Latest ArticlesIn 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.
This 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.