OpenPrint 20260728.0007v1MethodReleased: July 22, 202613 Views

LLaDA2.2: Enabling Agentic Diffusion Language Models via Levenshtein Editing

Tiwei Bie|Jiansheng Cai|Maosong Cao|Xiang Cao|Binbin Chen|Fuyuan Chen|Kun Chen|Yuqi Ding|Lun Du|Zhuocheng Gong|Yanmei Gu|Kaiyuan Guan|Xiaoxu Guo|Kairong Han|Zenan Huang|Zijin Huang|Zhenzhong Lan|Chengxi Li|Jianguo Li|Zhe Li|Ruiqi Liang|Xinyao Lin|Huabin Liu|Lin Liu|Xinyi Liu|Xuejie Liu|Guoshan Lu|Yuan Lu|Yuxin Ma|Chenhui Mao|Xiaojing Qi|Kaida Qiu|Xiaoyu Shi|Yiding Tian|Huanyu Wang|Qi Wang|Rui Wang|Shuaidi Wang|Lanning Wei|Tao Wu|Siyang Xiao|Yipeng Xing|Chenkai Xu|Song Yan|Ying Yan|Jingyi Yang|Yichun Yang|Liangyu Zha|Tianze Zhang|Yifan Zhang|Junbo Zhao|Wenqian Zhao|Da Zheng|Jianyuan Zhong|Jiahao Zhou|Junlin Zhou|Tianyu Zhou|Liwang Zhu|Yihong Zhuang

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.

Keywords

diffusion language modelsagentic AILevenshtein editingreinforcement learninglong-context modelsmixture-of-expertsparallel decoding

External Source

This is an externally sourced paper. It was originally published independently.
LLaDA2.2: Enabling Agentic Diffusion Language Models via Levenshtein Editing | OpenPrint 20260728.0007v1 — CSPaper