In recent years, pre-trained models (PTMs) like PhoBERT and ViSoBERT have demonstrated their potential power in natural language downstream tasks using Vietnamese social media text. In particular, ViSoBERT shows that it has surpassed the previous “state-of-the-art” for recognizing emotion from text. Inspired by these studies, we present a new method to leverage these PTMs’s performance on emotion recognition tasks. To do this, our approach utilizes Reinforcement Learning with Active Learning on the pre-trained language models for Vietnamese known as ViSoBERT and PhoBERT. Our proposed approach has been evaluated using cross-validation techniques on the Vietnamese Social Media Emotion Corpus (UIT-VSMEC) public dataset. This corpus fills a critical gap in emotion recognition research for the Vietnamese language, containing 6,927 sentences annotated with emotions. The experimental results showed that the proposed approach enhances the PTMs’s performance.