ResearchBenchmarksCode Generation on HumanEval (Acc and TPS)Code Generation on HumanEval (Acc and TPS)Follow39.6Accuracyd3LLM36.4837.2938.138.91Apr 1, 2026Evaluation resultsMethodMethodLinksAccuracyTokens Per Second (TPS)d3LLMBackbone=LLaDA-InstructBackbone=LLaDA-Instruct2026.0439.65.95dParallelBackbone=LLaDA-InstructBackbone=LLaDA-Instruct2026.04394.93LLaDABackbone=LLaDA-InstructBackbone=LLaDA-Instruct2026.0438.31Fast-dLLMBackbone=LLaDA-InstructBackbone=LLaDA-Instruct2026.0437.82.56TRIMSBackbone=LLaDA-InstructBackbone=LLaDA-Instruct2026.0437.23.98D2FBackbone=LLaDA-InstructBackbone=LLaDA-Instruct2026.0436.62.69