Variational Learning for Insertion-based Generation

This paper introduces the Insertion Process (IP), a stochastic generative model that utilizes a probabilistic framework for learning insertion order in variable-length insertion tasks. By formalizing a bijective correspondence between insertion trajectories and permutations, IP enables a reparameterization of data likelihood and supports adaptive insertion orders, significantly enhancing modeling quality and generalization in applications like goal-conditioned planning and molecular string generation. This advancement is crucial for practitioners as it allows for more flexible and efficient generation methods that go beyond traditional left-to-right autoregressive models.