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

Jian Zhang; Zhangqi Wang; Zhiyuan Wang; Weiping Fu; Yu He; Haiping Zhu; Qika Lin; Jun Liu

Source-linked AI summary

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

Jian Zhang, Zhangqi Wang, Zhiyuan Wang, Weiping Fu, Yu He, Haiping Zhu, Qika Lin, Jun Liu

arXiv:2601.13481v1Jan 20, 2026cs.AI

TL;DR

Mental-health emotion diagnosis requires reliable handling of ambiguous, intertwined emotions, yet existing approaches face prompt sensitivity and broader evaluation constraints. APOLO formulates instruction refinement as a POMDP and uses five collaborating agents for adaptive, risk-aware prompt optimization. Across six benchmark datasets, it reports the strongest overall performance, including Macro F1 of 49.25% and Micro F1 of 64.72%.

Problem
Mental-health emotion diagnosis is important for clinical decision-making, but prompt-sensitive LLMs must handle emotional comorbidity, inefficient exploration, annotation subjectivity, generalization challenges, and costly evaluation.
Method
APOLO formulates prompt optimization as a POMDP and combines Planner, Teacher, Critic, Student, and Target agents with risk-aware, cost-constrained planning and iterative feedback.
Results
Across six benchmark datasets, APOLO achieves average Macro F1 of 49.25% and Micro F1 of 64.72%, surpassing OPRO by 2.96% and 3.40%, respectively.
Takeaways & Limitations
APOLO provides a dynamic, interpretable, and risk-aware framework for broader emotion coverage with improved stability across multilingual and risk-sensitive benchmarks.
Takeaways & Limitations
Fine-grained emotion diagnosis remains bounded by annotation subjectivity, cross-lingual and cross-cultural generalization, class imbalance, rare emotions, and costly reliable evaluation.

Abstract

from arXiv · show

Linguistic expressions of emotions such as depression, anxiety, and trauma-related states are pervasive in clinical notes, counseling dialogues, and online mental health communities, and accurate recognition of these emotions is essential for clinical triage, risk assessment, and timely intervention. Although large language models (LLMs) have demonstrated strong generalization ability in emotion analysis tasks, their diagnostic reliability in high-stakes, context-intensive medical settings remains highly sensitive to prompt design. Moreover, existing methods face two key challenges: emotional comorbidity, in which multiple intertwined emotional states complicate prediction, and inefficient exploration of clinically relevant cues. To address these challenges, we propose APOLO (Automated Prompt Optimization for Linguistic Emotion Diagnosis), a framework that systematically explores a broader and finer-grained prompt space to improve diagnostic efficiency and robustness. APOLO formulates instruction refinement as a Partially Observable Markov Decision Process and adopts a multi-agent collaboration mechanism involving Planner, Teacher, Critic, Student, and Target roles. Within this closed-loop framework, the Planner defines an optimization trajectory, while the Teacher-Critic-Student agents iteratively refine prompts to enhance reasoning stability and effectiveness, and the Target agent determines whether to continue optimization based on performance evaluation. Experimental results show that APOLO consistently improves diagnostic accuracy and robustness across domain-specific and stratified benchmarks, demonstrating a scalable and generalizable paradigm for trustworthy LLM applications in mental healthcare.

1 INTRODUCTION

Mental-health emotion diagnosis is clinically important but difficult because medical language is ambiguous, emotionally comorbid, and highly sensitive to prompt design. APOLO addresses these challenges through POMDP-based, multi-agent prompt optimization with risk-aware planning and iterative feedback.

Mental-health texts contain emotions relevant to clinical triage, risk assessment, and intervention, but medical contexts involve semantic ambiguity, implicit expressions, and domain variability.
LLM diagnostic outcomes can change drastically with minor prompt or reasoning-path variations, while handcrafted prompts incompletely cover semantic cues and implicit intentions.
Emotional comorbidity: Emotional comorbidity intertwines states such as anxiety and depression, making fixed-template or single-label prompts prone to missing dependent or less dominant emotions.
Inefficient exploration: Generation–search methods explore limited prompt regions, whereas meta-prompt methods use rigid templates; both lack global planning and dynamic adaptation.
APOLO framework: APOLO formulates prompt optimization as a POMDP and combines Planner–Teacher–Critic–Student–Target roles for adaptive, collaborative instruction refinement.
Contributions: Risk-aware, cost-constrained planning targets emotional comorbidity and uncertain inference, while experiments evaluate effectiveness, robustness, and scalability across linguistic and clinical scenarios.

