Traditional quantitative investment research is encountering diminishing returns alongside rising labor and time costs. To overcome these challenges, we introduce the large investment model (LIM), a novel research paradigm designed to enhance both performance and efficiency at scale. LIM employs end-to-end learning and universal modeling to create an upstream foundation model, which is capable of autonomously learning comprehensive signal patterns from diverse financial data spanning multiple exchanges, instruments, and frequencies. These “global patterns” are subsequently transferred to downstream strategy modeling, optimizing performance for specific tasks. We detail the system architecture design of LIM, address the technical challenges inherent in this approach, and outline potential directions for future research.

Financial large language models (FinLLMs) offer immense potential for financial applications. While excessive deployment expenditures and considerable inference latency constitute major obstacles, as a prominent compression methodology, knowledge distillation (KD) offers an effective solution to these difficulties. A compre-hensive survey is conducted in this work on how KD interacts with FinLLMs, covering three core aspects: strategy, application, and evaluation. At the strategy level, this review introduces a structured taxonomy to comparatively analyze existing distillation pathways. At the application level, this review puts forward a logical upstream-midstream-downstream framework to systematically explain the practical value of distilled models in the financial field. At the evaluation level, to tackle the absence of standards in the financial field, this review constructs a comprehensive evaluation framework that proceeds from multiple dimensions such as financial accuracy, reasoning fidelity, and robustness. In summary, this research aims to provide a clear roadmap for this interdisciplinary field, to accelerate the development of distilled FinLLMs.

Alpha mining, which refers to the systematic discovery of data-driven signals predictive of future cross-sectional returns, is a central task in quantitative research. Recent progress in large language models (LLMs) has sparked interest in LLM-based alpha mining frameworks, which offer a promising middle ground between human-guided and fully automated alpha mining approaches and deliver both speed and semantic depth. This study presents a structured review of emerging LLM-based alpha mining systems from an agentic perspective, and analyzes the functional roles of LLMs, ranging from miners and evaluators to interactive assistants. Despite early progress, key challenges remain, including simplified performance evaluation, limited numerical understanding, lack of diversity and originality, weak exploration dynamics, temporal data leakage, and black-box risks and compliance challenges. Accordingly, we outline future directions, including improving reasoning alignment, expanding to new data modalities, rethinking evaluation protocols, and integrating LLMs into more general-purpose quantitative systems. Our analysis suggests that LLM is a scalable interface for amplifying both domain expertise and algorithmic rigor, as it amplifies domain expertise by transforming qualitative hypotheses into testable factors and enhances algorithmic rigor for rapid backtesting and semantic reasoning. The result is a complementary paradigm, where intuition, automation, and language-based reasoning converge to redefine the future of quantitative research.

Current financial large language models (FinLLMs) exhibit two major limitations: the absence of standardized evaluation metrics for stock analysis quality and insufficient analytical depth. We address these limitations with two contributions. First, we introduce AnalyScore, a systematic framework for evaluating the quality of stock analysis. Second, we construct Stocksis, an expert-curated dataset designed to enhance the financial analysis capabilities of large language models (LLMs). Building on Stocksis, together with a novel integration framework and quantitative tools, we develop FinSphere, an artificial intelligence (AI) agent that generates professional-grade stock analysis reports. Evaluations with AnalyScore show that FinSphere consistently surpasses general-purpose LLMs, domain-specific FinLLMs, and existing agent-based systems, even when the latter are enhanced with real-time data access and few-shot guidance. The findings highlight FinSphere’s significant advantages in analytical quality and real-world applicability.

