ResearchBenchmarksMulti-task Language Understanding on MMLU-Pro 5-shot (test)Multi-task Language Understanding on MMLU-Pro 5-shot (test)Follow54.56Throughputd2Cache27.821634.763341.70548.6467Sep 27, 2025Evaluation resultsMethodMethodLinksThroughputSpeedup RatioLatency (s)Scored2CacheModel=Dream-Inst, Few-...Model=Dream-Inst, Few-shot=5-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.0954.5625.46.4246.1d2CacheModel=LLaDA-Inst, Few-...Model=LLaDA-Inst, Few-shot=5-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.0928.8516.413.2435.1