As　the　core　of　the　rail　freight　flow　organization　process,　train　formation　problem　(TFP)　has　attracted　much　attention.　In　Chinese　practice,　car　flow　routing,　TFP,　and　train　routing　are　usually　optimized　sequentially　to　reduce　the　complexity　of　computation,　which　may　result　in　a　local　optimum,　and　even　no　feasible　solution.　To　address　this　issue,　this　paper　studies　the　integrated　optimization　of　the　three　sub　problems　with　aims　to　minimize　the　total　cost　of　transportation　cost,　accumulation　cost,　and　classification　cost.　An　integer　linear　arc-based　model　incorporating　the　unitary　and　in-tree　rules　of　a　shipment　is　first　formulated　and　solved　by　the　state-of-the-art　solver　GUROBI.　Since　GUROBI　can＇t　deal　with　the　large　test　cases,　a　path-based　model　is　built　and　solved　by　a　bespoken　two-phase　algorithm.　The　first　phase　of　the　algorithm　is　Benders-and-Price　approach　that　combines　Benders　decomposition　and　column　generation,　and　the　second　phase　is　to　solve　the　arc-based　model　with　some　variables　fixed　as　the　corresponding　values　fetched　from　the　first　phase.　The　results　show　that　the　proposed　algorithm　outperforms　GUROBI,　and　the　acceleration　techniques,　i.e.,　trust　region　and　Pareto-optimal　cuts,　can　improve　the　convergence　efficiency　of　Benders　decomposition　significantly.