FLUID Framework Bridges Autoregressive and Diffusion Models for Efficient Text Generation
Researchers propose FLUID, a framework that adapts autoregressive (AR) backbones to diffusion models for efficient parallel text generation. By enforcing Strictly Causal Alignment, FLUID enables initialization from standard GPT-style checkpoints, avoiding pre-training from scratch. The Elastic Horizons mechanism dynamically adjusts denoising strides based on local information density. Experiments show state-of-the-art performance with orders of magnitude reduction in training costs.
Key facts
- FLUID adapts AR backbones to diffusion models
- Strictly Causal Alignment enables GPT checkpoint reuse
- Elastic Horizons dynamically modulates denoising strides
- Achieves state-of-the-art performance
- Reduces training costs by orders of magnitude
- Published on arXiv as 2605.27387
- Cross announcement type
Entities
Institutions
- arXiv