Devices & Process
器件与工艺
L. Tan, X. Hu, M. Zhou, et al.
Advanced Science · 2026-06-12
The paper reports a photoelectric-coupled ferroelectric heterojunction that uses polarization memory to modulate the sensing interface for NO2 detection. The architecture links nonvolatile ferroelectric state to gas-response control, potentially reducing the need for continuous bias or external calibration. The candidate record contains no abstract or device metrics, so sensitivity, endurance, selectivity, and manufacturability require review in the full paper.
论文报道了一种光电耦合铁电异质结，利用极化记忆调制NO2传感界面。该结构把非易失铁电状态与气体响应控制相结合，有望减少持续偏置或外部校准需求。候选记录未提供摘要和器件指标，因此灵敏度、耐久性、选择性与可制造性仍需查阅全文。
M. Camacho
arXiv:2606.13609 · 2026-06-11T17:24:05Z
Camacho derives an exact selection rule for a time-modulated trilayer waveguide: odd frequency sidebands switch transverse parity, while even sidebands preserve it. Floquet eigenstate analysis and finite-section time-domain simulations suppress symmetry-forbidden channels to numerically negligible levels. The result offers a design rule for controlled photonic frequency conversion, but it remains simulation-based and has not yet been demonstrated in fabricated hardware.
Camacho为时间调制三层波导推导出精确选择规则：奇数频率边带会切换横向模态奇偶性，偶数边带则保持奇偶性。Floquet本征态分析和有限区段时域模拟将对称性禁戒通道压低至数值上可忽略的水平。该结果为可控光子频率转换提供了设计规则，但目前仍基于模拟，尚未在制造出的硬件上验证。
AI Research for Hardware
面向硬件的AI研究
G. Yu, W. Liu, Y. Hu, et al.
arXiv:2606.13657 · 2026-06-11T17:54:09Z
Across language and vision-language models, on-policy distillation produces small, coordinate-sparse updates distributed across layers and concentrated in feed-forward networks. Training only the identified subnetwork recovers nearly the same performance as full distillation, suggesting a route to lower optimizer-state and communication costs. The updates remain numerically full-rank and AdamW still beats a sparsity-inducing SGD variant, so the result does not reduce cleanly to low-rank adaptation.
在语言模型和视觉语言模型上，在策略蒸馏产生幅度较小、坐标稀疏的参数更新，这些更新分布于多层并主要集中在前馈网络。仅训练识别出的子网络即可恢复接近完整蒸馏的性能，为降低优化器状态和通信成本提供了路径。不过更新在数值上仍为满秩，且AdamW优于诱导稀疏的SGD变体，因此该结果不能简单归结为低秩适配。
J. Gan, X. Qiu, G. Chen, et al.
arXiv:2606.13624 · 2026-06-11T17:39:26Z
The paper compresses time-series tokens according to frequency-domain contribution and progressively drops prompt tokens as their influence fades with depth. Across forecasting, classification, imputation, and anomaly detection, the method reports up to 7.68x inference acceleration and performance gains in 78% of evaluated settings. The systems value is a variable token budget that reduces attention work, although the evidence is limited to time-series tasks rather than general-purpose LLM serving.
论文根据频域贡献压缩时间序列token，并随着提示词影响在深层网络中减弱而逐层减少prompt token。该方法在预测、分类、插补和异常检测任务上最高实现7.68x推理加速，并在78%的评测设置中提升性能。其系统价值在于以可变token预算减少注意力计算，但证据目前仅来自时间序列任务，尚不能直接推广到通用LLM服务。
S. Tang, Y. Zhu, M. Tao, et al.
arXiv:2606.13565 · 2026-06-11T16:50:58Z
A2D2 jointly fine-tunes token-insertion and unmasking policies for any-length discrete diffusion models, then uses a quality-based adaptive inference schedule. Its derivation targets the reward-tilted sequence distribution without requiring target samples, and experiments report better reward optimization and generation flexibility than fixed-length baselines. Hardware impact depends on whether adaptive schedules reduce real decoding work after accounting for control overhead; no accelerator measurements are reported.
A2D2面向任意长度离散扩散模型，联合微调token插入和去掩码策略，并采用基于质量的自适应推理调度。其推导无需目标样本即可逼近奖励倾斜的序列分布，实验显示相较固定长度基线可改善奖励优化和生成灵活性。其硬件价值取决于扣除调度控制开销后能否真正减少解码计算，论文未报告加速器实测结果。
S. Zhang, R. Liao, V. Tresp, et al.
arXiv:2606.13558 · 2026-06-11T16:41:25Z
BitResEdit performs training-free image editing by steering per-bit Bernoulli predictions inside a KL trust region and reinjecting localized residual codes across scales. Operating on the generator's native bitwise and residual structures preserves masked-out latent features exactly by code arithmetic, rather than relying only on token-level guidance. The approach is specific to bitwise-residual visual autoregressive models such as Infinity, so its utility depends on adoption of that representation.
BitResEdit通过在KL信赖域内引导逐比特Bernoulli预测，并跨尺度重新注入局部残差编码，实现免训练图像编辑。它直接利用生成器原生的逐比特和残差结构，通过编码运算精确保留掩码外的潜在特征，而不是只依赖token级引导。该方法针对Infinity等逐比特残差视觉自回归模型，其价值取决于这类表示方式的采用程度。
J. Hur, Y. Kwon, M.-G. Jo, et al.
arXiv:2606.13443 · 2026-06-11T15:05:02Z
AMGFNO learns how much historical state a neural operator should retain instead of applying a fixed memory weight across observation conditions. On Kuramoto-Sivashinsky and Burgers' equation benchmarks, it reports 55-79% lower normalized RMSE at low resolution, while the learned gate falls from about 0.7 toward zero as resolution rises. The adaptive state budget could reduce unnecessary memory traffic, but results are confined to simulated PDE workloads and do not include hardware measurements.
AMGFNO让神经算子学习应保留多少历史状态，而不是在不同观测条件下使用固定记忆权重。在Kuramoto-Sivashinsky和Burgers方程基准上，该方法在低分辨率时将归一化RMSE降低55-79%，并让学习到的门值随分辨率提高从约0.7降至接近0。自适应状态预算可能减少不必要的内存流量，但结果仅限于模拟PDE负载，且没有硬件实测。