開放模型與代理
LLaDA 2.2 Lets Diffusion Models Add and Delete Tokens Mid-Generation, Targeting Iteratively Correctable Agent Loops
inclusionAI has open-sourced the 100B-class LLaDA2.2-flash, using Levenshtein Editing to let parallel decoding insert, delete, and replace content. The model outperforms the company’s own autoregressive baseline in some agent benchmarks and throughput tests, but still trails in code repair and function-calling accuracy, while serving support remains incomplete.
Ant Group subsidiary inclusionAI officially released LLaDA2.2-flash on July 27, aiming to address a structural weakness that arises when diffusion language models are used as agents. Traditional autoregressive models commit to tokens from left to right. Although conventional text diffusion models can denoise in parallel, they struggle to change the length of an existing sequence after an environment returns an error. LLaDA 2.2 introduces Levenshtein Editing, adding `DELETE` and `INSERT` operations alongside retention and replacement. These operations let the model remove redundant spans or create new editable positions, enabling it to rewrite a draft’s structure during generation.
The model uses a 100B non-embedding-parameter MoE architecture with 32 layers and a 128K context window. The team also proposes Block Routing, which constrains expert activation at the diffusion-block level. For training, it introduces L-EBPO, which applies rewards from agent environments to block-level policy optimization. The goal is to teach the model to correct failed actions during multi-turn tool use, rather than merely improve single-turn text quality.
Using a Claude Code scaffold, a fixed 128K context, and scores averaged across five evaluations, inclusionAI reports 49.28 on SWE-bench Verified, 60.78 on BFCL-V4, and 80.33 on τ²-Bench. Compared with its own autoregressive Ling-2.6-flash, LLaDA 2.2 scores slightly higher on τ²-Bench, MCP-Atlas, and some agent tasks, but trails on SWE-bench Verified, Multilingual, and BFCL-V4. The throughput gap is more pronounced: LLaDA 2.2 reaches 703.82 tokens/s on BFCL-V4, versus 331.5 for Ling. However, the comparison does not fully disclose hardware, parallelism, or end-to-end latency, and some Ling scores were obtained with a different scaffold, so the results should not be treated as a strict like-for-like contest.
The Apache 2.0-licensed weights and technical report are now public. Transformers requires `trust_remote_code` to be enabled. The model card also recommends SGLang for long-context workloads, while the deployment section says support is still in preparation. Engineering teams should next watch for an official serving backend, KV/prefill costs, structured-output reliability, and whether independent testing can reproduce the model’s multi-turn self-correction capabilities. Previous research has found that diffusion agents can become trapped in retry loops and generate malformed JSON. Whether LLaDA 2.2 truly overcomes these limitations still requires direct comparative testing.