ArchiveAs semiconductor technology scales, chip physical design confronts escalating complexity and multi-objective optimization challenges, with traditional electronic design automation (EDA) tools encountering bottlenecks in computational efficiency and convergence. Artificial intelligence (AI), specifically deep learning (DL) and reinforcement learning (RL), offers a transformative paradigm for data-driven optimization in physical design. This survey systematically reviews the application of AI methods in the four core stages of chip physical design: floorplanning, placement, clock tree synthesis, and routing. First, from the perspective of optimization objectives, representative AI methods and their technical contributions in each stage are summarized. Second, a critical analysis of the gap between academic datasets and industrial scenarios is conducted, revealing key data bottlenecks. Third, the integration approaches between AI methods and traditional EDA flows across different stages are examined. Then, the evolutionary trends of AI methods in each stage are systematically reviewed, highlighting the paradigm shift from isolated, geometry-driven optimization toward cross-stage collaboration and AI-native chip physical design. Although AI methods have significantly improved design efficiency and quality, their widespread adoption is still limited by data scarcity, insufficient model interpretability, and challenges in integration with existing toolchains. Finally, we discuss the limitations of existing AI-driven approaches in physical chip design and outline promising research directions, including open-source AI-EDA infrastructure, large circuit foundation models, and AI-native EDA.
Deep learning has recently demonstrated strong capabilities in pattern discovery and structural inference, offering new possibilities for algorithm design—a field traditionally reliant on human experience and intuition. We focus on the core objective of automatically learning interpretable algorithmic rules under unsupervised settings, using two representative combinatorial optimization problems as cases: course scheduling problem (CSP) and traveling salesman problem (TSP). We propose two deep reinforcement–based methods, Neural-Greedy and Neural-3-OPT. For CSP, Neural-Greedy autonomously learns three optimal greedy rules from large collections of problem instances: prioritizing courses with the earliest finishing times, the latest starting times, and a hybrid of the two. Its underlying mechanism lies in parameter modulation and neuron cooperation, which precisely focuses on critical temporal features and enables the simultaneous acquisition of multiple rules. For TSP, Neural-3-OPT learns efficient 3-opt heuristic autonomously from large collections of problem instances, achieving optimality gaps as low as 0.00%, 0.08%, and 0.74% on random TSP20, TSP50, and TSP100 datasets, respectively. It significantly outperforms the industrial solver OR-Tools and maintains superior performance on the real-world TSPLIB benchmark. This advantage stems from an intelligent decision mechanism of early-stage precise selection, which prioritizes inefficient edges with substantial impact on global solution quality and identifies reconnection strategies adaptive to the entire iterative process, thereby avoiding myopic greedy choices and ineffective iterations to achieve fast convergence and high-quality solutions. Overall, our results demonstrate that deep learning can not only approach optimal solution quality but also autonomously learn interpretable algorithmic rules, offering a data-driven paradigm for algorithm design and providing important insights into the development of AI-assisted algorithm design.
In recent years, deep learning has achieved remarkable progress in meteorological forecasting and climate modeling, exerting a growing influence on agricultural applications. While numerous review articles have systematically examined related methodologies, most focus on model architectures, task types, or application domains, with limited attention given to the agricultural perspective. This paper provides a comprehensive review of the development of large meteorological models (LMMs), structured around model architecture. We trace the evolution from traditional numerical methods to deep learning approaches and categorize existing models into six major classes: classical deep neural networks, graph neural networks, Transformer-based architectures, state space models, generative models, and physics-informed/multimodal fusion models. Furthermore, we summarize representative applications of LMMs in agriculture, including meteorological forecasting services, extreme weather early warning and disaster mitigation, and decision support systems for precision agriculture. We also examine commonly used datasets and evaluation metrics in the field. This is followed by a discussion of current challenges, particularly in areas of generalization capability, uncertainty quantification, and physical consistency. Finally, we outline key research directions for future exploration in this rapidly evolving field. This work aims to provide a systematic reference for advancing both the research and application of LMMs in agricultural contexts.
