In real-world scenarios, many-objective optimization problems (MaOPs) are often affected by constraints, redundant information, or high non-linearity, resulting in irregular shapes of their Pareto front (PF). In addressing such issues, evolutionary algorithms face two major challenges: 1) the selection pressure of solutions significantly weakens; 2) the distribution of Pareto-optimal solutions is uneven. To tackle these difficulties, this paper proposes an adaptive reference vector enhanced generalized Pareto dominance many-objective evolutionary algorithm (ARP-MaOEA). This algorithm ensures population convergence to the true PF through a generalized Pareto dominance mechanism while maintaining population diversity by enhancing individual selection with reference vectors. To improve the algorithm’s adaptability and robustness to PF with different shapes, this paper introduces an adaptive reference vector strategy that automatically eliminates ineffective reference vectors based on population evolutionary information and dynamically adds new reference vectors based on non-dominated solutions. To validate the effectiveness of the ARP-MaOEA, a series of simulation experiments are conducted on MaOPs with different PF shapes. The experimental results show that the ARP-MaOEA outperforms other comparative algorithms in handling irregular MaOPs, demonstrating its advantages in solving such problems.