Towards Efficient and Robust Linguistic Emotion Diagnosis for Mental Health via Multi-Agent Instruction Refinement

Linguistic expressions of emotions, including depression, anxiety, and trauma-related states, are widespread in clinical notes, counseling dialogues, and online mental health communities. Accurate recognition of these emotions is crucial for clinical triage, risk assessment, and timely intervention in mental health related applications. Despite recent advances showing that large language models (LLMs) can generalize well to various emotion analysis tasks, their diagnostic reliability in high-stakes and context-intensive medical settings remains highly sensitive to prompt design. Moreover, existing approaches are challenged by two major issues: emotional comorbidity, where multiple intertwined emotional states complicate prediction, and inefficient exploration of clinically relevant cues. To address these issues, we propose APOLO (Automated Prompt Optimization for Linguistic emOtion diagnosis), a framework that systematically explores a broader and finer-grained prompt space to enhance diagnostic efficiency and robustness. APOLO models instruction refinement as a Partially Observable Markov Decision Process (POMDP) and introduces a multi-agent collaboration mechanism comprising the Planner–Teacher–Critic–Student–Target roles. This closed-loop design enables continuous optimization of prompt generation, evaluation, and evolution. After the Planner agent formulates a high-level optimization trajectory within the POMDP framework, the Teacher–Critic–Student agents collaboratively refine the prompts along this trajectory, iteratively enhancing the stability and effectiveness of the reasoning process. Finally, the Target agent determines whether to continue optimization or terminate the search based on performance evaluation. Experimental results demonstrate that APOLO improves diagnostic accuracy and robustness across domain-specific and stratified benchmarks, providing a generalizable and scalable paradigm for trustworthy LLM applications in mental healthcare.