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

A collaborative research team from multiple institutions has introduced S³ (Stratified Scaling Search), a novel method to significantly improve the output quality of diffusion language models (DLMs) at test time without modifying the model's weights. The core problem they address is the limitation of naive "best-of-K" sampling, where simply generating multiple final outputs fails because the model's high-probability outputs are often misaligned with high-quality, correct answers. S³ reallocates computational budget intelligently during the text generation process itself.

Instead of waiting for a final answer, S³ treats the denoising process—where the model gradually refines text from noise—as a search space. At each step, it expands multiple candidate trajectories, evaluates them with a fast, reference-free verifier, and strategically prunes and resamples to focus compute on the most promising paths while maintaining diversity. This approximates sampling from a reward-tilted distribution, favoring higher-quality outputs. In experiments, applying S³ to the LLaDA-8B-Instruct model yielded consistent gains across reasoning benchmarks, with particularly strong improvements on mathematical tasks like MATH-500 and GSM8K, while leaving the base model and its decoding schedule completely unchanged.

The results demonstrate that classical search techniques applied to the internal states of a generative model provide a powerful and practical lever for "test-time scaling." This approach allows a fixed model to produce better results when given more inference-time compute, moving beyond the plateau of simple output reranking. It highlights an under-explored axis for improving AI systems: optimizing the decision-making process during generation itself, which could be applied to various autoregressive and diffusion-based models to enhance reasoning, factuality, and code generation.