<p>Stance detection, a pivotal task in natural language processing, aims to identify the attitudinal orientation of texts toward specific targets, holding significant value in various scenarios such as public opinion analysis and opinion mining. However, existing lightweight large-model-based text stance detection suffers from issues including insufficient deep reasoning, limited semantic understanding, and low inference efficiency. To address these challenges, this paper proposes a multi-round feedback enhancement framework for lightweight large models, whose overall architecture includes an input module, a backbone language model, a hidden-state feedback module, and a classifier module. To fully exploit the potential of deep hidden states in large language models for discriminative tasks, a hidden-state feedback enhancement mechanism is proposed. By introducing a multi-round feedback optimization strategy, this mechanism realizes the semantic transformation and residual injection of deep hidden states, strengthening the model’s semantic representation capabilities in intermediate layers and thus improving overall recognition performance. Additionally, a lightweight linear classification head structure is proposed to aggregate and discriminate the hidden states output after multi-round reasoning. Experimental results on multiple public stance detection datasets demonstrate that the proposed method achieves high accuracy in stance detection and significantly improves classification performance. Comparative experiments across different large models further validate the method’s superiority, highlighting its practicality and application potential.</p>