the model freestyle a supportive reply, the framework forces each response through an explicit decision pipeline: detect mental states, score every DBT component’s fit, compute expected utility, and only then generate text grounded in the selected component’s skills. The researchers suggest that future work will add criteria-based evaluation of DBT components, expand the participant pool, and introduce long-term monitoring with PHQ-9 and GAD-7. For now, the results indicate that game-theoretic planning over therapy components is a viable way to make LLM well-being chatbots more goal-oriented, more personalized, and measurably more useful to students under academic pressure.

Subject of Research: A game-theoretic LLM chatbot framework that selects Dialectical Behavior Therapy components to support student mental health

Article Title: DBT-wellMind-GameTree: a game-theoretic framework for student mental health support using LLMs

Article References: Kaysar, M. N., Nagasawa, F., & Shiramatsu, S. (2026). DBT-wellMind-GameTree: a game-theoretic framework for student mental health support using LLMs. BMC Artificial Intelligence, 2(1), Article 8. https://doi.org/10.1186/s44398-026-00025-w

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00025-w

Keywords: large language models, game theory, Dialectical Behavior Therapy, mental health chatbot, student well-being, academic stress, GPT-4o, Claude 3.7, game tree search, expected utility, coping strategies, dialogue system