Cooperative multi-UAV search requires jointly optimizing wide-area coverage, rapid target discovery, and endurance under sensing and motion constraints. Resolving this coupling enables scalable coordination with high data efficiency and mission reliability. We formulate this problem as a discounted Markov decision process on an occupancy grid with a cellwise Bayesian belief update, yielding a Markov state that couples agent poses with a probabilistic target field. On this belief–MDP we introduce a segment-conditioned latent-intent framework, in which a discrete intent head selects a latent skill every K steps and an intra-segment GRU policy generates per-step control conditioned on the fixed intent; both components are trained end-to-end with proximal updates under a centralized critic. On the $50\times 50$ grid, coverage and discovery convergence times are reduced by up to 48% and 40% relative to a flat actor-critic benchmark, and the aggregated convergence metric improves by about 12% compared with a state-of-the-art hierarchical method. Qualitative analyses further reveal stable spatial sectorization, low path overlap, and fuel-aware patrolling, indicating that segment-conditioned latent intents provide an effective and scalable mechanism for coordinated multi-UAV search.