LLaDA2.2 brings diffusion language models into long-horizon agent tasks
量子位 · wechat · 2026-07-28
Alibaba’s inclusionAI pushes diffusion language models into long-horizon agent tasks
The inclusionAI team under Ant Group released LLaDA2.2, a trillion-parameter MoE diffusion language model with native 128K context support. The team says it is the first large-scale agentic diffusion model to enter long-horizon agent tasks.
Key pieces of the system include:
- Levenshtein editing inside diffusion decoding, allowing KEEP / SUBSTITUTE / DELETE / INSERT operations
- L-EBPO, which turns multi-turn editing decisions into a reinforcement-learning problem driven by environment feedback
- BlockRouting for MoE inference, reducing HBM traffic and communication overhead
On SWE-bench Verified, adding Levenshtein editing alone improved scores from 35.8 to 44.4. Across seven agent benchmarks, LLaDA2.2-flash came close to Ling-2.6-flash, trailing by less than 2 points on average, while outperforming it on τ²-Bench, PinchBench, and MCP-Atlas. The report also claims 1.64x higher average BF16 throughput, with another 18.6% gain under FP8.
Related event: Ant's LLaDA2.2 Brings Diffusion LLMs to Long-Context Agents(2 posts)→
More from Models
- Anthropic’s Opus 4.8 beats version 5 for non-coding work, despite weaker benchmarks — burkov · 2026-07-28
- Thinking Machines releases Inkling, a 975B open-weights multimodal model with 1M context — paraschopra · 2026-07-28
- Users Say Opus 5 Looks Better After Repeated Bug-Fix and Feature Requests — Rasmic · 2026-07-28
- A user says GPT-5.4, Opus 4.6, and Kimi k3 already cover most needs — haider1 · 2026-07-28
- Opus 5 looks perfect on benchmarks, but users say real-world quality is inconsistent — yunta_tsai · 2026-07-28
- ChatGPT vs Gemini debate centers on how each handles scandal coverage — Partygoer69420 · 2026-07-28