semi·news
Headlines要闻 / Research研究 / /
Research digest · Friday, November 20, 2026 研究摘要 · 2026年11月20日 星期五

Efficiency Depends on Scheduling, Precision and Memory 效率取决于调度、精度与存储

Eight preprints examine production inference dispatch, edge learning, hybrid attention, quantization verification and local engineering agents. Reported gains range from fewer H100s in a production A/B test to tighter error bounds, with benchmark scope and deployment assumptions kept explicit. 本期8篇预印本涵盖生产推理调度、边缘学习、混合注意力、量化验证及本地工程智能体。研究结果从生产A/B测试中减少H100用量,到收紧量化误差边界;摘要保留各自的评测范围与部署条件。

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

Edge Accelerators & Learning Systems 边缘加速器与学习系统

Hypnos Splits Edge Continual Learning Between a RISC-V Device and a Smart Charger Hypnos将边缘持续学习分配给RISC-V设备与智能充电器

Guilherme Guedes, Marco Apolinario Lainez, Charlotte Frenkel

arXiv:2611.24637 · 2026-11-18T15:46:58Z

Hypnos separates continual learning into a wake phase on a small RISC-V MCU-based system and a sleep phase on a smart charger. The device performs inference and lightweight gradient-free prototype updates, while the charger uses rehearsal-based gradient updates when power and memory budgets are larger. This partition targets adaptation without placing full backpropagation on the edge device; the supplied abstract mentions a ZCU104 FPGA deployment but ends before quantitative results. Hypnos将持续学习分为两个阶段:由基于RISC-V MCU的小型系统承担清醒阶段,由智能充电器承担睡眠阶段。设备负责推理和轻量级无梯度原型更新,充电器则在电力与内存预算更充足时,通过样本回放进行梯度更新。这种划分旨在实现在线适应,同时避免在边缘设备上执行完整反向传播;所提供的摘要提及ZCU104 FPGA部署,但在量化结果之前已被截断。

Inference Systems, Models & Quantization 推理系统、模型与量化

Cap-and-Track Adapts Inference Dispatch to Changing Backend Capacity Cap-and-Track让推理请求分发适应后端容量变化

Benjamin Novotny, Doug Szeto, Chang Wang, et al.

arXiv:2611.24891 · 2026-11-18T18:34:10Z

Cap-and-Track uses fast admission refusals to track available inference capacity, holding excess and retried work in durable priority queues outside execution. In a concurrent production A/B test at Meta, the traffic shaper served a classifier on 43% fewer H100s than its queued, fixed-rate predecessor, while lower-priority work waited longer during backfill peaks. Real-engine replay at steady capacity put goodput within 1% of a fixed rate tuned above measured capacity. The production result concerns asynchronous classifier serving rather than a general reduction in interactive LLM hardware needs. Cap-and-Track利用快速准入拒绝信号追踪可用推理容量,将超额任务及待重试任务保存在执行路径之外的持久化优先队列中。在Meta的同期生产A/B测试中,该流量整形器运行一个分类器时,比采用队列与固定分发速率的前代方案少用43%的H100,但低优先级任务在回填高峰时等待更久。在稳定容量下的真实引擎回放中,其有效吞吐与调至高于实测容量的固定速率方案相差不到1%。生产结果针对异步分类服务,不能直接推广为交互式大模型所需硬件的普遍降幅。

FlashMorph2 Selects Hybrid Attention Layers Under the Recovery Objective FlashMorph2按恢复训练目标选择混合注意力层

Disen Lan, Xuanda Wang, Jianbin Zheng, et al.

arXiv:2611.24809 · 2026-11-18T17:45:51Z

FlashMorph2 converts pretrained Transformers into hybrid full- and linear-attention models by evaluating discrete architectures on their own hidden states and jointly updating layer-selection scores and weights. Across tested Qwen3 scales and linear-attention variants, average scores improve over FlashMorph by 2.42 points on RULER, 1.08 on LongBench and 0.76 on recall-intensive tasks. Layer selection takes 0.72–3.57 GPU-hours for 0.6B–8B models. These results address selection quality and cost; the supplied excerpt does not give an end-to-end serving speedup. FlashMorph2通过使用离散架构自身产生的隐藏状态进行评估,并联合更新层选择分数与权重,将预训练Transformer转换为结合完整注意力和线性注意力的模型。在所测试的Qwen3规模与线性注意力变体上,相比FlashMorph,RULER、LongBench及高召回需求任务的平均分分别提高2.42、1.08和0.76分。对于6亿至80亿参数模型,层选择耗时为0.72至3.57 GPU小时。这些结果针对选择质量与成本,所提供的摘要尚未给出端到端推理服务加速比。

Zonotopic Analysis Tightens Quantized-Network Error Bounds Zonotope分析收紧量化神经网络误差边界

Robert Debbas, Éric Goubault, Sylvie Putot

arXiv:2611.24648 · 2026-11-18T15:52:07Z

This work verifies quantized ReLU networks with a zonotope-based abstract interpretation framework that handles rounding and clamping jointly. On ACAS Xu at the two smallest tested perturbation radii, constrained zonotopes reduce average error-bound widths by about 70–85% and runtimes by 65–80% relative to QEBVerif's symbolic DRA propagation. Plain zonotopes produce competitive sound bounds nearly 300 times faster on average. Gains vary with bitwidth, and symbolic DRA yields tighter bounds at larger radii, making the method's advantage regime-dependent. 研究提出基于zonotope(带状多面体)的抽象解释框架,联合处理舍入与截断,以验证量化ReLU网络。在ACAS Xu基准的两个最小测试扰动半径下,相比QEBVerif的符号DRA传播,约束zonotope将平均误差边界宽度缩小约70%至85%,运行时间减少65%至80%。普通zonotope方法也能给出有竞争力的可靠边界,平均速度接近快300倍。收益随位宽而变化,较大扰动半径下符号DRA的边界更紧,因此优势取决于具体参数区间。

