ResearchBenchmarksCode Generation on HumanEval dLLM-VarCode Generation on HumanEval dLLM-VarFollow31.7Pass@1LR-DLLM17.1420.9224.728.48Feb 7, 2026Evaluation resultsMethodMethodLinksPass@1LR-DLLMBackbone=LLADA-8B-Inst...Backbone=LLADA-8B-Instruct, Inference mode=training-free inference2026.0231.7dLLM-VarBackbone=LLADA-8B-Inst...Backbone=LLADA-8B-Instruct, Inference mode=trained checkpoint with block-diffusion inference2026.0217.7