Inference Workarounds Move Closer to Hardware 推理优化逼近硬件边界

This research slate is about reducing the cost of moving, sampling, and waiting: MoE communication, VLA action caching, diffusion early exits, optical receivers, THz beamforming, and low-power control-channel monitoring. The strongest papers treat efficiency as a cross-layer problem rather than a model-only problem. 今天的研究线索集中在降低移动、采样和等待的成本：MoE通信、VLA动作缓存、扩散模型提前退出、光接收机、THz波束成形，以及低功耗控制信道监听。最有价值的论文都把效率视为跨层问题，而不是单纯的模型问题。

Look-back window: 7 days · 8 paper(s) 回溯窗口: 7天 · 8篇

AI Accelerators & Systems AI加速器与系统

Hardware-Aware Design Space Exploration for Mixed-Signal Spiking Neural Networks 面向混合信号脉冲神经网络的硬件感知设计空间探索

S. N. Chowdhury, V. Nair, T. Forhad, et al.

arXiv:2607.06456 · 2026-07-07T16:15:03Z

Chowdhury et al. present an open-source hardware-aware framework for exploring mixed-signal SNN designs. It models LIF, Hodgkin-Huxley, and Axon-Hillock neurons plus floating-gate and ReRAM synapses inside a PyTorch training and inference loop. The useful contribution is cross-layer accounting: the framework reports accuracy together with area, power, and quantization sensitivity on benchmarks such as N-MNIST, DVS Gesture, and SHD. Caveat: it is a simulation and design-space tool, not measured silicon. Chowdhury等人提出了一个开源硬件感知框架，用于探索混合信号SNN设计。它在PyTorch训练和推理流程中建模LIF、Hodgkin-Huxley、Axon-Hillock神经元，以及floating-gate和ReRAM突触。其价值在于跨层核算：框架在N-MNIST、DVS Gesture和SHD等基准上同时报告准确率、面积、功耗和量化敏感度。需要注意的是，这是一套仿真和设计空间工具，并非实测硅结果。

UBEP Re-architects Expert Parallelism for Production Superpods UBEP为生产级Superpod重构专家并行通信

Y. Liu, C. Liu, S. Shen, et al.

arXiv:2607.06202 · 2026-07-07T12:25:16Z

UBEP targets the All-to-All communication path in MoE inference on high-bandwidth superpods such as NVIDIA NVL72/576 and Huawei CloudMatrix384. The paper attacks three bottlenecks: serialized BSP-style phases, synchronization overhead, and distance-agnostic token scheduling. It reports up to 52.4% lower All-to-All latency and up to 11.1% lower MoE inference time per output token. The result is important because sparse-model serving is increasingly limited by communication behavior, not just accelerator FLOPs. UBEP面向NVIDIA NVL72/576和Huawei CloudMatrix384等高带宽Superpod上的MoE推理All-to-All通信路径。论文针对三个瓶颈：BSP式阶段串行化、同步开销，以及不考虑距离的token调度。作者报告All-to-All延迟最高降低52.4%，MoE推理的每输出token时间最高降低11.1%。这一结果重要，因为稀疏模型服务越来越受通信行为限制，而不只是受加速器FLOPs限制。

Circuits & Communications 电路与通信

Constellation Shaping for Complexity-Constrained Direct-Detection Optical Links 面向复杂度受限直接探测光链路的星座成形

R. Fischer, S. Chimmalgi, A. Rode, et al.

arXiv:2607.06300 · 2026-07-07T14:05:07Z

Fischer et al. evaluate joint probabilistic and geometric constellation shaping with reinforcement learning for direct-detection optical signals. The demonstration is a simulated 56 GBd, 2.2 km C-band direct-detection system with complexity-constrained joint equalization and demodulation. The paper is relevant to short-reach optical links because receiver complexity is a hard deployment constraint, especially as bandwidth rises. Caveat: the reported evidence is simulation, so hardware margin and implementation cost still need validation. Fischer等人评估了用强化学习进行联合概率与几何星座成形的方法，目标是直接探测光信号。论文展示的是一个仿真的56 GBd、2.2 km C波段直接探测系统，并约束联合均衡与解调复杂度。它与短距光链路相关，因为随着带宽提升，接收机复杂度会成为真实部署约束。需要注意的是，当前证据来自仿真，硬件余量和实现成本仍需验证。

Bending Beams for THz Wireless Networks With a Prototype 带原型验证的THz无线网络弯曲波束

A. Liu, X. T. Hu, W. Mei, et al.

arXiv:2607.06294 · 2026-07-07T14:01:30Z

Liu et al. survey and prototype bending beams for THz wireless communication and sensing. These near-field beams are non-diffracting and self-healing, which could help THz links route around blockage. The paper compares beamforming design and hardware implementation methods, then presents a THz hardware demo that validates advantages over conventional beam focusing. The remaining challenges are practical: channel models, beam training, and codebook design are still early. Liu等人系统梳理并原型验证了用于THz无线通信与感知的弯曲波束。这类近场波束具备非衍射和自修复特性，可能帮助THz链路绕开遮挡。论文比较了波束成形设计和硬件实现方法，并给出THz硬件演示，验证其相对传统聚焦波束的优势。剩余挑战仍偏工程化：信道模型、波束训练和码本设计都处于早期阶段。

