A reinforcement learning framework with adaptive psychological state awareness for safe and effective intelligent psychological intervention

Abstract

Intelligent psychological intervention systems require accurate perception of latent emotional states and the formulation of safe long-term strategies, which poses significant challenges to the non-stationary adaptability of learning algorithms. However, most existing approaches rely on static unimodal classification paradigms, neglecting the time-varying evolution of psychological states, while conventional reinforcement learning struggles to balance policy flexibility with real-time ethical risk constraints during decision-making. To address these issues, this paper proposes the RL-MindFlow framework. The proposed method constructs an InfoNCE-based cross-modal attention mechanism to achieve semantic alignment across modalities, and innovatively introduces a PID-Lagrangian controller that dynamically adjusts safety boundaries according to the rate of risk variation. In addition, a hybrid paradigm combining offline pretraining and dataset-driven policy fine-tuning is adopted to improve sample efficiency. Experimental results demonstrate that the proposed model achieves a Weighted F1 score of 68.4% on the MELD task, outperforming the baseline by 5.7%. On the PsyQA task, the average cumulative reward is improved by 15.4%, and the number of interaction steps required for convergence is reduced by 38.5%. In a separate MELD-based audio-noise stress test, the downstream safety proxy remains at 0.891 at SNR = 5 dB. This study proposes a novel framework that combines multimodal affective perception with text-based long-horizon decision-making under dynamic risk constraints and validates the value of control theory in enhancing the safety of reinforcement learning, providing technical support for intelligent psychological service systems.