Adapt Diffusion Evolution to Open-Ended Evolution

Determine how to adapt the Diffusion Evolution algorithm, which relies on a finite denoising schedule with a predetermined number of sampling steps, to support open-ended evolution without a fixed termination, while preserving diversity and effective optimization over indefinite generational timescales.

Background

The paper establishes a mathematical and algorithmic equivalence between diffusion models and evolutionary algorithms, introducing Diffusion Evolution as an inverse denoising procedure aligned with evolutionary selection and mutation.

A key difference highlighted by the authors is that diffusion models operate with a finite number of sampling steps controlled by schedules, whereas biological evolution is inherently open-ended. Extending Diffusion Evolution to an open-ended setting would bridge this gap and more closely emulate natural evolutionary dynamics.