Circuits & Systems
电路与系统
Roberto Ammendola, Andrea Biagioni, Carlotta Chiarini, et al.
arXiv:2607.02429 · 2026-07-02T16:59:19Z
APEIRON describes a distributed heterogeneous processing framework that combines a multi-FPGA hardware architecture with a software stack for trigger and data acquisition. The interesting part is the full hierarchy: device drivers, communication infrastructure, and an HLS-based dataflow programming model are presented as one system. It is targeted at high-energy physics rather than AI, but the same integration problem shows up in low-latency accelerator clusters.
APEIRON提出一个分布式异构处理框架，把多FPGA硬件架构与触发和数据采集软件栈结合起来。值得关注的是它覆盖了完整层级：底层驱动、通信基础设施，以及基于HLS的数据流编程模型。论文面向高能物理而非AI，但低时延加速器集群也会遇到同类系统集成问题。
Sergey Egorov, Gregory Chockler, Brijesh Dongol, et al.
Royal Holloway, University of London / University of Surrey
arXiv:2607.02401 · 2026-07-02T16:34:39Z
FlintKV targets byte-addressable non-volatile memory with a skiplist storage engine that supports snapshots, consistent iterators, atomic batches, and durable linearizability. The paper reports up to a 75% end-to-end throughput improvement by co-designing concurrency control and persistence mechanisms. It is a systems paper, but it matters for memory hardware because NVM needs software semantics that make its latency advantage usable.
FlintKV面向字节寻址非易失性内存，提出支持快照、一致迭代器、原子批写和持久线性一致性的跳表存储引擎。论文称，通过协同设计并发控制和持久化机制，端到端吞吐最高提升75%。这是一篇系统论文，但对存储硬件有意义，因为NVM需要相应的软件语义才能真正释放低时延优势。
Raphael S. Steiner, Christos K. Matzoros, Pal Andras Papp, et al.
arXiv:2607.02324 · 2026-07-02T15:31:37Z
This paper introduces stale synchronous parallel execution for sparse triangular solves and a DAG scheduler called ElasticDivide. By overlapping synchronization with compute, the scheduler reports geometric-mean speedups of 7-30% over GrowLocal on a 48-core ARM machine and 19-60% over SpMP on a 48-core x86 machine. Sparse triangular solve is a stubborn HPC kernel, so these gains are relevant to CPU and accelerator runtime designers.
这篇论文把陈旧同步并行执行引入稀疏三角求解，并提出名为ElasticDivide的DAG调度器。通过重叠同步与计算，论文在48核ARM机器上相对GrowLocal实现7-30%的几何平均加速，在48核x86机器上相对SpMP实现19-60%的加速。稀疏三角求解是HPC中的顽固内核，因此这些结果对CPU和加速器运行时设计都有参考价值。
AI Research for Efficient Inference
面向高效推理的AI研究
Wentao Zhang, Liliana Hotsko, Woojeong Kim, et al.
arXiv:2607.02512 · 2026-07-02T17:59:50Z
Program-as-Weights compiles a natural-language function specification into parameter-efficient adapters for a frozen lightweight interpreter. The headline result is practical: a 0.6B Qwen3 interpreter running PAW programs matches direct prompting of Qwen3-32B while using about one fiftieth of the inference memory and running at 30 tokens/s on a MacBook M3. This is preprint evidence, but it points to a serving pattern where a large model builds reusable small artifacts instead of answering every call itself.
Program-as-Weights把自然语言函数规格编译成适配冻结轻量解释器的参数高效adapter。关键结果很实际：运行PAW程序的0.6B Qwen3解释器可匹配直接提示Qwen3-32B的表现，同时推理内存约为后者的五十分之一，并能在MacBook M3上达到30 tokens/s。这仍是预印本结果，但指向一种服务模式：大模型生成可复用的小型工件，而不是亲自处理每次调用。
Yanjun Zhao, Ruizhong Qiu, Tianxin Wei, et al.
arXiv:2607.02509 · 2026-07-02T17:59:26Z
ReContext is a training-free inference method that uses model-internal relevance signals to build a query-conditioned evidence pool and replay it before generation. On eight long-context datasets at 128K context length, the paper reports consistent gains in evidence use without external memory or context pruning. It does not remove the KV-cache cost of long prompts, but it attacks the separate problem of models failing to use evidence already paid for in memory.
