Zero-knowledge virtual machine (zkVM) is a powerful infrastructure for proving the correctness of a program execution with a succinct proof, attracting significant interest from researchers, developers, and users. It has been widely used in applications such as blockchain rollups, privacy-preserving machine learning, and off-chain computation. As the field grows, a wide range of zkVMs have been proposed. However, they adopt different choices in instruction formats, trace layouts, and proving backends, which results in a highly heterogeneous design landscape and makes it difficult to understand the relations among these systems.To bridge this gap, we provide a comprehensive study of zkVMs that covers both their theoretical foundations and practical implementations. We decompose zkVMs into three layers: (1) the ISA layer, which defines instruction semantics and determines the structure of the execution trace, (2) the VM layer, which captures program execution and organizes constraints through modular circuit components, and (3) the proving layer, which converts execution traces into algebraic constraints and generates the final proofs. This decomposition allows us to isolate the role of each layer while also examining how they interact in real systems. To give readers a more direct understanding of how these design choices affect performance, scalability, and usability, we conduct a comprehensive experimental evaluation of representative zkVMs following this layered framework. Finally, we conclude the paper by summarizing the main observations from our analysis and outlining several potential directions for zkVM design and implementation.

Smart contracts have expanded blockchain applications, but permissioned blockchain systems face severe through-put challenges, especially with the increasing complexity of nested contract transactions. These transactions, involving cross-contract interactions and deep call chains, intensify execution conflicts and rollback overhead, ultimately limiting parallelism. We propose Loom, a deterministic execution framework that enhances the efficiency of nested contract transactions. Loom employs snapshot-based concurrent pre-execution to decompose transactions into fine-grained subtransactions. To reduce rollback overhead, it introduces a two-phase rollback algorithm to minimize computational redundancy and fine-grained rescheduling to improve subtransaction-level parallelism during re-execution. Additionally, a multi-phase parallelism mechanism optimizes resource utilization across transaction blocks. Experimental results show that Loom achieves 6.1 × to$10.2\times$higher throughput while reducing rollback overhead by 89.9% to 98.4%, significantly outperforming state-of-the-art solutions.

As synthetic aperture radar (SAR) technology continues to evolve with a focus on miniaturization, reduced weight, and lower costs, its range of applications has broadened to encompass lightweight and slow-moving platforms, such as small autonomous aerial vehicles (AAVs) and airships. However, this advancement introduces a new challenge for SAR systems in the form of excessively long synthetic aperture time (LSAT). LSAT-SAR faces significant challenges caused by its longer integration time (10–1000 times more than conventional airborne SAR’s synthetic aperture time), including more stringent error tolerance of navigation system and more demanding trajectory control requirements. These issues often result in severe 2-D defocusing. Existing high-precision navigation systems and traditional autofocus algorithm, though adequate for conventional airborne SAR, often fail to meet the stringent requirements of LSAT-SAR, especially in low-contrast scenarios, necessitating more efficient and more robust compensation methods. To address these challenges, we propose a 2-D autofocus algorithm for LSAT-SAR using deep neural networks (DNNs). The proposed approach treats the 2-D focusing problem as a series of coupled 1-D curve estimation tasks, employing a weighted entropy loss function. The solution is optimized in an unsupervised manner using a DNN. Finally, the additional refinement is performed through a specialized fine-tuning correction module. Experiments on real LSAT-SAR images with synthetic aperture times of 150–270 s show that the proposed method substantially exceeds the performance of traditional approaches in low-contrast scenarios. It demonstrates excellent robustness, offering valuable insights and technical guidance for LSAT-SAR autofocusing.