The　model　search　based　on　the　distillation　model　framework　aims　to　train　the　candidate　models　adequately　and　guide　a　correct　evaluation　of　the　architecture.　This　NAS　method　can　easily　obtain　middle-level　monitoring　identification　indicators,　thus　significantly　improving　the　effect.　However,　the　model　search　based　on　distillation　framework　also　has　its　shortcomings.　First,　the　supervision　indicators　differ　greatly　for　various　teacher-student　models,　so　how　to　determine　a　highly　adaptable　supervision　indicator　is　a　very　important　issue.　Second,　different　teacher　models　will　introduce　biases.　Based　on　the　above　problems,　this　paper　proposes　the　following　measures.　Firstly,　this　paper　adopts　a　more　adaptable　supervision　index,　which　can　effectively　solve　the　problem　that　various　teacher-student　models　differ　greatly.　Secondly,　in　order　to　reduce　the　bias　introduced　by　the　teacher　model,　this　paper　adopts　the　largest　teacher　model　as　the　guidance　model　in　the　network　training.　Finally,　this　study　uses　a　reinforcement　learning　algorithm　to　guide　the　search　in　the　internal　network,　and　introduce　more　supervision　quantity,　which　makes　the　supervision　effect　between　layers　more　obvious.　It　can　be　concluded　that　the　above　methods　can　effectively　improve　the　model　performance　and　consistency.　©　2021　SPIE.