The　X-architecture　Steiner　Minimum　Tree　(XSMT)　is　the　best　connection　model　of　non-Manhattan　multi-terminal　nets　in　global　routing,　and　it　is　an　NP-hard　problem.　Particle　Swarm　Optimization　(PSO),　with　its　efficient　searching　ability　and　self-organizing　ability,　has　become　a　powerful　tool　for　constructing　the　XSMT.　However,　PSO　is　prone　to　fall　into　the　local　optimum　due　to　its　excessive　exploitation　intensity.　To　keep　a　smooth　trade-off　between　exploitation　and　exploration　capabilities　of　PSO,　maintain　the　diversity　of　the　population,　and　obtain　a　better　solution,　this　paper　proposes　an　XSMT　algorithm　based　on　Competitive　Swarm　Optimizer　(called　CSO-XSMT).　The　algorithm　utilizes　the　methods　of　pairwise　competition　and　roulette　wheel　selection　to　randomly　select　the　learning　objects　of　particles　so　as　to　enhance　the　exploration　ability　of　the　population　and　improve　the　algorithm　performance.　Meanwhile,　to　further　reduce　the　wirelength　of　the　Steiner　tree,　a　refine　strategy　based　on　sharing　edges　is　proposed,　which　adjusts　the　Steiner　tree　obtained　by　CSO　to　improve　the　quality　of　the　final　routing　tree.　Experimental　results　show　that　compared　with　other　Steiner　tree　construction　algorithms,　the　proposed　algorithm　has　better　wirelength　optimization　capability　and　superior　stability.　©　2021,　Springer　Nature　Switzerland　AG.