$S^3$: Stratified Scaling Search for Test-Time in Diffusion Language Models

Develop an inference engine for diffusion models that uses stratified sampling at test-time to improve generation quality. This is a game-changer for high-fidelity generative tasks that usually require massive prompt engineering.

Suggested repo: stratify-diff

"Better diffusion outputs through smarter test-time scaling."

Estimated effort: 35h