World's First Agentic Diffusion Model: Self-Corrects While Acting, 128K Context Matches Autoregressive Models

Ant Group's inclusionAI team has open-sourced LLaDA 2.2, a 100-billion-parameter MoE diffusion language model that supports 128K native context. It is the first large-scale agentic diffusion model, enabling the model to generate and correct its outputs simultaneously. The model uses a Levenshtein editing paradigm with four atomic operations (KEEP, SUBSTITUTE, DELETE, INSERT) to dynamically modify generated content. It also introduces L-EBPO (Levenshtein Editing Evidence Lower Bound Policy Optimization) to treat multi-turn editing decisions as reinforcement learning problems based on environmental feedback. To handle long contexts, the model employs BlockRouting, which selects a fixed expert pool at the block level before token-level routing, reducing HBM traffic and communication costs. On seven agent benchmarks, LLaDA 2.2-flash scored 53.83 vs. Ling-2.6-flash's 55.74, demonstrating comparable performance, while achieving 1.64x average throughput in BF16 and additional 18.6% improvement with FP8 quantization. This proves that diffusion models can match autoregressive models in agent tasks and will coexist in future agent systems.