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: An ICML 2026 paper proposes PRISM, a three-layer inference-time search using parallel generation, local rollback rewriting and self-scored selection, raising LLaDA-8B-Instruct accuracy on GSM8K from 67.58% to 85.30%.