Multiple autonomous underwater vehicles (AUVs) have been widely utilized for underwater search missions. However, traditional multi-AUV collaborative search methods based on multi-agent reinforcement learning (MARL) suffer from a cold start problem due to the lack of prior knowledge, which hinders efficient cooperation in the early stage of search operations. Furthermore, due to harsh environments and unstable underwater acoustic communication links, it is difficult for AUVs to obtain timely information, which reduces collaborative search efficiency among AUVs. To address these challenges, we integrate the reasoning capabilities of large language models (LLMs) into MARL framework, and propose an LLM-enhanced multi-AUV cooperative search scheme. In this scheme, to tackle with the cold start problem, we leverage LLMs to infer a target probability map (TPM) from the search task and underwater terrain information. Then, we use TPM to guide AUV’s search and alleviate the cold start problem caused by unfamiliarity with the new environment in the early stage of search operations. Moreover, we design an information forecaster (Forecaster) using LLMs, which infers the states of disconnected AUVs under unstable link interruptions from historical data and incorporates into TPM’s updates enhancing cooperation efficiency. Finally, we design the scheme based on multi-agent deep deterministic policy gradient (MADDPG), improving the efficiency of multi-AUV collaborative search. The effectiveness of the proposed scheme is verified through simulations.