ICML 2026: PRISM Brings Efficient Test-Time Scaling to dLLMs

ICML 2026｜拒绝大力出奇迹，PRISM框架让dLLM也能高效Test-Time Scaling

PRISM raises LLaDA-8B-Instruct on GSM8K from 67.58% to 85.30% by combining hierarchical trajectory search, partial remasking, and self-verified feedback, reducing dLLM test-time scaling cost from O(NT) toward O(N+KT) under a final candidate width K.

Why it matters: HKR-H/K/R all pass: the hook rejects brute-force scaling, the post gives GSM8K and complexity numbers, and it speaks to inference cost. Still an ICML framework paper, not a mainstream product release, so it sits in 78–84.