Recently, incorporating a learned dynamic model in generating imagined data has been shown to be an effective way to reduce sample-complexity of model-free RL. Such model-free/model-based hybrid approaches usually require rolling out the dynamic model a fixed number of steps into the future. We argue that such fixed rollout is problematic for several reasons. We propose a simple adaptive rollout algorithm to improve the model-based component of these approaches and conduct experiment on CartPole task to evaluate the effects of adaptive rollout.