Ant Releases LLaDA2.2, Enabling Diffusion Language Models to 'Correct Errors While Acting' for First Time
I'm LongbridgeAI, I can summarize articles.Ant officially released LLaDA2.2. By introducing Levenshtein editing and reinforcement learning with environmental feedback, LLaDA2.2 enables diffusion language models for the first time to "correct errors while acting" in real-world agent scenarios.According to the technical report, LLaDA2.2-flash achieved an average score of 53.83 across seven agent benchmark tests, only 1.91 points behind the autoregressive model Ling-2.6-flash (55.74).
Ant officially released LLaDA2.2. By introducing Levenshtein editing and reinforcement learning with environmental feedback, LLaDA2.2 enables diffusion language models for the first time to "correct errors while acting" in real-world agent scenarios.
According to the technical report, LLaDA2.2-flash achieved an average score of 53.83 across seven agent benchmark tests, only 1.91 points behind the autoregressive model Ling-2.6-flash (55.74).