Inference Efficiency Meets Physical-Layer Constraints 推理效率遇上物理层约束

This week's stronger papers connect AI system efficiency with the hardware limits underneath it: idle GPUs, carbon-aware data centers, RF interference, near-field arrays, and modulator physics. The common thread is that useful performance increasingly comes from co-design across models, communication channels, devices, and operations. 本周更值得读的论文把AI系统效率与底层硬件约束连接起来：闲置GPU、低碳数据中心、射频干扰、近场阵列以及调制器物理。共同主线是，有效性能越来越依赖模型、通信信道、器件和运营之间的协同设计。

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

Devices & Process 器件与工艺

Graphene Electric Double-Layer Transistors for Label-Free Albumin Detection 用于无标记白蛋白检测的石墨烯电双层晶体管

A. Liaquat, G. Baridi, F. Rapuzzi, et al.

arXiv:2607.03491 · 2026-07-03T16:58:53Z

The paper demonstrates a graphene electrolyte-gated FET that detects human serum albumin in real time under non-Faradaic operation. Albumin adsorption shifts the Dirac voltage in a concentration-dependent way, and Brownian Dynamics simulations connect the electrical response to adsorption orientation and interfacial charge. For semiconductor readers, the useful piece is the device-level transduction mechanism rather than the biomedical target alone. 这篇论文展示了一种石墨烯电解质栅FET，可在非Faradaic工作条件下实时检测人血清白蛋白。白蛋白吸附会以浓度相关方式移动Dirac电压，Brownian Dynamics模拟进一步把电学响应与吸附取向和界面电荷联系起来。对半导体读者来说，重点不只是生物医学应用，而是器件级换能机制。

Mid-IR Electro-Optic Comb Generation with an Ultrafast Modulator 基于超快调制器的中红外电光频梳生成

G.-L. Ngo, L. Lucia, S. Pirotta, et al.

arXiv:2607.03317 · 2026-07-03T13:37:10Z

The authors generate mid-infrared single- and dual-combs around 9 micrometers using room-temperature free-space electro-optic intensity modulators driven by short electrical pulse trains. They demonstrate spectroscopy on a germanium etalon and ammonia cell, with tunable repetition rates down to the megahertz range. The work is relevant to photonic sensing hardware because it pushes electro-optic comb techniques beyond the easier near-IR bands. 作者使用室温自由空间电光强度调制器，并以短电脉冲列驱动，在约9微米中红外波段生成单频梳和双频梳。他们在锗标准具和氨气池上演示了光谱测量，重复频率可调至MHz量级。该工作对光子传感硬件有意义，因为它把电光频梳方法推进到更难实现的中红外波段，而不只是近红外。

Circuits, RF & Sensing Systems 电路、射频与感知系统

Deep-Unfolded Wideband ISAC Beamforming for Dynamic Metasurface Antennas 用于动态超表面天线的深度展开宽带ISAC波束成形

A. S. Gharagezlou, P. Mobaraki, M. Monemi, et al.

arXiv:2607.03389 · 2026-07-03T14:46:16Z

This paper tackles wideband integrated sensing and communications with dynamic metasurface antennas under a frequency-selective Lorentzian element model. Instead of assuming frequency-flat DMA behavior, it optimizes digital beamforming, resonance frequencies, and damping factors, then unfolds projected-gradient updates into a trainable architecture. The result is useful for RF hardware co-design because it keeps the physical element response in the learning loop. 这篇论文研究采用动态超表面天线的宽带通感一体系统，并使用频率选择性的Lorentzian单元模型。它没有假设DMA单元是频率平坦的，而是联合优化数字波束成形、谐振频率和阻尼因子，再把投影梯度更新展开为可训练架构。其价值在于把真实射频单元响应保留在学习优化闭环中。

Joint Dictionary Learning and Channel Estimation for XL-MIMO 面向XL-MIMO的联合字典学习与信道估计

A. Arjas, I. Atzeni

arXiv:2607.03448 · 2026-07-03T16:06:18Z

The authors propose DL-ISTA for near-field channel estimation in extra-large MIMO arrays, where angle and distance both shape the array response. The method jointly learns sparse channel coefficients and continuous angle-distance parameters, reducing the grid mismatch that affects fixed dictionary methods. This is relevant to future base-station and sensing hardware as arrays grow large enough that far-field assumptions break down. 作者提出DL-ISTA，用于超大规模MIMO阵列中的近场信道估计；在这种场景下，角度和距离都会影响阵列响应。该方法联合学习稀疏信道系数和连续的角度-距离参数，从而减轻固定字典方法中的网格失配问题。随着基站和感知阵列规模增大、远场假设逐渐失效，这类方法对未来硬件系统更有价值。

