LLaDA2.2 pushes diffusion LLMs into agentic work with 128K context and self-editing
alifcoder · x · 2026-07-25
LLaDA2.2 is pitched as a large-scale agentic diffusion model that goes beyond text generation into planning, tool use, and self-correction on long multi-turn tasks.
- It uses Levenshtein Editing with KEEP / SUBSTITUTE / DELETE / INSERT so the model can rewrite its own sequence instead of freezing early mistakes into context.
- L-EBPO trains the model to decide when and where to correct errors using environmental feedback, aiming to reduce long-horizon drift and cascading failures.
- The system adds native 128K context and Block Routing for more controllable long-context inference.
- On 7 agentic benchmarks, LLaDA2.2-flash averages 53.83 vs. 55.74 for Ling-2.6-flash, while delivering 1.64× average BF16 throughput across 11 workloads.
More from Models
- FrontierML benchmark shows open models still far behind closed ones on end-to-end ML work — scaling01 · 2026-10-27
- Mirendil debuts Fern, its first from-scratch model with up to 10M-token context — shayan_ · 2026-10-27
- Week 3 of "Fable 5.5 launches this week" hype circles: no new frontier models this October — daniel_mac8 · 2026-10-27
- 125 poisoned samples backdoored Qwen2.5-7B — and open-weight supply chain risks debated — lxrjl · 2026-10-27
- Speculative leaderboard shows "Gemini 4 Argon" vs "GPT-6.1 Sol"; observer calls MiniMax parity suspicious — teortaxesTex · 2026-10-27
- Cohere sandbox: Qwen model tried to upload probe file to Beijing cloud server unprompted — simonguozirui · 2026-10-27