The proposed spiral interferometric laser considerably outperforms the best reports of integrated cavities to date, achieve a record on-chip fractional frequency noise of 5.6 × 10^−14, corresponding to a linewidth of 12 Hz centered at 1348 nm.
MIT researchers have demonatrated a new architecture for an ultranarrow linewidth integrated laser based on stabilization to a sinusoidal fringe of an interferometer having a long 25 meters unbalanced delay line.
arxiv.org/pdf/2602.16461
The results presented represent a practical step towards scalable quantum networks on a metropolitan scale, employing teleportation-based protocols.
The results hsow that, averaged over the three inputs, the fidelities were 90.1 ± 3.3% without co-propagating traffic and 85.9 ± 4.3% with traffic, quantifying polarization drift and classical-channel crosstalk under carrier-grade conditions.
The teleported photon was then transmitted through a 30-km field-deployed fiber loop under real-world environmental perturbations and, optionally, with co-propagating C-band classical data traffic.
Reseachers have demonstrated quantum teleportation over Deutsche Telekom’s metropolitan fiber testbed in Berlin using commercial components deployed at the telecom datacenter.
arxiv.org/pdf/2602.16613
The proposed system achieved a mean secret key rate of 1 kbit/s with high-efficiency SNSPDs and 200 bit/s with room-temperature SPAD detectors, demonstrating the practical viability of long-range urban free-space quantum communication.
Researchers have demonstrated real-time intermodal free-space QKD at telecommunication wavelength over an 18 km ground-to-ground low-altitude link, effectively exploiting adaptive optics in a 40 cm-class telescope to mitigate atmospheric turbulence.
arxiv.org/pdf/2602.16680
The authors demonstrate, through three popular applications (AES encryption, CNNs, LLMs), how it is possible to use and benefit from DARTH-PUM, with speedups of 59.4x, 14.8x, and 40.8x compared to an analog+CPU baseline.
DARTH-PUM provides a complete hybrid analog–digital PUM system with ISA, which can be used to accelerate a variety of applications from machine learning to cryptography.
In this paper is presented DARTH-PUM, a general-purpose hybrid processing-using-memory (PUM) architecture which combines both analog and digital PUM onto the same chip.
arxiv.org/pdf/2602.16075
The results of the evaluation on seven X-CT datasets show a 12.34x improvement in the relative compression ratio compared to standard compression.
ROIX-Comp applies intensity normalization and adaptive thresholding to improve object boundary detection and structure preservation in X-CT data, supporting both lossless and lossy compression methods, depending on the application requirements.
In this paper is presented ROIX-Comp, a region-of-interest driven extraction framework that intelligently compresses X-ray computed tomography data by identifying and retaining only essential features.
arxiv.org/pdf/2602.15917
The evaluation results of the impact of this API and implementation on a complex halo exchange benchmark running on up to 8,192 GPUs of the Frontier supercomputer show that the API can increase the strong scaling speedup of such communication patterns by up to 28%.
"(...) The API, which leverages persistent communication operations, preserves the matched two-sided communication semantics of MPI with only minimal changes; as a result, common HPC communication patterns such as halo exchanges remain easy to implement. "
In this paper, researchers have proposed a design of a new GPU communication API that can support CPU-free communication and mostly preserves familiar MPI two-sided communication abstractions.
arxiv.org/pdf/2602.15356
The results on 11 models show that FlashMem achieves 2.0x to 8.4x memory reduction and 1.7x to 75.0x speedup compared to existing frameworks, enabling efficient execution of large-scale models and multi-DNN support on resource-constrained mobile GPUs.
FlashMem reduces peak memory usage while improving execution efficiency by statically scheduling weight loading and balancing computation with data movement at a fine granularity.
In this paper is introduced FlashMem, a memory streaming framework designed to efficiently execute large-scale modern DNNs and multi-DNN workloads while minimizing memory consumption and reducing inference latency.
arxiv.org/pdf/2602.15379
The results show that the fractional skyrmion number evolves between adjacent integers through an abrupt change, thereby bridging discrete integer topological orders and reinforcing the integer nature.
In this paper, researches have demonstrated a robust framework for generating and controlling fractional optical skyrmions mediated by fractional OAM modes, paving the way for high-density optical information encoding.
arxiv.org/pdf/2602.15464
The evaluation results show that PRISM significantly boosts post-fabrication performance and yield while reducing calibration area and turnaround time.
PRISM synthesizes compact, informative calibration patterns to minimize required fabrication data, trains a physics-gounded differentiable fabrication model and performs photonics-informed inverse mask optimization that prioritizes critical features beyond geometry matching.
In this paper is introduced PRISM, a photonics-informed neural inverse lithography framework that turns inverse-designed photonic integrated circuits layouts into high-yield, manufacturable hardware.
arxiv.org/pdf/2602.15762
The results show that Scope achieves up to 1.73x throughput improvement while maintaining similar energy consumption for ResNet-152 inference compared to state-of-the-art approaches.
"By merging layers, Scope balances workloads and mitigates pipeline bubbles. Its flexible partitioning and distributed storage optimize computation, communication, and memory access."
In this paper is presented Scope, a novel merged pipeline framework for MCM-based NN inference that enhances hardware utilization and reduces network-on-package overheads, while also mitigating segment count tradeoffs in segmented pipelines.
arxiv.org/pdf/2602.14393
The results show that, on average, GLM-5 achieves about 20% improvement over the last version GLM-4.7, and is comparable to Claude Opus 4.5 and GPT-5.2 (xhigh), and better than Gemini 3 Pro.
