ArchiveThis paper discusses the expansion and transition of computing paradigms. Firstly, it expounds on the great historical significance and limitations of Turing computation, defines what computing paradigm is, distinguishes between four types of computing paradigms: CS1.0, CS1.x, CS2.x and CS2+, and points out that computing paradigms evolve from “discrete, algorithm-based, rule-fixed, externally-assigned semantics, and closed monomers” to “continuous, dynamical systems, rule generation, endogenous semantics, and open ecosystems” in multiple dimensions, moving towards open evolutionary computation. Next, it is emphasized that the distinguishing criterion between CS1.x and CS2.x is the rule closure and semantic closure; It introduces typical cases of open evolutionary computing paradigms, including generative AI, structural computing, and institutional computing; emphasizes that the core issue of open evolutionary computation is governability. Finally, it discusses the theoretical basis and key technologies of open evolutionary computation, clarifies the significant importance of conducting basic research on open evolutionary computation, and calls on academia and industry to pay attention to the transition of this technological paradigm.
With the rapid development of deep neural network (DNN) accelerators, building simulators for new architectures is costly and time-consuming. Although advances in large language models (LLMs) have opened possibilities for automated simulator generation, existing approaches suffer from limited generality, inability to construct complete systems, and high construction complexity. To address these challenges, we propose SimulatorGen, a multi-agent framework that generates DNN accelerator simulator code from natural language descriptions. First, we abstract the architecture of DNN accelerator simulators and extract twenty-three component specifications. Based on this abstraction, four collaborative agents are introduced to accomplish generation: the analyst agent retrieves domain knowledge from the simulator library via retrieval-augmented generation (RAG) and constructs structured prompts by leveraging chain-of-thought (CoT) reasoning; the coder agent generates or refines code using prompts and test feedback; the tester agent performs syntax checking, functional testing, and formal verification using the Z3 solver based on properties extracted from specifications; and the assembly agent conducts component integration, automated execution, and metric comparison to enable end-to-end construction. We evaluate SimulatorGen on twenty-three generation tasks covering diverse DNN accelerator modules and architectures. Experimental results show that SimulatorGen built on GPT-4o outperforms LLM baselines, including Claude-Sonnet-4, achieving a Pass@1 score of 82.39%. Furthermore, using the successfully generated components, we construct runnable simulators for tensor processing unit (TPU) and MAERI architectures. Compared with STONNE, the simulators built by SimulatorGen achieve relative errors ranging from 1.31% to 7.34% in energy, latency, and energy-delay product (EDP) across multiple DNN models, while maintaining functional consistency verified through testing and execution, demonstrating faithful modeling of accelerator behavior. In contrast to the single-agent SimulatorCoder, which only supports module replacement, SimulatorGen enables end-to-end generation of complete simulators, further validating the effectiveness of the proposed approach.
Register transfer level (RTL) code generation based on large language models (LLMs) has emerged as a promising research direction, as it targets the least automated stage in current chip design workflows. However, the substantial semantic gap between high-level specifications and RTL, combined with limited training data, poses significant challenges for generation accuracy. A natural approach is to leverage human design experience by integrating design and verification, yet RTL test data are even scarcer than design data, making this strategy less LLM-friendly. In contrast, LLMs exhibit stronger capabilities in high-level languages such as Python or C, which are more suitable for functional specification and appear promising for RTL verification. Nevertheless, a large semantic gap remains, with notable differences in spatiotemporal granularity between high-level languages and hardware code. Verifying RTL using Python or similar high-level languages requires the LLM not only to understand high-level functional semantics but also to ensure that low-level timing and operational details match the circuit behavior, which is nontrivial. To address this challenge, we propose function-abstracted verification for RTL generation with ReAct reasoning—Faver, a middleware framework that streamlines RTL verification in LLM-based workflows. By combining LLM-friendly code structures with rule-based templates, Faver decouples circuit verification details, enabling the LLM to focus on functionality itself. Experimental results on both SFT and open-source models show that Faver can improve RTL generation accuracy by up to 14%.