Clover Combines MoE Diffusion Generation with Recurrent Context Memory Clover结合MoE扩散生成与循环上下文记忆

DiffuSpace Technical Team

arXiv:2611.24636 · 2026-11-18T15:46:14Z

Clover-35B-A3B-Preview is a diffusion language model with 35B total parameters and 3B activated parameters, combining recurrent memory for prior context with bidirectional attention inside each generated block. The authors report scores of 73.6 on SWE-bench Verified and 51.7 on Terminal-Bench 2.1. On internal agent workloads, generation reaches a reported 680 tokens/s on one Nvidia GB300 at BF16 and batch size one. That throughput is tied to the stated internal workload and does not establish performance across serving conditions. Clover-35B-A3B-Preview是一款总参数350亿、激活参数30亿的扩散语言模型,将历史上下文的循环记忆与当前生成块内的双向注意力结合。作者报告,其SWE-bench Verified和Terminal-Bench 2.1得分分别为73.6和51.7。在内部智能体负载上,单颗英伟达GB300以BF16精度、批量大小为1运行时,报告的生成速度达到每秒680个词元。该吞吐对应指定的内部负载,不能据此判断各种服务条件下的表现。

GateWatch Uses Compact Boolean Rules for LLM Activation Monitoring GateWatch以紧凑布尔规则监测大模型激活

Firas Spanioli, Yaniv David

arXiv:2611.24619 · 2026-11-18T15:37:04Z

GateWatch expresses activation-monitor decisions as Boolean logic over a small set of binarized sparse-autoencoder features. GateWatchServe computes only selected features, streams binary values across layers and evaluates probes through compact lookup tables. Across 24 Gemma-2-9B tasks, the authors report 80.49% mean test accuracy with averages of 3.8 features and 3.5 logic gates per probe. The compact rule size is promising for monitoring cost, but the supplied excerpt does not quantify end-to-end inference overhead. GateWatch将激活监测器的决策表示为少量二值化稀疏自编码器特征上的布尔逻辑。GateWatchServe只计算选中的特征,跨层流式传递二值结果,并通过紧凑查找表评估探针。在Gemma-2-9B的24项任务上,作者报告平均测试准确率为80.49%,每个探针平均仅使用3.8个特征与3.5个逻辑门。紧凑规则有助于降低监测成本,但所提供的摘要尚未量化端到端推理开销。

Engineering Automation & Simulation 工程自动化与仿真

Local AI Agents Measure the Deployment Costs of Power Electronics Simulation 本地AI智能体评估电力电子仿真的部署成本

Yuzhuo Li, Pasan Gunawardena, Yunwei Li

arXiv:2611.24880 · 2026-11-18T18:27:40Z

This paper presents an evaluation framework for locally hosted AI agents that build, run, measure and revise power electronics simulations. It treats power, energy, token use and elapsed time as deployment costs, extending assessment beyond whether an agent completes a design task. The approach is relevant when engineering data and licensed simulators must remain on site. The supplied abstract is truncated before numerical comparisons, so it does not support a specific cost or efficiency claim. 论文提出评估框架,研究由本地部署AI智能体构建、运行、测量和修改电力电子仿真的流程。它将功率、能耗、词元用量与耗时纳入部署成本,评价范围超出了智能体能否完成设计任务。对于工程数据和授权仿真软件必须留在本地的场景,这种方法具有实际意义。所提供的摘要在数值对比之前已被截断,因此无法支持具体的成本或效率结论。

Quantum & Hybrid Learning 量子与混合学习

Soft-MoQE Learns to Combine Fixed Quantum Feature Maps Soft-MoQE学习组合固定量子特征映射

Valeria Repetto, Romina D'Aurizio, Piero Farinelli

arXiv:2611.24798 · 2026-11-18T17:41:45Z

Soft-MoQE replaces trainable quantum circuits with a pool of fixed quantum feature maps combined by an input-dependent classical gating network. On three PCA-reduced MNIST and FashionMNIST tasks, it outperforms the best individual expert, uniform expert averaging and a linear baseline, while reaching task-dependent performance comparable to a trainable quantum neural network. This avoids optimizing quantum parameters and targets the cost and optimization difficulties of variational training. The supplied results do not establish quantum advantage over strong classical models or performance on unreduced datasets. Soft-MoQE用一组固定量子特征映射替代可训练量子电路,再由依赖输入的经典门控网络组合这些映射。在经过PCA降维的3项MNIST与FashionMNIST任务上,它优于最佳单个专家、专家均匀平均和线性基线,并在部分任务中达到与可训练量子神经网络相当的表现。该方法无需优化量子参数,针对变分训练的成本与优化困难。现有结果尚不能证明其相对强经典模型的量子优势,也不能说明在未降维数据集上的表现。

Newsletter 邮件订阅

Daily semiconductor briefing. 每日半导体简报。