To construct genomic functional-element sequence generation models for major staple crops and evaluate their cross-species generalization ability, this study investigates fine-tuning and performance assessment of the Evo2 genomic language model for conditional sequence generation. Six representative crop genomes, including potato, peanut, indica rice cultivars (MH63 and ZS97), japonica rice, and wheat, are selected. A unified conditional generation and evaluation framework is established for six core genomic functional elements, including coding sequences (CDS), exons, introns, mRNA, 5′ untranslated regions (5′ UTR), and 3′ untranslated regions (3′ UTR). Based on this framework, Evo2 models are fine-tuned on crop-specific genomic data, while the pretrained Evo2 model without fine-tuning is retained as a zero-shot baseline. Experimental results show that fine-tuning consistently improves generation performance across all functional elements and crop species, an average absolute gain of about 0.49 percentage points, corresponding to a relative gain of approximately 0.79%. Cross-species evaluation further demonstrates that the fine-tuned models maintain good generalization ability when applied to unseen crop species, and the cross-species generalization gap remains below 1 percentage point for all six functional elements. Among them, coding-related elements such as CDS and mRNA show the most stable generation performance across species, whereas 5′ UTR is more sensitive to species differences. Additional bioinformatics analyses indicate that the generated sequences exhibit some degree of biological plausibility in ORF continuity, but still show limitations in codon preference, regulatory motifs in UTRs, and splice-site rules in introns. Overall, the results verify the effectiveness and cross-species applicability of Evo2-based fine-tuned genomic language models for functional-element sequence generation in major staple crops, providing a potential technical basis for genome-assisted breeding and crop improvement.
Ribonucleic acid (RNA) plays a central role in the regulation of biological processes, and its systematic investigation is essential for understanding disease mechanisms at molecular level and developing innovative therapeutic strategies. Nevertheless, progress in RNA research has been hindered by the structural and functional complexity of RNA molecules, the explosive growth of high-throughput sequencing data, and the intrinsic limitations of conventional experimental techniques. In recent years, artificial intelligence (AI) has emerged as a powerful enabling technology, offering novel solutions through its strengths in pattern recognition, high-dimensional data analysis, and predictive modeling. This review provides a comprehensive overview on the major applications of AI in RNA research, encompassing RNA sequence element optimization, RNA modification sites prediction, RNA secondary/tertiary structures prediction, and computational modeling of RNA-protein interactions. Furthermore, the key challenges facing AI-driven RNA studies are critically examined, including the limited availability of high-quality annotated datasets and the insufficient interpretability of complex learning models. Finally, future perspectives are outlined, emphasizing integrated data resource development, algorithmic and model-level innovations, and the construction of closed-loop translational frameworks, with the goal of fostering deeper integration between AI methodologies and RNA biology, accelerating both fundamental discoveries and the development of next-generation RNA-based therapeutics.
Machine learning interatomic potentials (MLIPs) bridge the gap between microscopic quantum mechanics and macroscopic property simulation by fitting high-dimensional potential energy surfaces. While maintaining near-first-principles accuracy, MLIPs extend molecular dynamics simulations to significantly larger spatial and temporal scales. Driven by the synergy of “model, training, and evaluation”, where diverse architectures need tailored training strategies, efficient optimizations can accelerate model development, and comprehensive task evaluation ensures the MLIPs’ reliability. This paper provides a systematic review of the recent representative advances in MLIPs’ architectures, training optimizations, and evaluation benchmarks. For model architecture, MLIPs can be categorized into invariant, equivariant, and unconstrained models based on their symmetry-keeping approach. We then analyze the evolution from “specialized models” to “pretrained models” and introduce several representative models. On the training strategies, we discuss key techniques such as framework-free implementations and kernel fusion for specialized MLIPs, alongside load balancing and two-stage training strategies for pretrained MLIPs. Finally, we list the evaluation metrics of modern benchmarks such as Matbench Discovery and FAIR Chemistry Leaderboard (OMol25).
Zoned namespaces solid-state drive (ZNS SSD) mitigates the performance overhead of the conventional “block interface tax,” opening new opportunities for high-performance storage system design. At the same time, ZNS SSDs fundamentally reshape the storage model by exposing zones to the host, enforcing sequential writes within each zone, and offloading flash-management functions—such as garbage collection and address mapping—from the device to host software. These shifts introduce significant challenges for ZNS-based systems. This survey distills three key challenges: 1) write amplification induced by host-managed placement and reclamation, where co-locating mixed-lifetime data within zones increase GC copying, compaction-GC mismatch leads to redundant data migration, and inefficient zone-reset policies cause unnecessary rewrites; 2) limited parallelism under fixed zone/resource mapping and sequential-write constraints, further exacerbated by intra- and inter-zone contention; and 3) elevated consistency and reliability overheads arising from wear imbalance, crash-recovery metadata costs, and parity maintenance in multi-device settings. For each challenge, we organize state-of-the-art optimization techniques, summarize recent advances, and highlight representative studies. We also review widely used simulators/emulators and auxiliary tools that support ZNS prototyping, debugging, and performance tuning. Finally, we discuss promising research directions, including cross-layer hardware-software co-design, multi-level heterogeneous integration, systematic software-stack refinement, and holistic, system-wide optimization.