Ensemble learning has been widely adopted to enhance the generalization performance of artificial intelligence models. However, the additional memory overhead and computational costs induced by ensemble models limit their deployment in resource-constrained scenarios. To address this challenge, we propose LEMON, a large language model (LLM) driven framework for ensemble reduction and Verilog HDL generation, designed to enable efficient hardware implementation. By leveraging the advanced capabilities of LLMs in semantic understanding, code generation, and complex problem solving, LEMON significantly reduces the memory footprint and computational demands of ensemble models. Extensive evaluations on 20 public datasets demonstrate that while maintaining a compact model size, LEMON achieves comparable or superior prediction accuracy to the original ensemble methods across the majority of test sets. Notably, compared with state-of-the-art techniques, LEMON achieves speedups exceeding 9 times and 338 times in ensemble reduction, showcasing exceptional scalability across diverse scenarios. Furthermore, LEMON integrates an LLM-powered automated hardware deployment pipeline that significantly simplifies the transition from high-level software models to optimized FPGA implementations. Compared with conventional random forest hardware implementations, LEMON reduces power consumption by over 64.7% and hardware resource utilization increased by more than 92.6%, making it highly suitable for edge computing and embedded system applications.
Large language models (LLMs) are increasingly being incorporated into chip and FPGA development workflows. However, existing studies have mainly focused on RTL generation, assertion generation, or localized debugging, while systematic case studies covering end-to-end engineering workflows remain limited. Using SegTracer, a data-path latency probing tool for multi-FPGA prototyping systems, as a case study, we investigate the role of a human-led, LLM-assisted, verification-driven development paradigm in practical IP implementation. During the design and deployment of SegTracer, LLMs are used for requirements refinement, solution comparison, design document consolidation, SpinalHDL implementation, generation of register-access and verification scripts, and troubleshooting during localized debugging; human developers are responsible for key mechanism definition, literature verification, clock-domain-crossing and handshake-semantics handling, waveform analysis, and experimental interpretation. Technically, SegTracer employs independent probe packets together with a packet generation and monitoring module and a unified timestamp module. Through stack-based recording on the forward path, in-place subtraction on the return path, and centralized endpoint statistics, it achieves segmented latency measurement and distribution analysis without relying on global clock synchronization. Simulation and FPGA board-level experiments show that the tool can support latency decomposition and tail-bottleneck localization in both QSFP direct-link and PCIe hardware-software collaborative scenarios. The case study demonstrates that large language models can significantly improve design iteration, script development, and joint debugging and verification efficiency, while human leadership remains essential in key aspects such as domain-specific semantic understanding, clock-domain-crossing handling, timing correctness, and result interpretation.
In recent years, the rapid advancement of deep learning and multimodal information fusion has significantly propelled the development of 2D virtual human driving technologies. Existing methods can generate highly realistic facial expressions and body movements for specified characters by leveraging diverse input data such as video, audio, facial cues, and poses. This remarkable progress offers substantial convenience for applications in virtual entertainment, online education, and intelligent interaction. This survey provides a systematic review of the current state and evolutionary trajectory of deep learning-based 2D virtual human driving methodologies. It thoroughly outlines the underlying principles and fundamental models commonly employed in this domain. The driving techniques are categorized and summarized according to both the target driving components (e.g., face, body) and the corresponding architectural frameworks, with a focused comparison and detailed elaboration of prominent methods developed in recent years. The research status and development prospects of holistic driving systems are also introduced. Furthermore, this survey delves into the critical technical challenges currently faced by the field, including generation realism, the quality of datasets, and the real-time performance of models. The aim of this survey is to offer a comprehensive technological overview and perspective for researchers in related fields, thereby contributing to the further advancement of this domain.
