MentraSuite: Post-Training Large Language Models for Mental Health Reasoning and Assessment
Mengxi Xiao, Kailai Yang, Pengde Zhao, Enze Zhang, Ziyan Kuang, Zhiwei Liu, Weiguang Han, Shu Liao, Lianting Huang, Jinpeng Hu, Min Peng, Qianqian Xie, Sophia Ananiadou
2025-12-16
Summary
This paper focuses on using large language models, or LLMs, to help with mental health support, but recognizes that current LLMs aren't always reliable in this sensitive area because their reasoning can be flawed.
What's the problem?
Many people seek mental health information and support online, and LLMs could potentially help a lot of people. However, existing LLMs often struggle with the kind of careful, step-by-step thinking that's needed for things like understanding someone's situation, making a diagnosis, planning treatment, and making sure their reasoning makes sense. They might give answers that are confusing, inaccurate, or just don't follow a logical path, which is dangerous when dealing with mental health.
What's the solution?
The researchers created a system called MentraSuite to improve LLMs' ability to reason about mental health. This includes a new way to test these models, called MentraBench, which looks at how well they perform on different mental health tasks and how clear and consistent their reasoning is. They also developed a new LLM called Mindora, which was specifically trained to be more reliable and coherent in its reasoning, using a special training process that rewards consistent and truthful answers. They also created a method for generating high-quality training examples for Mindora.
Why it matters?
This work is important because it takes a step towards making LLMs genuinely helpful and safe for mental health applications. By focusing on reliable reasoning, rather than just sounding empathetic or remembering facts, it could lead to tools that provide more accurate and trustworthy support to people who need it, and potentially assist mental health professionals.
Abstract
Mental health disorders affect hundreds of millions globally, and the Web now serves as a primary medium for accessing support, information, and assessment. Large language models (LLMs) offer scalable and accessible assistance, yet their deployment in mental-health settings remains risky when their reasoning is incomplete, inconsistent, or ungrounded. Existing psychological LLMs emphasize emotional understanding or knowledge recall but overlook the step-wise, clinically aligned reasoning required for appraisal, diagnosis, intervention planning, abstraction, and verification. To address these issues, we introduce MentraSuite, a unified framework for advancing reliable mental-health reasoning. We propose MentraBench, a comprehensive benchmark spanning five core reasoning aspects, six tasks, and 13 datasets, evaluating both task performance and reasoning quality across five dimensions: conciseness, coherence, hallucination avoidance, task understanding, and internal consistency. We further present Mindora, a post-trained model optimized through a hybrid SFT-RL framework with an inconsistency-detection reward to enforce faithful and coherent reasoning. To support training, we construct high-quality trajectories using a novel reasoning trajectory generation strategy, that strategically filters difficult samples and applies a structured, consistency-oriented rewriting process to produce concise, readable, and well-balanced trajectories. Across 20 evaluated LLMs, Mindora achieves the highest average performance on MentraBench and shows remarkable performances in reasoning reliability, demonstrating its effectiveness for complex mental-health scenarios.