Reinforcement learning (RL) with rule-based rewards has recently shown remarkable potential for improving large language model (LLM) reasoning, with approaches like SWE-RL demonstrating that software evolution data can serve as an effective training signal. However, existing methods treat pull requests as independent, identically distributed samples and rely on flat sequence-similarity rewards, ignoring the rich structural information embedded in the codebase’s dependency graph and the natural difficulty gradient across different types of code changes. In this paper, we introduce Evo-CuRL (Evolutionary Curriculum Reinforcement Learning), a framework that addresses both limitations through two synergistic innovations. First, we construct a code lineage graph from software evolution data that captures inter-file dependencies and modification histories, enabling a graph-aware hierarchical reward that evaluates generated patches at three levels of granularity: local sequence similarity, structural AST consistency, and global dependency coherence. Second, we introduce a curriculum-aware training scheduler that adaptively samples training instances from easy (single-file edits) to hard (cross-module refactors) using a multi-armed bandit formulation, ensuring the model builds foundational patch-generation skills before tackling complex multi-file changes. Trained on top of Llama-3, our resulting model, Llama3-Evo-CuRL-70B, achieves a 45.6% resolve rate on SWE-bench Verified and 38.7% on SWE-bench Lite, establishing new state-of-the-art results among open-source models under 100B parameters. Comprehensive ablation studies confirm that both the graph-aware reward and the curriculum scheduler contribute meaningfully: removing either component degrades performance by 2.5–3.8 percentage points. Furthermore, like SWE-RL, our model exhibits emergent out-of-domain generalization, with even larger improvements on coding, mathematical reasoning, and general language understanding benchmarks, suggesting that structured RL on software data cultivates broadly transferable reasoning capabilities.

Persistent transactional memory (PTM) exploits transactions to provide an easy crash-consistent interface for persistent memory (PM). However, because of the substantial reader-side overhead brought on by the low bandwidth and long persistence latency of PM, present PTM research cannot scale effectively. This paper proposes a highly scalable PTM system, DRPTM, which allows nearly non-overhead reads without lowering the isolation level. DRPTM decouples persistence latency from concurrency control and traces the read-only copy maintained in logs as a lightweight read set. The evaluation shows that DRPTM significantly outperforms the state-of-the-art PTM systems for various workloads and achieves near-linear scalability.

High-resolution (HR) synthetic aperture radar (SAR) image classification is a challenging task for the limitation of its complex semantic scenes and coherent speckles. Convolutional neural networks (CNNs) have been proven the superior local spatial features representation capability for SAR images. However, it is hard to capture global information of images by convolutions. To solve such issues, this letter proposes an end-to-end network named global–local network structure (GLNS) for HR SAR classification. In the GLNS framework, a lightweight CNN and a compact vision transformer (ViT) are designed to learn local and global features, and two types of features are fused in quality to mine complementary information through the fusion net. Then, our research devolves the twofold loss function to reduce the interclass distance of SAR images, which brings more compactness to classification features and less interference of coherent speckles. Experimental results on real HR SAR images indicate that the proposed method has more strong feature extraction capability and noise resistance performance. This method achieves the highest classification accuracy on both datasets compared with other related approaches based on CNN.

Multifrequency (MF) polarization synthetic aperture radar (PolSAR) systems can obtain more abundant and continuous earth resource information than single-frequency ones and have been widely used in the remote sensing community. However, there are relatively a few researches for the fine MF PolSAR image classification, which is an important part of remote sensing image interpretation. The main focus currently is on the single-frequency part. Therefore, for dual-frequency PolSAR image classification, this article proposes the dual-frequency attention fusion network (DFAF-Net). It is based on frequency-aware attention and adaptive feature fusion to improve classification performance. First, the dual-frequency PolSAR data is input into the joint feature extraction (JFE) module to obtain the joint feature representation. Meanwhile, two frequency-aware attention block (FAB) modules with the same structure are constructed, which, respectively, generate frequency-specific attention masks based on the guidance of different frequency data. Subsequently, these masks are used to weigh the joint features to highlight the description of different frequency-aware importance. The activated frequency-aware features can fully mine and utilize the complementary information provided by different frequencies, thereby enhancing the discrimination of similar landcover categories. Finally, the adaptive feature fusion block (AFFB) module is utilized to adaptively aggregate different frequency-aware features multiple times, which can effectively eliminate information differences. The obtained fusion features are more compact within classes and separable between classes, thereby effectively improving the classification performance. Experiments on three measured spaceborne and airborne dual-frequency PolSAR datasets verify that DFAF-Net can better perceive frequency characteristics and fully mine the complementary. Therefore, the classification accuracy is effectively enhanced, and the inaccuracy of single-frequency classification can be eliminated. Meanwhile, due to the introduction of the attention module, the classification performance of the proposed DFAF-Net is more competitive than the related deep learning networks. Quantitatively, the overall accuracy of DFAF-Net on the three datasets is respectively 98.30%, 97.42%, and 99.45%, which is better than other methods.

