Abstract: Abstract：Objectives The operational environment for carrier aircraft support is complex and dynamic, making it highly susceptible to accidents, especially scenarios with novel accident characteristics that impose higher demands on emergency responses. Traditional case-based reasoning (CBR) methods rely heavily on historical cases and commander experience, suffering from data sparsity and cold-start problems due to limited case coverage. Deep learning models partially alleviate data sparsity via multi-view graph contrastive learning, yet the absence of historical interactions in cold-start scenarios restricts their representation learning capability and decision-making performance.Method To address this issue, this paper proposes CILLM, a large language model (LLM)-enhanced emergency decision method for carrier aircraft support. CILLM improves cold-start recommendation performance from two perspectives: content representation enhancement and collaborative interaction enhancement. For content representation enhancement, it enhances the textual semantics of accident features and handling plans using LLMs, and fuses them with structured semantics derived from a domain knowledge graph via a gating mechanism. For collaborative interaction enhancement, it adopts prompts to guide the LLM in inferring potential interactions between cold-start features and candidate sub-plans. The enhanced representations are aligned through contrastive learning.Result Experiments on a simulated dataset demonstrate that CILLM improves Precision and Normalized Discounted Cumulative Gain (NDCG) by 2.08% and 1.67% respectively over ColdLLM, the LLM-enhanced cold-start recommendation baseline, and by 72% and 49% over the CBR method.Conclusion The experimental results validate its superior effectiveness in handling emergency decision-making under cold-start scenarios for carrier aircraft support operations.