多模态大模型解码新思路：用语义锚点驱动聚类生成，让长文本输出更连贯，EMNLP 2026 已收录，值得跟进。
arXiv:2608.22367v1 Announce Type: new Abstract: Diffusion multimodal large language models (dMLLMs) frequently produce long-form outputs marred by sem…
Kog专攻现有数据中心GPU的极致推理提速，演示单请求极速解码，还拿下200条销售线索，值得关注。
The idea that GPUs are poorly suited for agentic workflows may be a misconception, according to French startup Kog.
无需额外训练，通过未来价值引导解码就能让基础大模型找到正确推理路径，这篇论文提出了一种高效的采样方法。
arXiv:2605.02427v3 Announce Type: replace Abstract: A recurring pattern in "reasoning without training" is that base LLMs already assign non-trivial p…
DLLM解码计算并行化存在严重低效，只有少数token可解码，此文精准定位瓶颈并提出优化方向。
arXiv:2601.23278v2 Announce Type: replace Abstract: Diffusion Large Language Models (DLLMs) offer a compelling alternative to Auto-Regressive models, …
提出预测预填充方法，让扩散语言模型高效处理超长上下文，解码速度显著提升。
arXiv:2606.10537v1 Announce Type: new Abstract: Diffusion large language models (dLLMs) re-encode the entire prefix at every denoising step, causing r…