The development of large language models (LLMs) has created transformative opportunities for the financial industry, especially in the area of financial trading. However, how to integrate LLMs with trading systems has become a challenge. To address this problem, we propose an intelligent trade order recognition pipeline that enables the conversion of trade orders into a standard format for trade execution. The system improves the ability of human traders to interact with trading platforms while addressing the problem of misinformation acquisition in trade execution. In addition, we create a trade order dataset of 500 pieces of data to simulate the real-world trading scenarios. Moreover, we design several metrics to provide a comprehensive assessment of dataset reliability and the generative power of big models in finance by using five state-of-the-art LLMs on our dataset. The results show that most models generate syntactically valid JavaScript object notation (JSON) at high rates (about 80%-99%) and initiate clarifying questions in nearly all incomplete cases (about 90%-100%). However, end-to-end accuracy remains low (about 6%-14%), and missing information is substantial (about 12%-66%). Models also tend to over-interrogate—roughly 70%-80% of follow-ups are unnecessary—raising interaction costs and potential information-exposure risk. The research also demonstrates the feasibility of integrating our pipeline with the real-world trading systems, paving the way for practical deployment of LLM-based trade automation solutions.

Narrative economics suggests that financial markets are strongly influenced by evolving narratives, creating opportunities for forecasting emerging events and their economic impacts. However, existing large language model (LLM)-based approaches are inadequate in terms of systematic task decomposition and alignment with financial applications. We propose MENTOR, a multi-agent framework for event and narrative trend prediction that integrates teacher-student iterative reasoning with progressive subtasks: detecting and ranking trending events, forecasting future events from current narratives, and predicting industry index performance influenced by these events. Experiments on our self-constructed Chinese key opinion leader (KOL) articles dataset and English financial news dataset show that MENTOR consistently outperforms recent baselines such as the stakeholder-enhanced future event prediction (StkFEP) and summarize-explain-predict (SEP) frameworks in both event prediction and industry ranking tasks. In addition, the backtest results at the portfolio level show that improved event and industry forecasts can bring about a practical improvement in investment performance. These results demonstrate that incorporating structured reasoning and multi-agent feedback enables more reliable event forecasting and strengthens the connection between narrative dynamics and financial market outcomes.

Transformers have demonstrated considerable success across various domains but are constrained by their significant computational and memory requirements. This poses challenges for deployment on resource-constrained devices. Quantization, as an effective model compression method, can significantly reduce the operational time of Transformers on edge devices. Notably, Transformers display more substantial outliers than convolutional neural networks, leading to uneven feature distribution among different channels and tokens. To address this issue, we propose an adaptive outlier correction quantization (AOCQ) method for Transformers, which significantly alleviates the adverse effects of these outliers. AOCQ adjusts the notable discrepancies in channels and tokens across three levels: operator level, framework level, and loss level. We introduce a new operator that equivalently balances the activations across different channels and insert an extra stage to optimize the activation quantization step on the framework level. Additionally, we transfer the imbalanced activations across tokens and channels to the optimization of model weights on the loss level. Based on the theoretical study, our method can reduce the quantization error. The effectiveness of the proposed method is verified on various benchmark models and tasks. Surprisingly, DeiT-Base with 8-bit post-training quantization (PTQ) can achieve 81.57% accuracy with a 0.28 percentage point drop while enjoying 4×faster runtime. Furthermore, the weights of Swin and DeiT on several tasks, including classification and object detection, can be post-quantized to ultra-low 4 bits, with a minimal accuracy loss of 2%, while requiring nearly 8×less memory.

Predicting future heart rate (HR) not only helps in detecting abnormal heart rhythms but also provides timely support for downstream health monitoring services. Existing methods for HR prediction encounter challenges, especially concerning privacy protection and data heterogeneity. To address these challenges, this paper proposes a novel HR prediction framework, PCFedH, which leverages personalized federated learning and prototypical contrastive learning to achieve stable clustering results and more accurate predictions. PCFedH contains two core modules: a prototypical contrastive learning-based federated clustering module, which characterizes data heterogeneity and enhances HR representation to facilitate more effective clustering, and a two-phase soft clustered federated learning module, which enables personalized performance improvements for each local model based on stable clustering results. Experimental results on two real-world datasets demonstrate the superiority of our approach over state-of-the-art methods, achieving an average reduction of 3.1% in the mean squared error across both datasets. Additionally, we conduct comprehensive experiments to empirically validate the effectiveness of the key components in the proposed method. Among these, the personalization component is identified as the most crucial aspect of our design, indicating its substantial impact on overall performance.

