Stratified Scaling Search for Test-Time in Diffusion Language Models (April 2026)
FreeS³: inference-time procedure maintaining a population of partial denoising trajectories with verifier-based look-ahead and reward-tilted Gibbs distribution — first principled test-time scaling for discrete masked diffusion LMs
About Stratified Scaling Search for Test-Time in Diffusion Language Models (April 2026)
S³ (Stratified Scaling Search) is a test-time scaling method for discrete masked diffusion language models. It improves generation quality by reallocating compute during the denoising process using a verifier-guided search that maintains multiple candidate trajectories, evaluates them with a lightweight reference-free verifier, and selectively resamples promising candidates while preserving diversity. This approximates a reward-tilted sampling distribution that favors higher-quality outputs. Experiments with LLaDA-8B-Instruct on MATH-500, GSM8K, ARC-Challenge, and TruthfulQA show consistent performance improvements, especially on mathematical reasoning tasks, without modifying the underlying model or decoding schedule.
Key Features
Pros & Cons
- Improves performance without any retraining or fine-tuning
- Principled approach grounded in classical search and reward-tilted distributions
- Generalizable across multiple reasoning benchmarks
- Leaves the underlying model and decoding schedule unchanged
- Provides a practical mechanism for test-time scaling in diffusion LMs
- Requires additional inference-time compute for search and evaluation
- Effectiveness depends on the quality of the lightweight verifier
- Currently demonstrated only on discrete masked diffusion LMs (e.g., LLaDA-8B-Instruct)
- Still a research method; not yet validated across all model types or scales