2 RELATED WORK

Prior emotion-diagnosis methods range from lexicon and classical models to contextual language models, while APO research spans token-, vector-, semantic-, decision-, and multi-agent optimization. Persistent concerns include generalization, annotation quality, rare classes, and evaluation cost.

2.1 Linguistic Emotion Diagnosis: Linguistic emotion diagnosis identifies fine-grained emotions and affective dynamics from text, extending beyond sentiment polarity.
2.1 Linguistic Emotion Diagnosis: Lexicon, rule-based, and classical machine-learning methods require feature engineering and show limited cross-domain transferability, while sequence models capture sentence- and discourse-level dynamics.
2.1 Linguistic Emotion Diagnosis: Multimodal CNN-LSTM studies combine text with speech or facial cues, unlike the pure textual methods addressed here.
2.1 Linguistic Emotion Diagnosis: Transformer-based pretrained models improve contextual and implicit-emotion modeling but retain limitations in few-shot settings, long-tail classes, and cross-domain generalization.
2.1 Linguistic Emotion Diagnosis: Fine-grained emotion research remains constrained by annotation subjectivity, cross-lingual and cross-cultural transfer, class imbalance, rare emotions, and costly reliable evaluation.
2.2 Automated Prompt Optimization: APO evolved from discrete and continuous prompt optimization toward semantic generation-and-search, decision-based control, and multi-role conversational refinement.

3 METHODOLOGY

APOLO formulates linguistic emotion prompt optimization as a POMDP and combines risk- and cost-aware planning with multi-agent Socratic refinement. The framework guides prompt exploration, preserves clinical and semantic coherence, and uses theoretical analysis to characterize expected improvement under bounded penalties.

3.1 Task Formulation: APOLO models linguistic emotion diagnosis as constructing an optimized instruction for reliable multi-label predictions over clinically or psychologically grounded emotional comorbidity.The target model receives text and an initial prompt, while the objective improves completeness and consistency in emotion detection.
3.2 Risk- and Cost-Aware Trajectory Planning: The Planner generates localized sub-goals and selects trajectories using semantic likelihood together with explicit risk and computational-cost penalties.Candidate trajectories are sampled and evaluated under a regularized objective balancing semantic plausibility, clinical safety, and efficiency.
3.2 Risk- and Cost-Aware Trajectory Planning: Risk terms address missed comorbid emotions, ambiguous boundaries, and ignored self-harm cues, while cost terms account for trajectory length, model calls, and latency.The cost coefficients control trade-offs among trajectory length, computational budget, and execution time.
3.3 Joint Policy Optimization: The Teacher–Critic–Student triad iteratively proposes emotion-aware queries, evaluates clarity, relevance, and safety, and updates prompts while conditioning on interaction history.Historical context is included to maintain semantic continuity across interacting symptoms and multi-step refinement.
3.5 Theoretical Insights: The Planner and triad jointly produce stable, interpretable optimization dynamics, with the theoretical bound attributing improvement to Socratic pseudo-gradients minus bounded risk and cost offsets.The bound describes expected improvement after n refinement steps and states that reward improves monotonically up to explicit safety and efficiency penalties.

4 EXPERIMENTS

APOLO is evaluated across six emotion-diagnosis benchmarks, multiple task settings, and three LLM backbones using task-aligned metrics. It achieves stronger accuracy, robustness, and inference-time efficiency than baseline methods.

Experimental Setup: APOLO is evaluated on six public datasets spanning conversational, emotion-cause, noisy social-media, and fine-grained multi-label settings.The evaluation includes DailyDialog, EmoryNLP, PELD, RECCON, EmotionX, and DepressionEmo.
Evaluation Metrics: Macro F1 is primary for multi-class tasks, while Micro F1 provides an aggregate performance view and DepressionEmo additionally uses EMR and PMA.Macro F1 gives equal importance to frequent and rare emotions; EMR is strict exact-set matching, while PMA allows partially correct predictions.
Main Results: 49.25% average Macro F1 and 64.72% average Micro F1 surpass OPRO by 2.96% and 3.40%, respectively, across six benchmarks.APOLO also reaches 28.70% EMR and 93.78% PMA on the reported multi-label evaluation.
Main Results: APOLO improves both single-label and multi-label diagnosis, with gains up to +4.41% EMR and +3.57% PMA on DepressionEmo.The larger relative improvement in Macro F1 than Micro F1 indicates stronger sensitivity to infrequent emotions within the reported experiments.
Robustness Across Models: Across GPT-5-mini, DeepSeek-V3, and Qwen3-32B, APOLO improves average Macro F1 over the strongest baseline by +2.94%, +2.94%, and +3.00%, respectively.The reported gains extend across proprietary and open-source backbones, with comparable Micro F1 improvements.
Efficiency Analysis: At approximately 24.4 × 10^5 generated tokens, APOLO achieves about 7% higher average Micro F1 than OPRO.Conversely, OPRO requires roughly 7 × 10^5 more tokens to reach APOLO’s performance level, indicating a more favorable inference-time efficiency trade-off.

