Reproduce “Prism: Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models” on OpenResearch

About this paper

Jinbin Bai, Yixuan Li, Yuchen Zhu, Yi Xin, Qingyu Shi, Aosong Feng, Xiaohong Liu, Molei Tao and 3 more

PRISM introduces an efficient test-time scaling framework for discrete diffusion language models (dLLMs), integrating hierarchical search, partial remasking, and self-verification to achieve substantial performance improvements in reasoning tasks with significantly reduced computational cost compared to naive scaling methods.

Key insights

• Discrete diffusion models can achieve robust reasoning capabilities through adaptive test-time scaling strategies that leverage their parallel decoding for internal self-correction and guided exploration.
• The dLLM itself can effectively serve as an internal verifier for its intermediate outputs, providing crucial signals for pruning and re-allocation of computational resources without requiring external models or significant added memory.
• A hierarchical search approach combined with targeted re-exploration of uncertain solution segments (via partial remasking) offers a superior efficiency-accuracy trade-off compared to uniform expansion of all candidate trajectories.

Summary by alphaXiv.