EVA: A Single Linear Layer Brings VAEs Back to Sequence Generation

University of Tokyo researchers propose EVA, which adds a linear layer predicting the next prior on top of a causal decoder to approximate autoregressive priors, enabling VAE-based sequence generation that rivals diffusion quality.

2026-10-06 ~ 2026-10-07 · 2 related posts

• EVA: one linear layer turns VAEs into self-prior generative models rivaling diffusion — Kaede Shiohara · 2026-10-06
• EVA from UTokyo revives VAEs for sequence generation with a single extra linear layer — RichmanRonald · 2026-10-07