New S³ technique enables diffusion language models to generate higher-quality outputs during inference by strategically reallocating compute across the denoising process

S³ (Stratified Scaling Search) improves test-time scaling by moving beyond limitations of naive best-of-K sampling, which repeatedly draws from the same distribution misaligned with quality outputs

The method expands multiple candidate trajectories at each denoising step and uses a lightweight reference-free verifier to evaluate and selectively resample promising candidates

S³ approximates a reward-tilted sampling distribution that favors higher-quality outputs while maintaining diversity in the search frontier, enabling better generation without additional model training