Natural language inference (NLI) serves as a core downstream task for evaluating a model’s reasoning capabilities, and large language models (LLMs) have shown remarkable performance across such tasks. However, existing approaches often repurpose LLMs as feature extractors with discriminative classification heads, overlooking their inherently generative and autoregressive nature. To address this issue, we propose Self-Explanation-Augmented Generation (SEAG), a novel framework that reformulates NLI as a generative task to better align with the pretraining objectives of decoder-only LLMs. SEAG consists of two key components: (1) Generative Label Prediction, which prompts the model to directly generate textual labels as natural language tokens, and (2) Explanation-Augmented Attention, which introduces trainable [Exp] tokens that guide the model to construct internal reasoning pathways. This design activates the LLM’s latent reasoning capacity while improving interpretability. Evaluated on the SNLI dataset, SEAG outperforms two baselines, including the Bert-base and Llama-3.2-3B-Instruct variants. Extensive ablation studies confirm that both the generative label prediction and explanation-augmented attention are crucial to performance gains. Our results underscore the potential of generative approaches to not only match but surpass traditional classification strategies by embracing the intrinsic strengths of LLMs.