Under dynamic conditions, the smearing effect of star spots on the image plane reduces centroid extraction accuracy, which has an impact on attitude estimation. To enhance the dynamic performance of the star sensor, we propose a multiplication extended Kalman filter (MEKF)-aided non-blind star image restoration algorithm based on the heterogeneous blur kernel. The proposed algorithm consists of three procedures. First, the MEKF is used to estimate the attitude and gyro drift to eliminate the measurement error of the star sensor and gyro drift. Second, the attitude predicted by MEKF is used, which provides initial conditions and accelerates the subsequent algorithm. Finally, a gyro-assisted heterogeneous blur kernel estimation algorithm is presented for restoring non-uniform and nonlinear motion-blurred star images. In contrast to existing dynamic star image deblurring algorithms, which focus mostly on image content, the proposed method emphasizes the cause of motion blur by fusing MEKF and a heterogeneous blur kernel. This leads to significantly enhanced robustness against noise and improved restoration accuracy. Simulation results demonstrate that the proposed method significantly outperforms existing techniques, improving centroid extraction accuracy by up to 59.64% and pointing accuracy across all axes by more than 78.94%.

Static analysis presents significant challenges in alarm handling, where probabilistic models and alarm prioritization are essential methods for addressing these issues. These models prioritize alarms based on user feedback, thereby alleviating the burden on users to manually inspect alarms. However, they often encounter limitations related to efficiency and issues such as false generalization. While learning-based approaches have demonstrated promise, they typically incur high training costs and are constrained by the predefined structures of existing models. Moreover, the integration of large language models (LLMs) in static analysis has yet to reach its full potential, often resulting in lower accuracy rates in vulnerability identification. To tackle these challenges, we introduce BinLLM, a novel framework that harnesses the generalization capabilities of LLMs to enhance alarm probability models through rule learning. Our approach integrates LLM-derived abstract rules into the probabilistic model, using alarm paths and critical statements from static analysis. This integration enhances the model’s reasoning capabilities, improving its effectiveness in prioritizing genuine bugs while mitigating false generalizations. We evaluated BinLLM on a suite of C programs and observed 40.1% and 9.4% reduction in the number of checks required for alarm verification compared to two state-of-the-art baselines, Bingo and BayeSmith, respectively, underscoring the potential of combining LLMs with static analysis to improve alarm management.

A novel fuzzy sliding mode control (FSMC) strategy is proposed to enhance the robustness and stability of position control for underactuated quadrotor unmanned aerial vehicles (UAVs) in the presence of external disturbances and model uncertainties. To realize the adaptive ability and robustness of the system in complex dynamic environments, an intelligent two-dimensional fuzzy controller is designed based on traditional sliding mode control (SMC) to adjust SMC parameters in real time, thereby adapting to the variable structure parameters of the system. First, based on the designed filter variables regarding errors, traditional SMC is used to reduce tracking errors. Then, the fuzzy logic module (FLM) combined with SMC, i.e., the self-learning module (FLM+SMC), is developed based on the filter variables and their rate of change to adjust the two parameters of the above SMC. Subsequently, the output signals of the FLM are fed back into the SMC module, and then a closed-loop tuning system using FSMC is developed for the UAVs. Moreover, the stability of the FSMC is rigorously verified using the Lyapunov theory. Finally, comprehensive simulations demonstrate that the designed FSMC not only offers accurate trajectory precision but also has robustness and disturbance rejection, and comparative simulations using SMC and adaptive radial basis function neural network control (RBFNNC) are used to validate the result.

