Posterior Collapse

Posterior collapse studies why latent-variable models, especially variational autoencoders, learn representations that carry little or no input information. Researchers address it with changes to divergence penalties and latent noise, teacher-guided alignment constraints, initialization strategies, and objectives that encourage informative codes; spectral and entropy analyses help explain when latent dimensions become inactive. The problem appears in image and graph generation as well as language models, where a powerful decoder may explain outputs without using latent variables. A central challenge is that preventing collapse or retaining more latent information does not always improve reconstruction or downstream performance, so methods need diagnostics tied to the task and model.