Diffusion Models Outperform Autoregressive AI in Data-Constrained Settings
A Carnegie Mellon University paper found diffusion models are more robust to data repetition with sufficient compute.
A new paper from researchers at Carnegie Mellon University, published on September 22, 2025, addresses a fundamental question of AI model architecture and efficiency. The study found that diffusion models can outperform the more common autoregressive models, the basis for systems like GPT, in data-constrained settings, provided they are trained with sufficient compute.
## The Data Bottleneck
The research is motivated by the growing recognition that the AI field is approaching the limits of high-quality training data available on the internet. The paper sought to answer the question: "how can we trade off more compute for less data?"
Archive Access
This article is older than 24 hours. Create a free account to access our 7-day archive.
Archive Access
This article is older than 24 hours. Create a free account to access our 7-day archive.
Share this article
Related Articles
DeepMind unveils WeatherNext 2 for probabilistic forecasting at NeurIPS
Google DeepMind reveals WeatherNext 2 at NeurIPS, an AI model that forecasts weather 8x faster …
Google DeepMind partners with CFS to control fusion plasma with AI
Google DeepMind partners with Commonwealth Fusion Systems to use AI for stabilizing plasma in tokamak …
TwelveLabs Launches Marengo 3.0 on Amazon Bedrock
TwelveLabs launched Marengo 3.0 on Amazon Bedrock. The new AI model improves video understanding by …