Internal Guidance as a Sampling Method for Tilted Diffusion Models
Reward guidance methods sample an exponential tilt of a base density using a diffusion model trained with only samples from the base distribution.
ProofPaper ↗
Key points
- We propose internal guidance which leaves the output of the base denoiser unmodified and instead changes the point at which it is evaluated.
- For discrete distributions embedded in the hypercube $\{-1,1\}^{d}$, the denoiser of any linearly tilted density is exactly the base denoiser evaluated at a shifted and rescaled input.
- We modify the training algorithm with random diagonal rescalings of the data, so that the denoiser network learns both in- and out-of-distribution evaluations of the denoiser relative to the base distribution.
- A single denoiser network samples binary reward tilts, class labels, and infilling with no fine-tuning and no external guidance term.
Sources (1)
- [1]Internal Guidance as a Sampling Method for Tilted Diffusion ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 02:21 AM
Reward guidance methods sample an exponential tilt of a base density using a diffusion model trained with only samples from the base distribution.
We propose internal guidance which leaves the output of the base denoiser unmodified and instead changes the point at which it is evaluated.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 21, 2026Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation
- Oct 21, 2026DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations
- Oct 20, 2026When Do Gaussian Priors Fail? A Latent Geometry Investigation of RL Fine-Tuning
- Oct 20, 2026Do Diffusion LLMs Scale? Testing Mercury 3 + Decide
- Oct 20, 2026DiffResearch: Accelerating Deep Research with Diffusion Language Models
- Oct 20, 2026Large Flow Language Models