With the development of the RISC-V open-source ecosystem, driving domain-specific architectures (DSAs) with custom extension instructions based on RISC-V has become a widely recognized design approach. However, under traditional hardware description languages and design paradigms, integrating DSAs with processors and completing instruction design require significant manpower and time costs to understand processor design details and modify them. Currently, some work supports developers in accessing DSAs by implementing fixed data paths and interfaces in sequential processors. Yet, for high-performance out-of-order processors, there is no comprehensive solution available. We summarize the key issues that need to be addressed through practice when driving and integrating DSAs with out-of-order processors using extension instructions. We design a hardware programming framework for instruction extension targeted at high-performance out-of-order RISC-V processors, assisting developers in integrating DSAs and completing instruction extensions on out-of-order processors. The framework is implemented in Chisel, applying object-oriented and functional programming features to flexibly describe hardware and assist in completing signal connections, thereby saving developers’ effort. Our evaluation shows that with the help of this framework, developers can save up to
With the rapid development of deep learning applications, deep learning processors (DLPs) such as TensorCore GPU, TPU, MLU have risen rapidly and demonstrated excellent performance in matrix operations. To efficiently handle complex matrix data, these processors typically partition sub-matrices and load them into on-chip memory with fixed stride, establishing a distinct pattern of strided memory access. However, devices like NVIDIA H100, Huawei Ascend, and Cambricon MLU currently depend on complex explicit conflict management, lacking dedicated hardware controls for strided memory access conflicts, thus making it challenging to ensure the atomicity and consistency of instruction execution. Traditional memory conflict detection methods focus on scalar operations and have difficulty capturing dependency conflicts among matrix data. To address this issue, we propose a strided matrix memory access conflict determination method that transforms conflict detection into the problem of solving a binary linear Diophantine equation, thereby accurately determining the dependency relationships between memory access instructions. Simulation experiments demonstrate that this approach markedly enhances the memory performance of typical operators in large language models: memory bandwidth utilization can peak at 94% (LPDDR) and 91% (HBM), while the average determination overhead is reduced to 4.68% and 0.31% when compared with the Power test and set intersection methods, respectively. In addition, hardware evaluation based on a 12 nm process shows that the determination design occupies only
Large language models (LLMs) have demonstrated remarkable capabilities across various domains, including natural language processing, computer vision, chat bots and content generation. Deploying LLMs on edge devices has attracted significant attention due to its advantages in service stability and data privacy. However, the decoding inference of LLMs is composed of memory-bound operations, and with model sizes growing and complex tasks demanding increasingly longer context lengths, the resulting memory access overhead becomes critical bottleneck limiting the deployment of LLMs on edge platforms. To lower the memory access demand of long context decoding, locality aware attention mechanism is proposed. By leveraging the numerical locality of attention scores during decoding, this mechanism compresses the key-value cache (KV cache), significantly reducing redundant memory access without compromising model accuracy. We present a locality-aware inference accelerator for LLMs with hybrid-bonded 3D DRAM. The accelerator not only reduces the overall memory traffic but also significantly improves the main memory bandwidth. It integrates load-balancing and prediction-based prefetching mechanisms tailored to the structural characteristics of hybrid-bonded 3D DRAM, thereby enhancing the bandwidth utilization. On the evaluated models, the proposed accelerator achieves on average
The rapid expansion of social networks has made data transmission over public channels ubiquitous, raising serious security and privacy concerns. Although encryption can guarantee confidentiality, its distinctive format may arouse suspicion. Steganography offers a covert alternative by hiding secret data within ordinary media, such as images, text, or audio, thereby reducing the risk of detection. Traditional pixel-editing steganographic methods are vulnerable to steganalysis or compression artifacts, while existing generative approaches often depend on tailored prompts or specific data sources, limiting their generalization and flexibility. To address these challenges, we propose a source-agnostic, target-domain image steganography framework based on a dual diffusion model. First, this method employs a latent diffusion model (LDM) to invert and optimize arbitrary secret images, obtaining high-fidelity latent representations. To eliminate residual structural cues from the LDM inversion that could compromise subsequent hiding and extraction, this method introduces a Gaussian structural reordering alignment module which encrypts the secret’s latent representation via pseudo-random orthogonal perturbation and spatial shuffling, preserving its norm while enforcing a Gaussian-like distribution that aligns with downstream inputs and enhances security. Finally, we use a denoising diffusion implicit models (DDIM) sampler to denoise and decode the encrypted latent vectors, producing stego images that faithfully reflect the target domain’s appearance while carrying the hidden content. Experiments on four public datasets demonstrate that our method requires no source-specific training, achieves strong cross-domain generalization, and significantly outperforms competing techniques in extraction accuracy, imperceptibility, and resistance to steganalysis.
