LLaDA2.2-flash Brings Diffusion Language Models Into Agentic Tool Use With DELETE/INSERT Editing Tokens and 128K Context
Ant Group's inclusionAI released LLaDA2.2-flash, a 100B non-embedding-parameter MoE diffusion language model that adds Levenshtein Editing — explicit DELETE and INSERT control tokens — so diffusion decoding can restructure a sequence, remove redundant content and open insertion slots during parallel generation rather than only unmasking in place. That editing capability is what makes multi-turn tool use and error correction tractable for a diffusion model, and it's trained with L-EBPO, a block-level policy optimization using agentic environment rewards. Context extends to 128K via Block Routing, which bounds MoE expert activation at the diffusion-block level; notably, the team evaluated the SWE-bench series using the Claude Code scaffold.
↳ Follow the thread