5 SUPPLEMENT ANALYSIS

The supplementary analyses examine APOLO’s component contributions, convergence, initialization robustness, and data efficiency. Results indicate that Socratic refinement is central, Planner structure is important, optimization converges rapidly and stably, and one-shot prompting can outperform baselines using much more data.

5.1 Ablation Study: Removing the Socratic module produces the largest average degradation: Macro F1 decreases 7.19%, Micro F1 8.69%, EMR 11.22%, and PMA 12.07%.The module drives iterative reflection and prompt refinement through multi-round question–answer interactions.
5.1 Ablation Study: Removing the Planner causes the second-largest average declines: Macro F1 decreases 3.75%, Micro F1 4.90%, EMR 6.66%, and PMA 6.73%.The Planner decomposes optimization into subgoals, enabling focused multi-stage refinement.
5.1 Ablation Study: Removing the Critic causes smaller but consistent declines of 2.00% in Macro F1, 3.00% in Micro F1, 4.08% in EMR, and 3.72% in PMA.The Critic monitors Teacher–Student interactions and supports stable, principled refinement.
5.2 Convergence Analysis: APOLO shows rapid early improvement followed by smooth stabilization across six datasets, with most gains occurring during iterations 1–4.Performance plateaus around iterations 6–10 as optimization shifts from coarse structural changes to fine-grained affective refinement.
5.3 Initial Prompt Analysis: APOLO’s final RECCON Micro F1-scores across six initial prompts have a standard deviation of only 0.6647.The result indicates limited sensitivity to initialization quality.
5.4 Sample Size Analysis: APOLO with one exemplary dialogue consistently outperforms baselines using up to 100 samples; on DepressionEmo, its 1-shot Micro F1-score is 82.45%.Increasing from one to three shots yields only marginal improvements across the six tasks.

6 DETAILS OF DATASETS AND OPTIMIZED PROMPTS

This section describes the six evaluation datasets, APOLO’s role-specific prompts, and a case study of multi-agent prompt refinement. Together, the materials show how the framework is applied across varied emotion-diagnosis settings and transforms an ambiguous instruction into a structured prompt.

6.1 Datasets: DepressionEmo is a multilabel corpus of naturalistic online text with fine-grained depressive emotions including sadness, hopelessness, loneliness, worthlessness, and emptiness.
6.1 Datasets: DailyDialog contains 1,000 test dialogues with emotion and communicative-intent labels, while EmoryNLP provides 1,328 context-preserving utterances across seven categories.
6.1 Datasets: PELD combines Big Five personality profiles with response emotions across more than six thousand dialogue triples.
6.1 Datasets: RECCON jointly examines target-utterance emotion and contextual antecedents that plausibly triggered it, making the task causality-aware and discourse-oriented.
6.1 Datasets: EmotionX evaluates Friends and EmotionPush test portions, enabling robustness analysis across scripted dialogue and spontaneous chat registers.
6.2 Prompts for Agents: APOLO assigns structured system prompts to the Planner and Socratic agents, while the Target prompt is the refined final output rather than a predefined instruction.
6.3 Case Study: In the DepressionEmo case study, the Planner decomposes an ambiguous single-label prompt into six steps, after which Critic feedback and Socratic dialogue produce a nuanced prompt handling multiple emotions.
6.3 Case Study: The completed optimization yields P* with Chain-of-Thought structure, role-playing, and precise output constraints after all Planner steps are completed.

7 CONCLUSION

APOLO addresses emotional comorbidity and inefficient exploration through POMDP-based, risk-aware multi-agent prompt optimization. Across multilingual and risk-sensitive benchmarks, it consistently outperforms existing methods in diagnostic accuracy, calibration, and robustness.

APOLO uses Planner–Teacher–Critic–Student–Target collaboration to optimize prompts through adaptive reasoning and feedback alignment.
The framework targets emotional comorbidity and inefficient exploration while seeking more comprehensive emotion coverage, stability, and interpretability.
Across multilingual and risk-sensitive benchmarks, APOLO consistently outperforms existing methods in diagnostic accuracy, calibration, and robustness.