Multimodal emotion recognition aims to integrate data from multiple modalities for accurate inference of emotional states. Existing research either treats all modalities equally or adopts fixed fusion strategies based on a single modality, failing to adequately address the imbalance of modality contributions. To tackle this, we propose a dynamic multimodal fusion method MoMFE based on the mixture of experts (MoE) framework, incorporating an adaptive router module and a multimodal fusion expert module. This method dynamically evaluates the contribution of each modality and selects appropriate fusion strategies accordingly. The adaptive router mechanism dynamically assesses modality contributions through inter-modal correlation analysis and dynamic weighting to guide the selection of fusion experts, while an expert-guided loss function is integrated to further optimize the expert selection process. The multimodal fusion experts perform complementary fusion for different modality contribution combinations and introduce a shared expert to mitigate the loss of global information and reduce parameter redundancy. Comparative experiments on three benchmark datasets (MER2024, CMU-MOSEI, and CH-SIMS) for multimodal emotion recognition demonstrate that MoMFE method outperforms state-of-the-art (SOTA) multimodal fusion methods in core metrics, including binary emotion recognition accuracy (Acc-2) and F1 score. Notably, the method achieves an average improvement of approximately 2% on CH-SIMS dataset.
The core challenge in estimating individual causal effects (ICE) from observational data lies in confounding bias and the unobservability of counterfactual outcomes. To eliminate confounding bias, existing methods typically balance all covariates as potential confounders. However, erroneously balancing instrumental variables and adjustment variables introduce estimation bias and increase variance. Although recent studies address this via covariate decomposition to separate confounders, their sample re-weighting approach for balance can lead to extreme weights while neglecting the sample imbalance between treatment and control groups, resulting in inaccurate ICE estimates. To address these limitations, we propose an individual causal effect estimation method based on adversarial learning, which optimizes representation learning and distribution balancing via an adversarial mechanism. Specifically: 1) Sample processing. The SMOTE algorithm preprocesses data to mitigate sample imbalance between treatment and control groups. 2) Representation learning. Two adversarial probability constraint modules are designed to achieve confounder separation via variable disentanglement. 3) Confounder balancing. A feature balancer and an adversarial discriminator dynamically align the spatial distributions of confounders across treatment and control groups, constructing an adversarial balancing network to avoid the extreme weight problem inherent in sample re-weighting. Experimental results demonstrate that our algorithm can precisely decompose confounders and significantly improve the accuracy of ICE estimation.
With the widespread application of lightweight models in resource-constrained scenarios such as edge computing, their insufficient adversarial robustness has become increasingly prominent. While adversarial distillation serves as a primary approach to enhance lightweight models’ robustness, existing methods predominantly employ KL (Kullback-Leibler) divergence-based rigid constraints for robust knowledge transfer, which still suffer from flawed knowledge modeling and limited transfer efficiency. To address these issues, we propose an adversarial ranking distillation (ARD) method that significantly improves lightweight models’ robustness through a priority ranking constraint mechanism and a multi-level consistency framework. Specifically, the priority ranking constraint mechanism organizes the output elements of vision models on adversarial samples by importance ranking, enforces teacher-student ranking consistency, and implements differentiable loose coupling constraints by approximating discontinuous ranking processes through hyperbolic tangent functions. Building upon this, the multi-level ranking consistency distillation framework models transfer robustness knowledge from three perspectives: categorical semantic correlation, sample semantic correlation, and adversarial discrepancy correlation, enabling multi-perspective transmission of teachers’ adversarial defense capabilities while synergistically improving lightweight models’ clean sample accuracy and adversarial robustness. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets validate the effectiveness of our method, demonstrating substantial performance advantages over state-of-the-art adversarial distillation approaches. Furthermore, the proposed ARD exhibits enhanced adaptability and generalization capabilities under scenarios of limited training data and black-box attacks, providing an efficient security solution for lightweight model deployment in edge computing environments.
