Against the backdrop of innovation fault-tolerance, government compensation for failed generic technological innovation projects is beneficial to stimulate reinnovation behaviour. However, considering the information asymmetry, a collusion tendency exists between the compensated party and the evaluator during the process of compensation. To prompt the government to build collusion-proof mechanisms to reduce collusion loss, the evolutionary game method was used to build replicated dynamic equations and a Jacobian matrix of both sides based on the information topology between the conspirators and regulator. Through the evolutionary equilibrium analysis and numerical simulation, evolutionary stability strategies (ESS) under different topological relationship information (non-intersect, partial intersect, and inclusive type) were found. Results show that the collusion behaviour can be effectively restrained when the government is unaware of collusion information, the net defense income is positive, and the penalty threshold is the product of the net collusive income and the ratio of the collusive and regulatory information. With the increasing amount of collusive information available to the government, the conspirators tend to adhere to moral principles subject to strict regulation. In addition, the moderating effect of penalty factor is positively correlated with the private information possessed by both sides. The conclusion is beneficial to provide theoretical support for optimizing the governmentled compensation mechanism for failed generic technological innovation projects.