With the development of large language models (LLMs), efficient inference through Key-Value (KV) cache compression has attracted considerable attention, especially for long-context generation. To compress the KV cache, recent methods identify critical KV tokens through static modeling of attention scores. However, these methods often struggle to accurately determine critical tokens as they neglect the *temporal patterns* in attention scores, resulting in a noticeable degradation in LLM performance. To address this challenge, we propose **AttentionPredictor**, which is the **first learning-based method to directly predict attention patterns for KV cache compression and critical token identification**. Specifically, AttentionPredictor learns a lightweight, unified convolution model to dynamically capture spatiotemporal patterns and predict the next-token attention scores. An appealing feature of AttentionPredictor is that it accurately predicts the attention score and shares the unified prediction model, which consumes negligible memory, among all transformer layers. Moreover, we propose a cross-token critical cache prefetching framework that hides the token estimation time overhead to accelerate the decoding stage. By retaining most of the attention information, AttentionPredictor achieves **13$\times$** KV cache compression and **5.6$\times$** speedup in a cache offloading scenario with comparable LLM performance, significantly outperforming the state-of-the-arts. The code is available at https://github.com/MIRALab-USTC/LLM-AttentionPredictor.

Synthetic aperture radar Interferometry (InSAR) has proven to be an effective landslide monitoring technique, especially for very slow landslides without observable morphological features. Integration of multi-source InSAR observations from different satellites/tracks could help reduce omissions and misjudgements of potential landslides, showing a promising trend toward automatic landslide detection and monitoring at regional or national scale. However, existing methods present poor error suppression performance, and are not capable of describing and solving potential information conflicts, due to the neglect of observation uncertainties during the integration. Here we propose a new integrated method based on Dempster-Shafer evidence theory to address these deficiencies. The two key steps are multi-source InSAR integration based on DST and a two-step decision rule. We apply the proposed method to the entire Jinsha River Basin using both ascending and descending Sentinel-1 SAR data from 2014 to 2023. The preliminary result on the Luoshui-Baini section based on two sources (ascending and descending orbits) identified 68 landslides, which is nearly the same between our method and the existing mosaic method.

Logic Synthesis (LS) aims to generate an optimized logic circuit satisfying a given functionality, which generally consists of circuit translation and optimization. It is a challenging and fundamental combinatorial optimization problem in integrated circuit design. Traditional LS approaches rely on manually designed heuristics to tackle the LS task, while machine learning recently offers a promising approach towards next-generation logic synthesis by neural circuit generation and optimization. In this paper, we first revisit the application of differentiable neural architecture search (DNAS) methods to circuit generation and found from extensive experiments that existing DNAS methods struggle to exactly generate circuits, scale poorly to large circuits, and exhibit high sensitivity to hyper-parameters. Then we provide three major insights for these challenges from extensive empirical analysis: 1) DNAS tends to overfit to too many skip-connections, consequently wasting a significant portion of the network's expressive capabilities; 2) DNAS suffers from the structure bias between the network architecture and the circuit inherent structure, leading to inefficient search; 3) the learning difficulty of different input-output examples varies significantly, leading to severely imbalanced learning. To address these challenges in a systematic way, we propose a novel regularized triangle-shaped circuit network generation framework, which leverages our key insights for completely accurate and scalable circuit generation. Furthermore, we propose an evolutionary algorithm assisted by reinforcement learning agent restarting technique for efficient and effective neural circuit optimization. Extensive experiments on four different circuit benchmarks demonstrate that our method can precisely generate circuits with up to 1200 nodes. Moreover, our synthesized circuits significantly outperform the state-of-the-art results from several competitive winners in IWLS 2022 and 2023 competitions.

Tropospheric delays have been a major limitation on the precision and accuracy of Interferometric Synthetic Aperture Radar (InSAR). InSAR data-based tropospheric delay correction methods, especially the local window methods, could estimate heterogeneous tropospheric delays in the same resolution as InSAR data, thus, are increasingly desired for modern high-resolution InSAR products. However, the phase-elevation relationship estimation at local windows can be severely contaminated by topography-correlated deformation. In this paper, we present a new InSAR phase-based tropospheric delay correction method based on texture correlation, which is shown to be relatively insensitive to topography-correlated deformation. The texture information represents the spatially high-frequency component of a two-dimensional image, which can be obtained using high-pass filtering. The method first produces a low-resolution tropospheric delay estimation using the window-wised texture correlation in the space domain, then refines it to high resolution by fitting the residual with the previously estimated tropospheric phase-elevation slope in the time domain. We apply the proposed method to ALOS-2 data over the Kirishima volcanic complex in Japan, encompassing typical topography-correlated deformation. The estimated phase-elevation slope shows seasonal oscillation in agreement with independent ERA5 prediction. The proposed method reduces the median spatial standard deviation of the residual phase from 1.3 cm to 0.7 cm, showing superior performance compared with other existing methods without compromising the deformation signal.

