Identifying personality accurately can improve interpretability and adaptability for human–computer interaction and psychological assessment. Recent studies have introduced evidence utterance into large language models (LLMs) for personality recognition, providing supporting evidence to explain personality labels. However, the mixed background and social expressions in dialogue scenarios interfere with the identification of evidence utterance. At the same time, the overconfidence of LLMs leads to biased personality inference. Inspired by the lens theory in psychology, how to systematically explore the mutual influence between evidence utterance identification and personality recognition while improving the performance of both tasks remains a key challenge. In this paper, we reframe personality recognition as an Evidence-Personality Iterative Reasoning problem (EPIR). A graph model is employed to select an initial evidence utterance set, and personality recognition is performed based on a Generative-Discriminative Feedback mechanism. Then, the inferred results are fed back to the graph mapping network to dynamically adjust evidence utterance filtering. This interactive iterative process continues until the evidence and personality results across different iterations converge. Experimental results on the two public datasets demonstrate the effectiveness of the proposed method.

As social media grows, harmful information spreads rapidly across platforms and evolves over time, showing cross-platform and cross-temporal variations. Existing methods rely on fixed model parameters during training, which fail to handle substantial semantic discrepancies, leading to Out-Of-Distribution (OOD) problems. While test-time tuning enables dynamic parameter adjustment, it may lead to excessive adaptation to individual samples. The key challenge is how to adapt to semantic variations during testing while preventing overfitting from continuous tuning. To tackle this issue, this paper proposes RLAT, a reinforcement learning (RL)–guided adaptive tuning method for harmful text detection. First, a tuning joint optimization module is designed to update parameters and adapt to semantic variations during testing. It tunes the model by optimizing consistency loss and applying word-level attention constraints to reduce over-reliance on local words and learn a more robust global representation. Then, to mitigate overfitting caused by continuous tuning, a RL–guided adaptive decision model is introduced to direct the tuning process. It reduces the influence of local samples by selecting data and controlling parameter updates, thereby improving overall test performance. Experimental results show that the RLAT outperforms state-of-the-art baselines in cross-platform and cross-temporal scenarios across multiple public datasets.