Large Language Models (LLMs) fine-tuning techniques not only improve the adaptability to diverse downstream tasks, but also mitigate adverse effects of model quantization. Despite this, conventional quantization suffers from its structural limitation that hinders flexibility during the fine-tuning and deployment stages. Practical on-device tasks demand different quantization precisions (i.e. different bit-widths), e.g., understanding tasks tend to exhibit higher tolerance to reduced precision compared to generation tasks. Conventional quantization, typically relying on scaling factors that are incompatible across bit-widths, fails to support the on-device switching of precisions when confronted with complex real‑world scenarios. To overcome the dilemma, we propose OTARo, a novel method that enables on-device LLMs to flexibly switch quantization precisions while maintaining performance robustness through once fine-tuning. OTARo introduces Shared Exponent Floating Point (SEFP), a distinct quantization mechanism, to produce different bit-widths through simple mantissa truncations of a single model. Moreover, to achieve bit-width robustness in downstream applications, OTARo performs a learning process toward losses induced by different bit-widths. The method involves two critical strategies: (1) Exploitation-Exploration Bit-Width Path Search (BPS), which iteratively updates the search path via a designed scoring mechanism; (2) Low-Precision Asynchronous Accumulation (LAA), which performs asynchronous gradient accumulations and delayed updates under low bit-widths. Experiments on popular LLMs, e.g., LLaMA3.2-1B, LLaMA3-8B, demonstrate that OTARo achieves consistently strong and robust performance for all precisions.

The performance of multi-turn, agentic LLM inference is increasingly dominated by KV-Cache storage I/O rather than computation. In prevalent disaggregated architectures, loading the massive KV-Cache from external storage creates a fundamental imbalance: storage NICs on prefill engines become bandwidth-saturated, while those on decoding engines remain idle. This asymmetry severely constrains overall system throughput.

Due to the escalating demand to analyze large graphs, many organizations are now collecting billion-level property graph datasets, concurrently executing many complex graph queries against them, and expecting interactive-level response latency. However, such requirements are particularly challenging because of the notoriously irregular data access pattern and complex dependencies between heterogeneous subtasks. Despite the widespread availability of many-core CPUs and high-speed networking in modern datacenters, existing distributed graph query systems struggle with their inherent inefficiencies, resulting in low hardware utilization and poor query performance on these state-of-the-art hardware. To address these challenges, we introduce the Partitioned Stateful Traversal Machine (PSTM), which extends the Gremlin graph traversal machine. PSTM retains the expressive power of the Gremlin query language, enabling it to accommodate a wide range of graph query tasks, including traversal, pattern matching, filtering, and result aggregation. It additionally introduces query memoranda, allowing for more efficient implementation and execution of numerous graph queries in distributed environments. Moreover, PSTM facilitates various system-level optimizations, such as massively parallel execution, overlapping computation with communication, locality-aware data access, and lightweight progress tracking. Building upon PSTM, we develop GraphDance, a distributed graph database featuring an efficient asynchronous PSTM run-time. Our evaluations, conducted on an 8-node cluster, show that GraphDance achieves millisecond-level query latency for complex queries on terabyte-scale graphs, with an average latency reduction of 89.2% across all interactive complex queries in the LDBC SNB benchmark compared to existing distributed graph query systems.

Due to the sparse nature of Mixture-of-Experts (MoE) models, they are particularly suitable for hybrid CPU/GPU inference, especially in low-concurrency scenarios. This hybrid approach leverages both the large, cost-effective memory capacity of CPU/DRAM and the high bandwidth of GPU/VRAM. However, existing hybrid solutions remain bottlenecked by CPU computation limits and CPU-GPU synchronization overheads, severely restricting their ability to efficiently run state-of-the-art large MoE models, such as the 671B DeepSeek-V3/R1.

Despite the long-term records of satellite observations, factors controlling the seasonal and interannual variations in canopy reflectance remain poorly understood. Leaf optical properties (LOP, including leaf reflectance and transmittance) changes as leaves age, and thus impact the seasonal pattern of canopy reflectance, i.e., the “leaf age effect”. Here, we combined the Geometric Optical Radiative Transfer (GORT) model with continuous field measurements of leaf- to stand-scale characteristics to simulate canopy reflectance in a Chinese fir plantation with stand development (1-33 yr). We found that canopy structure controls the variations in canopy reflectance during young stages (<; 10 yr) and that leaf age controls the variations in canopy reflectance after canopy closure. Moreover, we found that the “leaf age effect” get enhanced with stand development, with R2 increased from about 0.1 to 0.56, 0.67, 0.92, and 0.82 for young, half-mature, near-mature, and mature stages, respectively. This study reveals the stand age dependence of leaf age effect on canopy reflectance which improves our interpretation and understanding of satellite observations to study the ecosystem function of forests.

A passive ultra-high frequency (UHF) RF identification (RFID) system is proposed that employs multiple antennas at the reader and single antenna at each tag. The gain due to multiple antennas in terms of the maximum reader interrogation range is first quantified. Then, a blind adaptive beamforming (BABF) algorithm that does not require the channel state information (CSI) is presented to improve both the interrogation range and the data transmission performance of the system in both the full- and half-duplex configurations. Moreover, an estimator for the number of tags within the reader interrogation range is proposed to enhance the system throughput. Simulation results show that both the interrogation range and the packet error rate (PER) performance for data transmission can be improved significantly using multiple antennas at the reader. A Universal Software Radio Peripheral (USRP)-based prototype system is implemented to validate the interrogation range improvement by employing multiple-antenna readers with several beamforming schemes. The proposed multiantenna RFID reader with the BABF complies with the existing RFID standard and can be readily implemented in existing systems.

The ill-posed inversion problem is to be solved urgently. Priori knowledge is introduced to increase the inversion information to improve the inversion quality. MODIS time-series leaf area index (LAI) data is used to extract the priori knowledge. Savitzky-Golay (SG) filter and bi-Gaussian curve-fit are performed to reconstruct the original time-series LAI data, then, the upper envelope of the smoothed time-series curve is achieved to get a fixed range of LAI in a certain growing season. The variation of LAI was restricted with the range based on Look-up Table (LUT) method with the PROSAIL model. The validation results show that the priori knowledge extracted from LAI time-series data is efficiency on improving the LAI inversion precision.

This paper extends the standard support vector regression (SVR) into multi-dimensional case to estimate different biophysical parameters simultaneously. The improvement is made by the Vapnik loss function of L2 form. The proposed multi-output SVR (MO-SVR) is implemented in the joint inversion of Leaf Area Index (LAI) and vegetation cover fraction (fCover) over SPOT2/HRV1 data in the Fundulea site (VALERI), Romania. Comparison between standard SVR and MO-SVR, and the validation using biophysical maps both indicate the better fitness and accuracy of MO-SVR than the traditional form.