Granular-ball generation (GBG) serves as the foundation for various learning tasks in granular-ball computing (GBC), such as classification and clustering. However, existing adaptive GBG methods introduce significant uncertainty due to the strategy of randomly selecting the division center, resulting in a highly unstable granular-ball classification result. Moreover, the k-division based adaptive GBG method requires repeated overlap detection and de-overlapping of granular balls, which significantly increases the computational overhead during the iterative process and impairs the efficiency of GBG. To address the above issues, a new method for stable, accelerated and adaptive GBG (SAAGBG) is proposed in this paper, building upon the existing k-division strategy. On the one hand, SAAGBG selects the sample point with the smallest difference with the median vector of the data cluster as the division center, which effectively enhances the robustness to outliers and avoids the inconsistency and instability caused random center selection. On the other hand, a new adaptive division condition based on granular-ball coverage is introduced into SAAGBG, which replaces the minimum quality threshold and overlap detection steps in the original k-division based GBG method, significantly improving the efficiency of the GBG process. Experimental results on benchmark datasets and datasets with varying noise ratios demonstrate that SAAGBG achieves greater stability and efficiency without loss of classification accuracy compared with the mainstream adaptive GBG methods.
In multi-agent reinforcement learning (MARL), partial observability poses a significant challenge to effective collaboration among agents. Although the centralized training with decentralized execution (CTDE) paradigm alleviates this challenge by leveraging global state information during training, agents must still rely solely on local observations during execution. To address this limitation, existing research has primarily followed two directions. The first direction introduces communication mechanisms to enable information sharing among agents; however, these methods often suffer from high communication overhead and redundant information. The second direction leverages mask reconstruction techniques to infer the missing global information; however, due to sample sparsity, the inferred information depends only on local observations, resulting in low inference accuracy and potential interference with decision-making. To overcome these limitations, we propose an active communication masked reconstruction method. During centralized training, the method dynamically groups agents based on global observation trajectories and allows each agent to select the most relevant communication targets from these groups. During decentralized execution, agents proactively communicate according to their selected target sets, enabling accurate reconstruction of global information and ultimately supporting their decision-making. Extensive experiments conducted on the StarCraftⅡ Multi-Agent Challenge (SMAC) demonstrate the effectiveness of the proposed approach in improving coordination and overall performance in partially observable multi-agent scenarios.
In recent years, large language model (LLM) has shown potential in multivariate time series forecasting. However, existing cross-modal methods based on LLM still face the challenges of insufficient modality alignment and limited generalization capability in few-shot learning and long-term forecasting tasks. To address the above issues, we propose an enhanced cross-modal LLM fine-tuning framework, termed Enhanced CALF, which integrates dynamic attention and hierarchical distillation. Firstly, the framework constructs a dynamic attention cross-modal matching module, introducing adaptive weight generation and alignment prediction mechanisms. This module can dynamically adjust the allocation of attention weights based on data distribution characteristics and inter-modality correlation strengths, thereby improving the precision of cross-modal feature alignment. Secondly, a multi-level knowledge distillation and contrastive learning module is built. By introducing projection mappings at each Transformer layer and combining hierarchical distillation loss with an adaptive temperature contrastive loss, it achieves hierarchical feature transmission covering aspects from local details to global semantics, aiming to enhance the consistency of cross-modal representations. Finally, an adaptive alignment mechanism is designed, which dynamically adjusts the weights of the total loss function through quantitative evaluation of inter-modality alignment scores, thereby optimizing the model training process. Experiments on seven real-world datasets demonstrate that Enhanced CALF outperforms existing baseline models in long-term forecasting and few-shot learning tasks.
Compilers can introduce security vulnerabilities during the compilation process. These compiler-introduced vulnerabilities are often subtle and can impact critical software such as operating system kernels, databases, and browsers, posing significant security risks. Nowadays, various softwares, including compilers, are rapidly iterated and updated, which means that new vulnerabilities can be introduced into important software binary code by compilers at any time. Our research indicates that such vulnerabilities often persist in software for a considerable period before being addressed. Therefore, timely detection of newly introduced compiler vulnerabilities is crucial for ensuring software security. There has been extensive research on detecting security vulnerabilities introduced by compilers. However, existing detection methods often focus on specific known vulnerability types and struggle to efficiently handle the frequent updates of compilers and the programs being tested. Consequently, we propose a novel detection technique based on general security code modeling and differential testing, which can efficiently identify security vulnerabilities introduced by compilers in the target program before and after version updates (either in the compiler or source code). This method identifies security code through error handling functions and determines compiler-introduced vulnerabilities by examining the differences in the removal of security checks across different versions. Evaluation experiments conducted on the Linux kernel demonstrate that our approach achieves a low false negative rate, maintains a practical false positive rate, and efficiently detects security vulnerabilities newly introduced by compilers during the compilation process of binary code updates.