Extends variational active inference to planning during learning by assigning factor-specific priors to avoid losing information gains; uses Bethe-augmented SP-MP to incorporate epistemic priors and learn model parameters, demonstrated on grid POMDPs where learning-enabled planners find safer routes and reach goals reliably.
Semantic grounding of variable names lets LLMs perform data-efficient numerical predictions with little task-specific training, matching or surpassing traditional models across many datasets; they combine domain knowledge with in-context examples in a Bayesian-like manner, using priors when needed and adjusting to contradicted examples.
A researcher describes pre-training a new VAE variant, Primal VAE, with compact latent channels and specific down/up-sampling dynamics, using LPIPS and MAE now with plans for a critic, and applying EQVAE-inspired latent flips and rotations to ease diffusion.
The paper adds a Rényi-divergence regularization to variational flow-based Bayesian calibration to improve multimodal posterior coverage, showing partial success in reducing mode collapse but noting trade-offs and the absence of a general solution.
A soft ontology prior is injected into neuro-symbolic learning to mitigate concept-identification shortcuts, by embedding the prior into a probabilistic circuit and keeping hard axioms in the data-generating process; this approach reshapes shortcut mass without losing the concept space, and its behavior is analyzed and validated on driving benchmarks and controlled tasks.
A step-by-step guide for using DENO’s MiniMax H3 2K video upscaler to convert MP4 videos to 2K by encoding to VAE latents, upscaling, and refining with specialized models, while managing GPU memory with spatial tiling and preserving original audio; includes model links and required components.
This paper studies audio representation autoencoders (AuRA) by pairing frozen SSL encoders with trained decoders to assess their effectiveness for text-to-audio generation, comparing SSL-based and baseline latent spaces, and finding that SSL-space generation can match or surpass baseline performance even when reconstruction quality differs.
OpenAI announces that its models have proven the local smoothing conjecture in all dimensions, which, if validated, would also settle the Kakeya and Stein restriction conjectures. The proof is machine-checked in Lean, formally published on GitHub with a human-readable explanation, and centers on a new auxiliary partial differential equation as part of a broader release of AI-generated mathematical preprints.
The post reports early results from an overnight experiment with Primal VAE training on MOSS Image, using a 2x1 configuration and 64x spatial compression to achieve 4096x4096 outputs on the same compute as 1024x1024. The author questions why some outputs, like an eye, look smudged or not realistic, seeking understanding rather than critique.
OpenAI and collaborators release results on wave-equation smoothing, Kakeya/Stein-type conjectures, and a counterexample to ETH with a subexponential CNF algorithm; an independent group will review the proof and materials are being made public.