Spatial-temporal trajectory mining plays a crucial role in location-based services. Trajectory representation learning transforms complex trajectory features into low-dimensional representation vectors, enabling various downstream tasks. This process requires effectively extracting spatial-temporal features and motion patterns from trajectories. However, trajectory representation learning faces numerous challenges due to limitations in model capacity and issues with the scale of trajectory datasets. Large language model (LLM), pre-trained on massive high-quality datasets, have demonstrated remarkable capabilities in multi-task applications. Given the similarities between trajectories and natural language, it is feasible to develop an efficient trajectory representation method leveraging LLM. Nevertheless, vanilla LLMs cannot directly handle the unique spatial-temporal characteristics of trajectories or effectively extract motion patterns. To address these issues, we propose a trajectory representation learning model, MPTraj, which leverages LLMs to extract motion patterns. MPTraj introduces a novel trajectory semantic embedder that enables LLMs to process spatial-temporal features and extract movement patterns effectively. Additionally, a trajectory encoder with prompt tuning is integrated to generate high-quality trajectory representations. Furthermore, two pre-training tasks—next trajectory point prediction and movement pattern classification—are designed to enhance the model’s performance on downstream tasks. Experimental results on two real-world datasets and four representative tasks demonstrate the effectiveness of MPTraj.
Neuro-oriented speaker extraction is an intelligent speech processing technique that emulates human auditory attention mechanisms, enabling target speaker extraction from mixed speech by decoding the attentional focus embedded in listeners’ EEG (electroencephalogram) signals. This technology offers novel insights into resolving the “cocktail party problem” and holds significant implications for developing intelligent hearing-assistive devices. However, existing approaches confront three critical challenges: inefficient multimodal feature fusion, inadequate local speech feature extraction, and high computational complexity, which collectively constrain practical application efficacy. We propose GALENet, a novel end-to-end time-domain speaker extraction model, featuring three innovative modules: 1) A gated cross-attention fusion mechanism that establishes cross-modal correlations between EEG and speech through bidirectional cross-attention, while adaptively modulating modal contributions via dynamic gating weights to overcome the information interaction limitations of conventional linear fusion; 2) An E-DPRNN network incorporating Conv-U modules within the dual-path architecture to reinforce local short-term feature capture, synergized with residual connections to enhance gradient propagation efficiency; 3) A single-path global modulation (SPGM) block that replaces traditional inter-block modeling with parameter-free pooling and feature modulation, preserves global semantic modeling capability while reducing parameters. These modules collectively achieve efficient multimodal fusion and precise speaker extraction. Experimental validation demonstrates that the proposed method attains average SI-SDR scores of 13.59 dB, 10.44 dB, and 10.17 dB on the Cocktail Party, AVED, and MM-AAD datasets respectively, outperforming the baseline model MSFNet by 5.4%−8.2% while reducing parameters by 49.7%−55.6%. These results substantiate the model’s application potential and effectiveness in neuro-oriented speaker extraction.
With the continuous development of big data technology, high-dimensional datasets have been increasingly widely applied in various fields. These datasets typically feature high dimensionality and excessive redundancy, which tend to induce model overfitting and increase computational costs. Feature selection helps extracting an informative subset from them, reducing dimensionality and enhancing model interpretability. Although deep neural networks have shown potential in feature extraction, their performance is vulnerable to noise and outliers. To address these challenges, a novel feature selection algorithm based on the intuitionistic fuzzy deep autoencoder (IFDAE) is proposed. First, the algorithm fuses membership and non-membership degree information and applies intuitionistic fuzzy weights to handle uncertainty, thereby suppressing the impact of noise and outliers. Then, it fully learns the highly non-linear latent representations in the deep autoencoder, uses weight transfer to migrate pre-trained knowledge into the target network, and imposes a regularizer on the connection weight matrix between the input layer and the first hidden layer via the structured sparse norm to obtain a structurally sparse weight matrix, thus deriving the weight of each feature. Finally, to verify the effectiveness of the proposed IFDAE algorithm, experiments are conducted on twelve public datasets and one real-world schizophrenia dataset. The results demonstrate that the proposed algorithm exhibits significant superiority in both reconstruction capability and classification performance.