Ambient IoT Backscatter as Passive Anchors for NLOS Cellular Positioning 将环境IoT反向散射器件用作NLOS蜂窝定位的被动锚点

H. Yiğitler, M. F. Keskin, O. Kaltiokallio, R. Jäntti

arXiv:2607.03459 · 2026-07-03T16:15:50Z

The paper studies whether ambient IoT backscatter devices at known locations can serve as passive anchors for non-line-of-sight cellular positioning. It derives Fisher-information limits under calibrated, partially calibrated, and uncalibrated device assumptions, showing how unknown phases strip away carrier-phase information. The result is a sober boundary condition for low-cost localization schemes that rely on weak, unpowered RF devices. 这篇论文研究位于已知位置的环境IoT反向散射器件，能否作为非视距蜂窝定位的被动锚点。作者在已校准、部分校准和未校准假设下推导Fisher信息极限，并说明未知相位如何剥离载波相位信息。该结果为依赖低成本、无源射频器件的定位方案给出了较清醒的物理边界。

AI Accelerators & Compute Systems AI加速器与计算系统

SPORK: Self-Speculative Forking to Accelerate Agentic LLM Inference SPORK：用自预测分叉加速Agentic LLM推理

H. Bai, W. Lv, H. Zheng, Y. Lu, J. Shu

arXiv:2607.03333 · 2026-07-03T13:51:32Z

SPORK targets a concrete inference-system inefficiency: agentic LLMs often idle the GPU while waiting for tool results, which the paper says consumes 16-37% of wall time in its workloads. The method forks a lightweight probe early in generation to predict the coming tool call, then overlaps tool execution with the remaining decode. The reported Qwen3-32B tool-name prediction accuracy of 74.6-99.6% suggests useful latency hiding without an auxiliary model. SPORK瞄准了一个具体的推理系统低效点：Agentic LLM在等待工具结果时经常让GPU空转，论文称在其工作负载中这占用16-37%的墙钟时间。该方法在生成早期分叉出轻量探针预测即将发生的工具调用，并把工具执行与后续解码重叠起来。Qwen3-32B上74.6-99.6%的工具名预测准确率表明，不依赖额外模型也可能实现有效的延迟隐藏。

Carbon-Aware Multi-Agent Scheduling for AI Data Centers 面向AI数据中心的低碳多智能体调度

H. Lee, P. Prabawa, D.-H. Choi, J. Kim

arXiv:2607.03324 · 2026-07-03T13:41:00Z

The paper proposes a hierarchical multi-agent RL framework that shifts and allocates AI training and inference jobs across data centers using nodal carbon-intensity signals from the power distribution system. It models both a workload-manager agent and local data-center agents, tying GPU scheduling to grid emissions rather than only electricity price or utilization. The work matters because accelerator deployment is increasingly constrained by power availability, not just chip supply. 这篇论文提出一种分层多智能体强化学习框架，利用配电系统中的节点碳强度信号，在多个AI数据中心之间移动和分配训练与推理任务。它同时建模工作负载管理智能体和本地数据中心智能体，把GPU调度与电网排放联系起来，而不只是电价或利用率。该工作重要，因为加速器部署正越来越受电力可获得性约束，而不只是芯片供给。

AI Systems & Model Behavior AI系统与模型行为

Decoding Hidden Computation Across Filler Tokens 解码填充Token中的隐藏计算

K. Brauer, C. M. Verdun, S. Marks

arXiv:2607.03502 · 2026-07-03T17:18:34Z

This paper shows that frontier open-weight LLMs can perform structured computation over content-free filler tokens such as dots or counting sequences. The authors use attention analysis, logit-lens readouts, and KV-cache transplants to show that hidden intermediate computation remains partly legible even when the visible chain of thought carries no information. For inference-system designers, it is a reminder that token streams and KV-cache state can encode more computation than surface text suggests. 这篇论文表明，前沿开源权重LLM可以在点号或计数序列等无内容填充token上执行结构化计算。作者通过注意力分析、logit-lens读出和KV-cache移植显示，即使可见思维链不携带信息，隐藏的中间计算仍部分可读。对推理系统设计者而言，这提醒我们token流和KV-cache状态承载的计算可能远多于表面文本所显示的内容。