Abstract: ObjectivesEmergency response for carrier-based aircraft fires is characterized by high complexity and strong coupling. However, the “black-box” nature of firefighting emergency decision-making models makes it difficult for commanders to understand the underlying decision logic, leading to trust barriers. To address this issue, this study proposes an interpretable emergency decision-making framework based on the collaboration between a small model and a large model.MethodsFirst, the framework employs a Conv2D-based intent-disentanglement small model to adaptively separate latent intents—such as “resource allocation” and “risk control”—from the features of firefighting emergency response. These disentangled intent embeddings are then aligned to the representation space of a large language model through linear mapping. Using carefully designed prompts, the aligned intent embeddings are injected into the large model for collaborative fine-tuning, enhancing its understanding of complex carrier-based aircraft fire scenarios. During inference, the predefined prompts incorporate both fire-event features and the small model’s recommended response actions to guide the large model in generating structured and logically coherent interpretable emergency response plans.Results Experimental results on a carrier-based aircraft fire emergency response dataset show that the proposed framework improves action-recommendation accuracy (Recall@7) by 5.8%. Furthermore, on interpretability evaluation metrics such as GPTScore and BERTScore, the generated responses demonstrate greater consistency and rationality compared with baseline models of similar parameter size.ConclusionsThe findings indicate that the proposed method effectively alleviates the black-box issue in carrier-based aircraft firefighting emergency decision-making, improving both the accuracy and interpretability of emergency response recommendations. This provides strong support for carrier-based aircraft firefighting operations and related research.