Ambient Discrete Diffusion: Using the Wrong Data at the Right Time for Data Efficient Learning

Name: Ambient Discrete Diffusion: Using the Wrong Data at the Right Time for Data Efficient Learning
Uploaded: 2026-10-14T23:57:15.187600+00:00
Duration: 344 s
Description: RefineMix is a framework for training discrete diffusion models with very little data by strategically using out-of-distribution data at specific noise levels. It exploits the fact that at low noise, different domains have distinct data supports, allowing the model to learn from both in-domain and source data without biasing the final sampling distribution.

RefineMix is a framework for training discrete diffusion models with very little data by strategically using out-of-distribution data at specific noise levels. It exploits the fact that at low noise, different domains have distinct data supports, allowing the model to learn from both in-domain and source data without biasing the final sampling distribution.

Open the full episode