ReContext是一种无需训练的推理方法，利用模型内部相关性信号构建与查询相关的证据池，并在生成前进行重放。论文在8个128K长上下文数据集上报告了稳定的证据利用提升，且不依赖外部记忆或上下文裁剪。它并没有消除长提示带来的KV-cache成本，但针对的是另一个实际问题：模型已经为上下文付出内存成本，却未能有效使用其中证据。
Dezheng Wang, Tong Chen, Wei Yuan, et al.
arXiv:2607.02344 · 2026-07-02T15:49:14Z
Self-Gating Attention targets the quadratic time and memory cost of transformer attention in time-series forecasting. The method replaces part of the attention-score computation with a shared learnable matrix plus an input-dependent residual, exploiting repeated temporal patterns. The caveat is domain scope: the redundancy assumption is natural for time series, but it may not transfer directly to general LLM workloads.
Self-Gating Attention瞄准时间序列预测中Transformer注意力的二次时间和内存开销。它用共享可学习矩阵加输入相关残差来替代部分注意力分数计算，利用许多时间序列中重复的时间模式。需要注意的是适用范围：这种冗余假设很适合时间序列，但未必能直接迁移到通用LLM负载。
Ziyun Qiao, Yue Min, Ruining Chen, et al.
arXiv:2607.02266 · 2026-07-02T14:51:42Z
HERMES proposes a hierarchical labeling substrate for pre-training corpora, using a learned semantic transform and three-stage residual vector quantization to label documents once across multiple granularities. In a 1B-parameter, 25B-token pre-training run, one prefix length exposes a data-mixing rule that improves a 16-task capability macro-average by +0.0253, while a finer level loses the gain. The result is relevant because data efficiency changes training compute demand, but the paper also shows how sensitive these rules are to granularity.
HERMES为预训练语料提出层级标签基底，用学习到的语义变换和三阶段残差向量量化，对文档进行一次标注并支持多种粒度。在1B参数、25B token的预训练实验中，某一前缀长度下的数据混合规则让16项能力宏平均提升+0.0253，但更细粒度下该收益消失。结果值得关注，因为数据效率会改变训练算力需求，同时也说明这类规则对粒度非常敏感。
Quantum & Physics Computing
量子与物理计算
Quoc Bao Phan, Tuy Tan Nguyen
arXiv:2607.02426 · 2026-07-02T16:54:35Z
QFedAgent adds a variational quantum circuit fusion module to personalized federated learning for multi-agent activity recognition. The paper reports 72 quantum rotation parameters versus 33K parameters for a classical MLP fusion module, roughly a 10x total parameter reduction, with 97.7% mean test accuracy on OPPORTUNITY under subject-based non-IID partitions. Treat it as an algorithmic quantum-classical result, not evidence of near-term quantum hardware deployment.
QFedAgent在面向多agent活动识别的个性化联邦学习中加入变分量子电路融合模块。论文报告量子融合模块只需72个量子旋转参数，而经典MLP融合模块为33K参数，总参数量约降低10倍，并在OPPORTUNITY数据集的按受试者非IID划分下达到97.7%平均测试准确率。应把它视为量子-经典算法结果，而不是近期量子硬件部署证据。
Juan Agustin Duque, Sergio Garcia Heredia, Vinicius Hernandes, et al.
arXiv:2607.02292 · 2026-07-02T15:09:50Z
This paper reframes variational energy minimization for neural quantum states as an advantage policy-gradient problem over the Born distribution. The proposed Proximal Wavefunction Optimization clips probability-ratio changes in the amplitude channel and phase increments in the phase channel, aiming to avoid costly matrix inversion while keeping trust-region behavior. It is relevant to quantum simulation workloads, with the usual caveat that algorithmic scaling still needs validation beyond benchmark spin systems.
这篇论文把神经量子态的变分能量最小化重新表述为Born分布上的优势策略梯度问题。作者提出的Proximal Wavefunction Optimization会裁剪振幅通道中的概率比变化和相位通道中的相位增量，目标是在避免昂贵矩阵求逆的同时保留信赖域行为。它对量子模拟负载有意义，但仍需在基准自旋系统之外验证算法扩展性。