Abstract
Diffusion language models (dLLMs) demonstrate strong performance and high efficiency across general tasks, yet their block-parallel decoding process makes them susceptible to error accumulation in multi-turn, long-horizon agentic settings. LLaDA2.1 partially mitigates this via token-to-token (T2T) editing, but its fixed-length substitution mechanism remains a critical bottleneck in agentic workflows. To address this, we present LLaDA2.2, which equips dLLMs with flexible Levenshtein editing through four primitive edit operations—KEEP, SUBSTITUTE, DELETE, and INSERT—with training labels derived via longest common subsequence (LCS) alignment between intermediate drafts and ground-truth sequences. We further propose L-EBPO, an agentic RL algorithm that optimizes editing decisions based on environmental feedback. For practical long-horizon deployment, LLaDA2.2 extends the context window to 128K tokens and introduces a block-routing mechanism to mitigate MoE inference overhead. Empirical results demonstrate that LLaDA2.2 achieves performance competitive with autoregressive baselines on long-horizon agentic benchmarks.