Multi-agent reinforcement learning faces dual challenges of robust exploration and efficient collaboration.Existing algorithms predominantly employ fixed or state-independent exploration mechanisms,neglecting the differentiated exploration needs across diverse decision contexts.This hinders the maintenance of an effective balance between exploration and exploitation.Simultaneously,they typically adopt a flattened information aggregation approach when handling agent relationships,overlooking their inherent interactive topological structure.To address these challenges,this paper proposes a multi-agent deep deterministic policy gradient algorithm based on state-adaptive maximum entropy and graph attention networks.First,a maximum entropy reinforcement learning framework is constructed.By designing a state-dependent temperature coefficient as an adaptive weight for policy entropy,agents dynamically adjust exploration intensity according to real-time interaction states and exploration needs,achieving adaptive exploration-exploitation balance.Second,a graph attention network explicitly models the interactive topology among agents.By distinguishing the importance of different neighbors through attention mechanisms,it generates more expressive state representations to enhance inter-agent collaboration.Experimental results in multi-agent simulation environments demonstrate that the proposed algorithm exhibits significant superiority in both learning efficiency and final collaborative performance.