LLaDA2.2 adds delete and insert to diffusion LLMs for 128K agent workflows

机器之心 · wechat · 2026-07-28

Ant Group’s LLaDA2.2 pushes diffusion language models from static generation into agentic workflows by adding delete and insert to the edit space, not just keep and substitute.

The report describes three big changes: Levenshtein-style editing with four atomic operations, an RL method driven by environment feedback for tool-use and multi-turn trajectories, and native 128K context plus a fixed-capacity block routing design for MoE inference. On benchmarks, LLaDA2.2-flash is roughly on par with comparable autoregressive models on agent tasks, beats them on some tool-use benchmarks, and reaches 1.64x the average BF16 throughput.

Related event: Ant's LLaDA2.2 Brings Diffusion LLMs to Long-Context Agents(2 posts)→

Original post →

More from coding & agent

coding & agent channel →