The nonstationary and non-Gaussian distribution of the high-resolution (HR) synthetic aperture radar (SAR) image provides much valuable information. However, the current methods, especially deep learning models, directly learn spatial features from HR SAR data while ignoring global statistical information. Combining the local spatial features and global statistical properties of HR SAR images is urgently needed to capture complete HR SAR characteristics. In this paper, a feature fusion network (Fusion-Net) using both deep spatial context encoder and nonstationary joint statistical model (NS-JSM) is proposed for the first time. Fusion-Net realizes the fusion description of local spatial and global statistical features in an end-to-end supervised classification framework. First, a deep spatial context encoder network (DSCEN) is designed based on multiscale group convolution (MSGC) module and channel attention (CA) module. The DSCEN expands the scope of context information extraction with few parameters and increases the interaction between high- level feature channels. Then, the NS-JSM is adopted to capture the unique SAR statistical information. Specifically, the SAR image is transformed into the Gabor wavelet domain. The produced sub-band magnitudes and phases are modeled by the log-normal and uniform distribution. The covariance matrix (CM) is calculated for mapped sub-band data to capture the interscale and intrascale nonstationary correlation. Finally, the group compression and smooth normalization units are introduced into Fusion-Net to fuse the statistical features and spatial features, which not only exploits the complementary information between different features but also optimizes the fusion feature representation. Experiments on four real HR SAR images validate the superiority of the proposed method over other related algorithms.

A pixel-wise classification for high-resolution (HR) synthetic aperture radar (SAR) images is a challenging task, due to the limited availability of labeled SAR data, as well as the difficulty of exploring context information affected by coherent speckle. In this article, we propose a novel supervised classification method for HR SAR images, which combines a context-aware encoder network (CAEN) and a hybrid conditional random field (HCRF) model. First, a new CAEN architecture is developed based on the intrinsic property of HR SAR pixel-wise labeling. The proposed architecture follows an encoder-decoder structure, wherein the residual context encoder (RCE) block and the global context-aware (GCA) block are proposed in the encoder module to capture local to global semantic contexts. The multiscale skip connections and feature compression structures are designed in the decoder module to preserve precise object structures while improving computational efficiency. Then, a patch sampling strategy is adopted to ensure that the training and test data are completely separated. It can achieve a less-biased estimate of the test error of CAEN. In addition, the overlapped sampling and data augmentation techniques are used to solve the problem of limited labeled data. Finally, the HCRF model is constructed and combined with the previous CAEN to further enhance the spatial label consistency. Our HCRF integrates pixel-level and region-level potentials into a unified Bayesian framework, making two spatial supports come to a more accurate decision on pixel categories. Experiments on four HR SAR images validate the superiority of the proposed method over other related algorithms.

The registration of synthetic aperture radar (SAR) and optical images is a challenging task due to the potential nonlinear intensity differences between the two images. In this paper, a novel image registration method, which combines nonlinear diffusion and phase congruency structural descriptor (PCSD), is proposed for the registration of SAR and optical images. First, to reduce the influence of speckle noise on feature extraction, a uniform nonlinear diffusion-based Harris (UND-Harris) feature extraction method is designed. The UND-Harris detector is developed based on nonlinear diffusion, feature proportion, and block strategy, and explores many more well-distributed feature points with potential of being correctly matched. Then, according to the property that structural features are less sensitive to modality variation, a novel structural descriptor, namely, the PCSD, is constructed to robustly describe the attributes of the extracted points. The proposed PCSD is built on a PC structural image in a grouping manner, which effectively increases the discriminability and robustness of the final structural descriptor. Experimental results conducted on SAR and optical image pairs demonstrate that the proposed method is more robust against speckle noise and nonlinear intensity differences and improves the registration accuracy effectively.

Triplet Markov fields (TMF) model is widely used to deal with non‐stationary synthetic aperture radar (SAR) images. However, its ability to capture global information remains limited due to the non‐causal property. A hierarchical TMF model is proposed in this study based on the non‐linear diffusion (ND) strategy, which is denoted as ND‐hierarchical TMF (HTMF). ND is adopted to generate multiscale decomposition according to local image content, and that is superior to traditional wavelet decomposition in reflecting hierarchical nature of image structure and detailed features. The auxiliary field in ND‐HTMF is redefined and initialised on the finest scale to characterise edge information and that enhances the prior modelling ability for non‐stationary local image features. The multiscale likelihood and multiscale causal prior energy functions are then defined respectively in bottom‐up and top‐down procedures to capture local and global information for performing segmentation. Segmentation experiments on simulated and real SAR images demonstrate the effectiveness of ND‐HTMF in both edge characterisation accuracy and robustness against speckle noise.