Benefiting from the wide coverage and high resolution of synthetic aperture radar (SAR) data, interferometric SAR (InSAR) has significant advantages in wide-area deformation detection and monitoring. In order to improve computational efficiency and save computational resources, it is expected to perform the deformation area identification first as accurately as possible, and then perform local time series inversion for the specific deformation area. However, with the interference of tropospheric delay, especially its systematic component, the accurate identification of the deformation area becomes challenging. To address this issue, we propose a two-step tropospheric delay removal method including time domain correction and spatial domain correction. The time domain correction is used to avoid the effect of the systematic component of tropospheric delay so as to derive an accurate deformation rate map. The purpose of the spatial domain correction is to finely remove the effect of tropospheric delay in local area to recover the correct deformation time series. Applying our method, external data based method and existing classical SAR data based method to the reservoir area of the Lianghekou hydropower station with Sentinel-1A ascending data for comparison, the results demonstrate the advantages of our method in deformation area identification and deformation monitoring.

Phase synchronization plays a crucial role in the LuTan-1 (LT-1) bistatic synthetic aperture radar (BiSAR) system, as it aims to eliminate additional azimuthal phase modulation caused by oscillator differences. However, for the pulse alternating transmission system operating in the L-band, the presence of radio frequency interference (RFI) poses an inevitable challenge. Serious RFIs introduce phase jitters, compromising the accuracy of synchronization. In this letter, an effective method for detecting and removing phase jitters is proposed to enhance synchronization accuracy. In the proposed method, the jitter features are separated by the iteratively reweighted least squares (IRLS) in the instantaneous frequency domain based on the established synchronization phase model. Then the jitter positions are detected by correlated peaks between the designed convolutional kernels and jitters. Finally, a polynomial model is utilized to remove jitters and assist in phase unwrapping. The synchronization phases acquired by the LT-1 mission are used to verify the feasibility of the proposed algorithm. The improved imaging quality demonstrates the effectiveness of the proposed method and confirms its ability to ensure the high-precision generation of the LT-1 BiSAR images.

The innovative bistatic synthetic aperture radar (BiSAR) mission LuTan-1 (LT-1) uses a noninterrupted synchronization scheme to achieve high-precision phase synchronization. While radio frequency interference (RFI) is a major factor in deteriorating phase synchronization performance. To present the effect of RFI on the synchronization phase clearly, the characteristics of RFI in the synchronization link are described in detail, and a precise analytic expression between the amplitude, frequency, and phase of RFI and synchronization phase error is established. Furthermore, a novel pulse-compression-based notch (PCN) method is proposed to eliminate the phase error introduced by RFI. In the proposed method, the saturated distortion signals resulting from strong interferences are detected and discarded by the distribution features of their modes. Then, inspired by contrary thinking, the synchronization signal after pulse compression is notched instead of RFI. A fast missing data iterative adaptive approach (Fast MIAA) is performed to recover the gaped RFI signal and remove it from the compressed signal. Finally, the correct synchronization phases can be extracted from the peak positions of the remaining signals. Experimental results derived from using simulated and real synchronization data of the LT-1 system validate the performance of the proposed RFI mitigation method.

An accurate baseline of synthetic aperture radar (SAR) interferometry (InSAR) is an important parameter for the geodetic application of the InSAR data. Although some advanced SAR satellites have precise orbit determination, there are still many SAR satellites suffering from baseline inaccuracies, such as GF-3. In this article, an image-based estimator for baseline correction is proposed, which requires only the external digital elevation model (DEM) data. The idea of the method is to project the orbit error phase onto the phase components carrying the baseline error information, which is called orbit error phase bases in this article, and to correct the baseline according to the projection coefficients. Since the pure orbit error phase is unavailable, the residual phase of the interferogram is used to approximate the orbit error phase, and a series of processes are introduced to weaken the effect of this approximation. Both the simulated and real data from GF-3 SAR are used to validate the proposed method, and a comparison with the conventional nonlinear least-square and the latest proposed flat-Earth phase-based baseline refinement methods are made. The results indicated the superior accuracy and robustness of our method, especially in areas with higher relief and wider coverage.