Keyframe selection is an important method for long-form video question answering, as it accurately identifies key content from redundant information and establishes an interpretable reasoning path. However, existing keyframe selection methods face challenges with insufficient semantic sensitivity during end-to-end training, leading to the introduction of a significant amount of irrelevant frame noise, which affects the model’s accuracy and interpretability. To address this, we propose a pseudo-labels guided keyframe selection (PGKS) model. This model first utilizes large language models (LLMs) to semantically integrate questions and answers, generating global descriptive text. It then employs a multimodal alignment model to compute semantic similarity scores between the description and video sampled frames, thereby constructing frame-level pseudo-labels. By guiding the calculation of frame-level similarity scores with these pseudo-labels, the model effectively suppresses irrelevant frame noise, enhancing both the accuracy and interpretability of keyframe selection. The decoupling between the utilization of LLMs and the training of the backbone network is architecturally designed to provide a cost-effective solution for leveraging LLMs in resource-constrained scenarios. Furthermore, our method balances differentiable and non-differentiable sorting results, and introduces a sliding time window mechanism to further improve the model’s understanding of temporal relationships in videos. Experimental results demonstrate that PGKS model achieves an accuracy of 62.67% on the long-form video question answering dataset NExT-QA, outperforming existing methods of comparable size and improving by 8.51% over the baseline without pseudo-label guidance.
With the rapid advancement of artificial intelligence technologies, large language models (LLMs) have emerged as foundational components in modern intelligent systems. However, the ever-increasing model size—ranging from billions to hundreds of billions of parameters—poses significant challenges in terms of memory footprint and computational overhead, particularly on latency-sensitive and resource-constrained platforms. While low-bit integer quantization (e.g., INT8) has demonstrated effectiveness in reducing memory consumption and improving inference efficiency, it suffers from notable quantization limitations when applied to large-scale models, including quantization-induced accuracy degradation and inefficient storage utilization. To address these issues, we propose a numerical quantization framework, termed low bit sharing (LBS). Built upon conventional integer quantization, LBS provides a structured high-low bit decomposition scheme that preserves only the most significant bits for each parameter and shares the less significant bits within a tensor group. The shared solution can reduce storage requirements while upholding the numerical representational capability. Furthermore, to mitigate accuracy degradation caused by the quantization of high-impact weights, we develop a salient value aware quantization strategy. By employing a Top-K selection algorithm, we isolate the most influential weights—typically residing in the tail of the parameter distribution—and assign them dedicated scaling factors. This targeted treatment effectively suppresses error accumulation during quantization and improves robustness in downstream performance. Experimental results across several state-of-the-art LLMs show that LBS achieves substantial storage reduction while preserving high model accuracy, making it a practical and scalable solution for deployment of large-scale models.
In recent years, edge-cloud-end collaboration (ECEC) has emerged as an efficient and promising technology for empowering various computation-intensive applications, effectively bridging the gap between data analytics and physical states. Within ECEC, there is an urgent need for a reliable and optimal task scheduling scheme to maximize resource utilization and provide satisfactory services to end users. However, existing scheduling schemes still face numerous challenges, such as the instability and complexity of network topologies, the highly dynamic nature of task priorities, and the lack of trust mechanisms among end users. We propose a blockchain-assisted dynamic priority scheduling strategy capable of transmitting large-scale tasks to edge or cloud servers in dynamic scenarios. We first model the resource allocation and task scheduling problem, establish a dynamic priority assignment model, and employ the analytic hierarchy process to determine the priority of sub-tasks generated from large-scale tasks. Subsequently, we adopt a blockchain-based Lagrangian multiplier distributed island genetic algorithm, which transforms the formulated problem into a convex optimization form to achieve optimal resource allocation under specific schemes. Furthermore, the proposed architecture incorporates a blockchain verification mechanism to enhance system stability and strengthen data privacy protection. Extensive simulation results demonstrate that, compared with several baseline schemes, the proposed solution achieves a 10% reduction in both total system delay and privacy overhead in ECEC.
