Open-domain dialogue generation is an important research area and is drawing more and more attention in the past years. Nowadays, the vast amount of conversational corpora and the popularity of seq2seq models have made it possible to design data-driven, end-to-end trainable dialogue systems without verbose handcrafted rules. However, vanilla seq2seq models stochastical variations only at the token level, seducing the system to gain immediate short rewards and neglect the long-term structure. One way of attenuating this problem is by introducing “latent variables”, which stand for high-level sentence representations to help guide the generating process. This talk explains the application of latent variable models on dialogue generation and two main challenges: uninterpretability of latent variables and difficulty of training. We propose some solving strategies to these challenges respectively and validate the effectiveness.