Circuits & Architecture
电路与架构
R. Nather
arXiv:2606.13623 · 2026-06-11T17:36:54Z
The work adds an await_delete construct to future-based runtimes so storage can be safely reused after all consumers finish, addressing a practical limitation of immutable future values. A block-wise symmetric-matrix inversion implementation in an extended Taskflow runtime approaches linear scaling on large matrices. Futures still impose substantial overhead on smaller problems, so the benefit depends on crossing an amortization threshold.
该工作为基于Future的运行时加入await_delete机制，使所有消费者完成后可安全复用存储，解决Future值不可变带来的实际限制。作者在扩展版Taskflow中实现分块对称矩阵求逆，大矩阵上接近线性扩展。小规模问题仍承受明显的Future开销，因此收益取决于能否跨过开销摊销阈值。
J. Guichemerre, C. Studer
arXiv:2606.13417 · 2026-06-11T14:46:56Z
The paper develops two affine models for low-resolution SAR-ADC behavior in massive MU-MIMO, one based on Bussgang decomposition and one optimized for signal-to-distortion ratio. The models explicitly include converter non-idealities that ideal quantization analyses usually omit, then support low-complexity digital mitigation. The available abstract does not report measured converter silicon or system-level power savings, so the result remains a modeling and algorithm contribution.
该论文为大规模MU-MIMO中的低分辨率SAR-ADC建立两种仿射模型，一种基于Bussgang分解，另一种针对信号失真比优化。模型显式纳入理想量化分析常忽略的转换器非理想效应，并据此设计低复杂度数字补偿。现有摘要未报告ADC实测硅片或系统级功耗节省，因此主要贡献仍是建模与算法。
AI Research for Hardware
面向硬件的AI研究
E. Lumer, S. Sen, K. Paul, et al.
arXiv:2606.13643 · 2026-06-11T17:47:30Z
Recursive Agent Harnesses let a parent agent generate executable workflows that spawn tool-using subagents in parallel rather than recursively calling a bare model. With GPT-5 held constant, the approach improves the Codex baseline from 71.75% to 81.36% on 199 Oolong-Synthetic tasks spanning contexts up to 4 million tokens. The gain shifts scaling pressure toward orchestration, parallel inference, and structured communication, but evidence is currently limited to a controlled long-context benchmark.
Recursive Agent Harness让父agent生成可执行工作流，并行启动具备工具能力的子agent，而不是递归调用裸模型。在固定使用GPT-5时，该方法在199个Oolong-Synthetic任务上把Codex基线从71.75%提升至81.36%，上下文长度最高达400万token。它把扩展压力转向编排、并行推理与结构化通信，但当前证据仍局限于受控长上下文基准。
M. S. Preetam, M. Bhaskar
arXiv:2606.13589 · 2026-06-11T17:11:51Z
Simplex-Constrained Sparse Bagging jointly prunes and calibrates ensemble models using out-of-bag loss plus a concave quadratic sparsity penalty. The method reports up to 96% ensemble compression with linear inference speedups while preserving or improving accuracy and calibration. Results cover bagged model families rather than modern transformer workloads, so transfer to accelerator-heavy inference still needs demonstration.
Simplex-Constrained Sparse Bagging使用袋外损失与凹二次稀疏惩罚，联合完成集成模型剪枝和概率校准。该方法报告最高96%的集成压缩，并获得线性推理加速，同时保持或改善准确率与校准表现。实验面向Bagging模型族而非现代Transformer负载，因此能否迁移到加速器密集型推理仍需验证。
H. Jung, X. Wang
arXiv:2606.13550 · 2026-06-11T16:30:45Z
UMG-RAG treats retrieval granularity as a query-specific reliability problem, combining dense and sparse retrievers across multiple chunk sizes using entropy-derived confidence. A parent-promotion variant retrieves with fine chunks but returns broader, non-redundant context to the generator. The design is training-free and could reduce wasted context processing, although the candidate abstract does not expose latency, token, or energy savings.
UMG-RAG把检索粒度视为按查询变化的可靠性问题，利用熵生成的置信度融合多种分块尺度上的稠密与稀疏检索器。其parent-promotion变体用细粒度分块定位证据，再向生成器返回更宽且不重复的上下文。该设计无需训练，可能减少无效上下文计算，但候选摘要未给出延迟、token或能耗节省数据。
X. Zhang, L. Shen, M. Chen, et al.
arXiv:2606.13513 · 2026-06-11T16:00:51Z
CloudCons evaluates forecasting models by their downstream value in cloud resource consolidation rather than prediction error alone. Its datasets span Huawei Cloud, Microsoft Azure, and Google Borg workloads with periodic, bursty, and noisy demand patterns. This is useful for infrastructure efficiency because a statistically better forecast may not produce a better placement decision; the available abstract, however, truncates before reporting the full comparative results.
CloudCons不只比较预测误差，而是按模型在云资源整合决策中的下游价值进行评测。数据集覆盖Huawei Cloud、Microsoft Azure和Google Borg负载，包含周期型、突发型与高频噪声需求。该基准有助于基础设施效率研究，因为统计上更准的预测未必带来更好的放置决策；但现有候选摘要在完整对比结果前被截断。
A. Soen, H. Husain, V. De Bortoli, et al.
arXiv:2606.13426 · 2026-06-11T14:54:13Z
The paper adapts block verification from speculative LLM decoding to continuous diffusion sampling while preserving the target distribution through acceptance and rejection. Its training-free Free Drafter reports up to a 6.3% speedup over existing speculative methods with negligible additional overhead. The gain is modest, but the method matters because it avoids training a separate draft model and attacks repeated diffusion steps directly.
该论文把推测式LLM解码中的块验证移植到连续扩散采样，并通过接受-拒绝机制保持目标分布。无需训练的Free Drafter相对现有推测方法最高加速6.3%，额外开销可忽略。增益幅度不大，但其价值在于无需训练独立草稿模型，并直接减少重复扩散步骤。