全球首个Agentic扩散模型来了：边行动边纠错，128K上下文追平自回归
Merged summary
TL;DR - InclusionAI released LLaDA2.2, an open-source MoE diffusion language model designed for long-horizon agent tasks. It approaches an autoregressive baseline on seven agent benchmarks while delivering higher throughput and native 128K context.
- Levenshtein-style KEEP, SUBSTITUTE, DELETE, and INSERT operations let the model revise block-parallel generations during denoising.
- L-EBPO reinforcement learning uses environmental feedback to optimize editing decisions across multi-turn interactions.
- BlockRouting limits each block’s expert pool to reduce memory traffic and communication costs at long context lengths.
- LLaDA2.2-flash averaged 53.83 versus Ling-2.6-flash’s 55.74 across seven agent benchmarks, with 1.64× average BF16 throughput across 11 workloads.
Sources (1)
全球首个Agentic扩散模型来了：边行动边纠错，128K上下文追平自回归
TL;DR - InclusionAI released LLaDA2.2, an open-source MoE diffusion language model designed for long-horizon agent tasks. It approaches an autoregressive baseline on seven agent benchmarks while delivering higher throughput and native 128K context.
- Levenshtein-style KEEP, SUBSTITUTE, DELETE, and INSERT operations let the model revise block-parallel generations during denoising.
- L-EBPO reinforcement learning uses environmental feedback to optimize editing decisions across multi-turn interactions.
- BlockRouting limits each block’s expert pool to reduce memory traffic and communication costs at long context lengths.
- LLaDA2.2-flash averaged 53.83 versus Ling-2.6-flash’s 55.74 across seven agent benchmarks, with 1.64× average BF16 throughput across 11 workloads.