1
FASTER: Rethinking Real-Time Flow VLAs
聚焦实时流VLA架构创新，重新思考并加速Flow VLA推理效率，适合AI研究者。
arXiv:2603.19199v3 Announce Type: replace-cross Abstract: Real-time execution is crucial for deploying Vision-Language-Action (VLA) models in the phys…
2
Roll Out and Roll Back: Diffusion LLMs are Their Own Efficiency Teachers
扩散LLM无需外部教师，通过“展开回退”策略自我提升推理效率，开辟模型加速新方向。
arXiv:2605.16941v1 Announce Type: new Abstract: Diffusion Large Language Models (DLLMs) promise fast parallel generation, yet open-source DLLMs still …
3
ReasonCache: Accelerating Large Reasoning Model Serving through KV Cache Sharing
一篇关于通过KV缓存共享加速大型推理模型服务的系统优化论文，解决推理内存瓶颈问题。
arXiv:2507.21433v3 Announce Type: replace-cross Abstract: Large Reasoning Models (LRMs) are becoming integral to many AI inference systems, enhancing …
4
BEAM: Binary Expert Activation Masking for Dynamic Routing in MoE
一种新型动态路由方法通过二进制专家激活掩码减少MoE冗余计算，无需重训即可加速推理。
arXiv:2605.14438v1 Announce Type: new Abstract: Mixture-of-Experts (MoE) architectures enhance the efficiency of large language models by activating o…
5
When RL Meets Adaptive Speculative Training: A Unified Training-Serving System
强化学习驱动自适应投机训练，统一训练与推理流程，消除部署延迟，加速大模型服务。
arXiv:2602.06932v4 Announce Type: replace Abstract: Speculative decoding can significantly accelerate LLM serving, yet most deployments today disentan…
6
Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution
开源项目Orthrus-Qwen3实现高达7.8倍前向推理加速，且保证输出分布完全一致，Qwen3模型效率飞跃。
Article URL: https://github.com/chiennv2000/orthrus Comments URL: https://news.ycombinator.com/item?id=48154865 Points: 217 # Comments: 43
7
Dynamic Mixed-Precision Routing for Efficient Multi-step LLM Interaction
提出动态混合精度路由方法，在多步LLM交互中实现高效推理，在精度与效率间取得平衡。
arXiv:2602.02711v2 Announce Type: replace Abstract: Large language models (LLMs) achieve strong performance in long-horizon decision-making tasks thro…