ABR (adaptive bitrate) algorithms play a crucial role in enhancing the QoE (quality of experience) for online video streaming. Existing ABR algorithms rely on network characteristics observed at the application layer for bitrate decisions. However, this approach has inherent limitations: accurate video chunk download times cannot be fully derived from application layer observations. Specifically, these algorithms overlook critical factors such as RTT (round-trip time) and packet loss rate, which impact video chunk transmission, and they exhibit limited responsiveness to rapid network fluctuations. To address this, we propose Prophet, a adaptive bitrate algorithm based on transport layer information. Unlike traditional ABR algorithms, Prophet directly calculates network parameters such as bandwidth, packet loss rate, and RTT at the transport layer, enabling a more accurate assessment of network conditions. Additionally, a download time prediction model is developed based on transport-layer feedback, taking factors into account, like packet loss retransmission and tail latency to achieve precise download time predictions. Experiments conducted in real-world network environments demonstrate that the Prophet algorithm performs well under various network conditions. Compared with existing ABR algorithms, Prophet achieves an average QoE improvement of 0.3%−117.9%, with a notable average QoE increase of 31.7%−117.9% in cellular network conditions.
An efficient and robust Byzantine fault tolerance (BFT) consensus protocol serves as the cornerstone of consortium blockchain applications. In existing proposal processes, most protocols require the primary node to transmit O(n2) signatures, which significantly impairs protocol efficiency. Some protocols achieve linear view change, where the primary node only needs to send O(n) signatures to verify the validity of proposals, but at the cost of sacrificing optimistic responsiveness or introducing an extra round of message exchange. To address these inherent trade-offs, we propose Climber, a novel BFT consensus protocol that integrates linear view change, optimistic responsiveness, and two-phase commit into a unified framework. Specifically, during view change, the newly-elected primary node enumerates all unlocked commitments generated in the previous view to prove proposal validity without adding extra commit phases, ensuring the protocol’s operational efficiency. Furthermore, the new primary node leverages aggregate signature technology to aggregate unlocked commitments from the same view into an unlocked certificate, reducing the communication complexity of view change to O(n). Additionally, a blacklist mechanism for malicious nodes is introduced, enabling the primary node to verify signatures with a time complexity of O(n). Simulation experiments are conducted under different network scales and malicious node ratios, and the results show that Climber not only maintains competitive throughput and latency performance but also exhibits superior robustness compared with Fast-HotStuff, making it more suitable for large-scale consortium blockchain scenarios.
The defense methods against adversarial perturbations holds significant importance for the secure application of recognition models. Existing adversarial defense methods typically either leverage fixed adversarial samples to augment classifier training or perform sample reconstruction prior to model input. However, in practical defense scenarios, such approaches suffer from inherent limitations, including inadequate generalizability, loss of critical sample components, and unstable data distribution. To tackle this problem, we propose probability distribution bidirectional guidance for adversarial purification with diffusion model. Specifically, we propose a novel bidirectional guidance strategy encompassing inter-class loops and intra-class aggregation. For the inter-class loop component, cross-domain mapping between adversarial and clean domains is achieved by leveraging the inverse time sequence of stochastic differential equations (SDEs), with targeted interference elimination enabled via the circular consistency principle. Subsequently, latent variable refinement is employed for distribution regularization, enhancing the integrity of noise-free components and the stability of distribution characteristics. For the intra-class aggregation component, it is exploited to strengthen the confidence of category attributes at each time step along the path from the clean domain to the adversarial domain, thereby facilitating the shift toward high-density intervals of the probability distribution. Extensive experiments conducted on benchmark datasets including CIFAR-10, CIFAR-100, and ImageNet demonstrate that our proposed method achieves state-of-the-art (SOTA) performance. Specifically, it yields a 10.9 percentage point (pp) improvement in Standard Acc and a 4.36 pp gain in Robust Acc compared with existing approaches. Furthermore, our method exhibits remarkable advantages in preserving the integrity of noise-free components and demonstrates robust stability throughout the purification process.