Abstract
Latent diffusion models almost universally place a Gaussian prior on the VAE latent space. For base generation this default is unproblematic; under reinforcement-learning fine-tuning, however, it carries a structural limitation. The standard VAE's KL term does not merely ignore mode structure - it actively taxes separation and tightness. When data are multimodal, the ELBO optimum compresses distinct modes into an overlapping latent. RL fine-tuning amplifies this overlap: the score network is under-resolved on the saddle region between modes, dynamic-to-variational amplification rises sharply, and the collapse boundary beta_c is pushed far above its variational scale. On the real alanine-dipeptide Boltzmann distribution, a Gaussian-prior VAE collapses at beta_c = 0.224 while a mode-tracked Gaussian-mixture prior collapses at 0.00316, a 71x gap from the same pipeline. We trace this gap to a single geometric mediator, d_min/(2 sigma_0), and verify the causal chain through controlled interventions: moving geometry at fixed prior, and switching prior at matched geometry (full mediation; prior-label coefficient 0.995x, 95% CI [0.79, 1.25]). A three-source decomposition localises the amplification to the learned policy (zero ODE-solver contribution, bounded anchor mismatch); the coverage plateau is an unconditional benefit of mode-separating priors that no amount of RL capacity recovers. The remedy is conditional: a non-Gaussian prior is necessary precisely when the latent geometry would fall into overlap and the task is coverage-sensitive. A machine-checked theory skeleton supports the KL pricing structure - a divergent compactness tax under a Gaussian prior versus a finite assignment fee under a mixture - and the position-versus-plateau distinction delimits what is finite-resource from what requires changing the prior.
Abstract (translated)
URL
https://arxiv.org/abs/2610.28849