Fractional calculus is considered a useful tool for gaining deeper insights into systems with memory effects or history. Fractional-order modeling of nonlinear systems may increase the stiffness and complexity of the system, but also provides better insights. This study introduces a swarm intelligence-based parameter estimation of the fractional Hammerstein autoregressive exogenous noise (fractional-HARX) model. The Grünwald-Letnikov finite difference formula is used to develop the fractional-HARX model from the standard HARX model. This study presents the design of a swarm intelligence-based electric eel foraging optimization algorithm (EEFOA) for parameter estimation of the fractional-HARX model under multiple noise scenarios for second- and third-order polynomial type nonlinearity. The key-term separation principle is also incorporated in the system model to reduce the occurrence of redundant parameters due to cross-product terms in the information vector. The designed methodology is examined, and the superiority of EEFOA is endorsed in terms of convergence, robustness, stiff parameter estimation, and deviation from the mean point in comparison with state-of-the-art optimization heuristics such as the whale optimization algorithm, the African vulture optimization algorithm, Harris hawk’s optimizer, and the reptile search algorithm. The statistical significance of the EEFOA for the estimation of fractional-HARX models is also established using statistical indices of best, mean, and worst fitness values along with standard deviation for multiple noise scenarios.

Tibetan Jiu chess, recognized as a national intangible cultural heritage, is a complex game comprising two distinct phases: the layout phase and the battle phase. Improving the performance of deep reinforcement learning (DRL) models for Tibetan Jiu chess is challenging, especially given the constraints of hardware resources. To address this, we propose a two-stage model called JFA, which incorporates hierarchical neural networks and knowledge-guided techniques. The model includes sub-models: strategic layout model (SLM) for the layout phase and hierarchical battle model (HBM) for the battle phase. Both sub-models use similar network structures and employ parallel Monte Carlo tree search (MCTS) methods for independent self-play training. HBM is structured as a hierarchical neural network, with the upper network selecting movement and jump capturing actions and the lower network handling square capturing actions. Human knowledge-based auxiliary agents are introduced to assist SLM and HBM, simulating the entire game and providing reward signals based on square capturing or victory outcomes. Additionally, within the HBM, we propose two human knowledge-based pruning methods that prune parallel MCTS and capture actions in the lower network. In the experiments against a layout model using the AlphaZero method, SLM achieves a 74% win rate, with the decision-making time being reduced to approximately 1/147 of the time required by the AlphaZero model. SLM also won the first place at the 2024 China National Computer Game Tournament. HBM achieves a 70% win rate when playing against other Tibetan Jiu chess models. When used together, SLM and HBM in JFA achieve an 81% win rate, comparable to the level of a human amateur 4-dan player. These results demonstrate that JFA effectively enhances artificial intelligence (AI) performance in Tibetan Jiu chess.

Deep learning has empowered traffic prediction models to integrate diverse auxiliary data sources, such as weather and temporal features, for enhanced forecasting accuracy. However, existing approaches often suffer from limited generality and scalability, and the field lacks a unified benchmark for fair model comparison. This absence hinders consistent performance evaluation, slows the development of robust and adaptable models, and makes it challenging to quantify the incremental benefits of different auxiliary data sources. To address these issues, we present MltAuxTSPP, a unified benchmark framework for deep learning-based traffic state prediction with multi-source auxiliary data. The framework features a standardized data container and a fusion embedding module, enabling consistent utilization of heterogeneous data and improving scalability. It produces unified hidden representations that can be seamlessly adopted by various downstream models, ensuring fair and reproducible comparisons under identical conditions. Extensive experiments on real-world datasets demonstrate that MltAuxTSPP effectively leverages weather and temporal features to improve long-term forecast performance and offers a practical and reproducible foundation for advancing research in traffic state prediction.

