{"slug": "when-do-gaussian-priors-fail-a-latent-geometry-investigation-of-rl-fine-tuning", "title": "When Do Gaussian Priors Fail? A Latent Geometry Investigation of RL Fine-Tuning", "summary": "A latent-geometry study posted to arXiv (2610.28849v1) finds that the Gaussian prior standard in latent diffusion models structurally limits reinforcement-learning fine-tuning when data are multimodal, because the VAE's KL term taxes mode separation and tightness. On the real alanine-dipeptide Boltzmann distribution, a Gaussian-prior VAE collapsed at beta_c = 0.224 while a mode-tracked Gaussian-mixture prior collapsed at 0.00316 — a 71x gap from the same pipeline — with the authors tracing the effect to the geometric mediator d_min/(2 sigma_0) and confirming full mediation through controlled interventions (prior-label coefficient 0.995x, 95% CI [0.79, 1.25]). The authors conclude a non-Gaussian prior is necessary precisely when latent geometry would fall into overlap and the task is coverage-sensitive, and support the KL pricing structure with a machine-checked theory skeleton.", "body_md": "arXiv:2610.28849v1 Announce Type: new \nAbstract: 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.\n  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.", "url": "https://wpnews.pro/news/when-do-gaussian-priors-fail-a-latent-geometry-investigation-of-rl-fine-tuning", "canonical_source": "https://arxiv.org/abs/2610.28849", "published_at": "2026-10-22 04:00:00+00:00", "updated_at": "2026-10-22 04:14:11.350219+00:00", "lang": "en", "topics": ["machine-learning", "generative-ai", "ai-research", "neural-networks"], "entities": ["arXiv", "VAE", "alanine-dipeptide", "Gaussian prior", "Gaussian-mixture prior"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/when-do-gaussian-priors-fail-a-latent-geometry-investigation-of-rl-fine-tuning", "markdown": "https://wpnews.pro/news/when-do-gaussian-priors-fail-a-latent-geometry-investigation-of-rl-fine-tuning.md", "text": "https://wpnews.pro/news/when-do-gaussian-priors-fail-a-latent-geometry-investigation-of-rl-fine-tuning.txt", "jsonld": "https://wpnews.pro/news/when-do-gaussian-priors-fail-a-latent-geometry-investigation-of-rl-fine-tuning.jsonld"}}