MoE Reinforcement Learning for Energy-Efficient 5G PDCCH Monitoring 用于5G PDCCH节能监听的MoE强化学习

Y. Xiu, N. Wei, Z. Song, et al.

arXiv:2607.06084 · 2026-07-07T09:52:53Z

Xiu et al. formulate 5G IoT-device PDCCH monitoring as a reliability-constrained energy minimization problem. The receiver decides whether to monitor the downlink control channel or sleep before it knows the actual scheduling outcome, while keeping missed-grant risk below a threshold. The MoE deep reinforcement learning setup matters because idle control-channel listening is a recurring energy drain in low-power radios. Caveat: the candidate abstract frames the method and objective clearly, but real-device silicon and RF-chain overhead are not established here. Xiu等人将5G IoT设备的PDCCH监听建模为带可靠性约束的能耗最小化问题。接收机需要在不知道实际调度结果前，决定监听下行控制信道还是进入低功耗状态，同时把漏检授权的风险控制在阈值以下。MoE深度强化学习方案的意义在于，空闲控制信道监听是低功耗无线设备的长期能耗负担。需要注意的是，候选摘要清楚描述了方法和目标，但没有证明真实设备硅片与RF链路开销。

AI Research for Efficient Inference 面向高效推理的AI研究

ActionCache Accelerates Vision-Language-Action Models Without Training ActionCache无需训练加速视觉语言动作模型

R. Oi, H. Otsuka, K. Matsushima, et al.

arXiv:2607.06370 · 2026-07-07T15:10:15Z

ActionCache reuses intermediate actions to warm-start flow-matching VLA models, reducing the iterative denoising cost in robotic manipulation. The method stores compact multimodal keys and retrieves similar past contexts across episodes or tasks. The paper reports up to 11.75x acceleration for pi_0.5 and 34.43x for GR00T-N1.6 while maintaining high task success in simulation and real-world settings. The hardware relevance is direct: lower denoising latency relaxes accelerator and edge-compute pressure for real-time robots. ActionCache通过复用中间动作来热启动flow-matching VLA模型，从而降低机器人操作中的迭代去噪成本。该方法存储紧凑的多模态key，并在不同episode或任务之间检索相似上下文。论文报告在保持较高任务成功率的同时，pi_0.5最高加速11.75倍，GR00T-N1.6最高加速34.43倍，覆盖仿真和真实环境。其硬件意义很直接：降低去噪时延可以缓解实时机器人对加速器和边缘计算的压力。

Truncated Jump Sampling Cuts Diffusion NFEs Without Retraining Truncated Jump Sampling无需重训练降低扩散模型NFE

X. Peng, A. Gao

arXiv:2607.06114 · 2026-07-07T10:24:38Z

Peng and Gao show that standard affine probability paths expose enough x0 information to early-exit an ODE sampler and decode the clean sample. Their Truncated Jump Sampling method needs no retraining, distillation, or architecture change. Across SDXL, SD3.5M, Z-Image-Turbo, and three class-conditional benchmarks, it reduces neural function evaluations by 20% to 70% with near-matched quality. The result is useful for inference systems because NFE count maps directly to serving cost and accelerator occupancy. Peng和Gao指出，标准仿射概率路径在中间状态中已经暴露了足够的x0信息，可让ODE采样器提前退出并解码干净样本。他们提出的Truncated Jump Sampling不需要重训练、蒸馏或改变架构。在SDXL、SD3.5M、Z-Image-Turbo和三个类别条件基准上，该方法将神经函数评估次数降低20%到70%，同时质量接近原始结果。这个结果对推理系统有用，因为NFE数量会直接映射到服务成本和加速器占用。

Quantum 量子

Entanglement as a Complexity Axis in Quantum Policies 作为量子策略复杂度轴的纠缠

J. Xu, D. Zeng, J. Paisley, et al.

arXiv:2607.06230 · 2026-07-07T12:57:46Z

Xu et al. give a PAC-Bayesian account of generalization in parameterized quantum circuits used as policies and value functions. The key claim is that entanglement inflates the effective Fisher dimension, making connectivity a separate complexity axis from parameter count. In controlled experiments, more-entangled circuits show larger train-test gaps than non-entangled circuits with the same number of trainable rotations. This is a useful caution for quantum ML hardware and software co-design: more entanglement is not automatically better when sample efficiency matters. Xu等人从PAC-Bayesian角度分析了作为策略和值函数的参数化量子电路泛化能力。其核心观点是，纠缠会放大Fisher几何的有效维度，使连接性成为独立于参数数量的复杂度轴。在受控实验中，纠缠更多的电路相较于可训练旋转数量相同的非纠缠电路，呈现更大的训练到测试间隙。这对量子ML软硬件协同设计是一个有用提醒：当样本效率重要时，更多纠缠并不自动等于更好。