Using the Global Positioning System (GPS) and the mobility of marine surface vehicles, this paper addresses the navigation problem between unmanned surface vehicles (USVs) and autonomous underwater vehicles (AUVs). We propose a moving AUV state estimation method based on the trajectory optimization of the USV. In particular, by exploring the Doppler effect on the frequency of arrival (FOA) of the acoustic signals received by a single-surface USV, the position and velocity of the AUV can be estimated simultaneously, offering a robust solution that eliminates the need for time synchronization. Moreover, the USV trajectory is dynamically adjusted to achieve optimal USV-AUV measurement geometry, thereby improving the AUV’s observability and enhancing state estimation performance. The innovation lies in a tailored cost function grounded in observability analysis via the Cramér-Rao lower bound (CRLB) and geometric constraints. It integrates (1) the CRLB to optimize system observability, thereby enhancing estimation accuracy, (2) a distance term to ensure that the USV maintains appropriate proximity to the AUV, and (3) a turning rate term that adjusts the USV’s orientation to improve following capability. The cost function is then minimized using a particle swarm optimization algorithm, balancing these components to achieve a robust AUV tracking framework. We conduct comprehensive simulations to examine the potential influences of different factors, including the complexity of the USV trajectory, AUV depth, measurement frequency, packet loss rate, and noise levels, on navigation performance. Simulation results demonstrate the effectiveness of the proposed method in estimating and tracking the AUV.

To address the problem of underwater multi-sensor multi-target passive tracking in clutter, a distributed kernel mean embedding-based Gaussian belief propagation (DKME-GaBP) algorithm is proposed. First, a joint posterior probability density function (PDF) is established and factorized, and it is represented by the corresponding factor graph. Then, the GaBP algorithm is executed on this factor graph to reduce the computational complexity of data association. The factor graph of the GaBP consists of inner and outer loops. The inner loop is responsible for local track estimation and data association. The outer loop fuses information from different sensors. For the inner loop, the kernel mean embedding (KME) with a Gaussian kernel is designed to transform the strong nonlinear problem of local estimation into a linear problem in a high-dimensional reproducing kernel Hilbert space (RKHS). For the outer loop, a multi-sensor distributed fusion method based on KME is proposed to improve fusion accuracy by accounting for the distance among different PDFs in RKHS. The effectiveness and robustness of the DKME-GaBP are validated in the simulations.

A metasurface-loaded 1×2 patch array antenna assisted by a deep-learning optimization method is proposed to realize port and radiation pattern decoupling simultaneously to enhance the isolation among elements in multi-input multi-output (MIMO) systems. The deep-learning-assisted optimization method uses an artificial neural network (ANN) and a particle swarm optimization (PSO) algorithm to seek the optimal structure of the antenna to achieve port decoupling with undistorted radiation patterns. The ANN is trained to describe the nonlinear relationship between the geometric parameters and the responses of the antenna. The PSO algorithm, guided by the cost function and number of iterations, is used to optimize the structure of the antenna according to the cost function combined with the trained ANN. Finally, by constraining the cost function, we obtain a 1×2 patch array antenna with a metasurface fixed above by studs, which achieves port and radiation pattern decoupling simultaneously. To validate the principle and design method, we designed, fabricated, and measured an antenna prototype with dimensions of 0.88λ₀×0.47λ₀×0.21λ₀ (λ₀ is the wavelength in free space at the center frequency). The measured fractional bandwidth is 8% (4.8-5.2 GHz). The isolation of the two-element patch antenna increases from 7.6 dB to 24.3 dB with an envelope correlation coefficient (ECC) of < 0.0005 at 0.35λ₀. Moreover, the H-plane radiation pattern of each element is consistent and symmetric in the broadside direction. These characteristics make the proposed antenna suitable for MIMO antenna systems with close spacing.

Monolithic three-dimensional integrated circuits (M3D ICs) have emerged as an innovative solution to overcome the limitations of traditional 2D scaling, offering improved performance, reduced power consumption, and enhanced functionality. Inter-layer vias (ILVs), crucial components of M3D ICs, provide vertical connectivity between layers but are susceptible to manufacturing and operational defects, such as stuck-at faults (SAFs), shorts, and opens, which can compromise system reliability. These challenges necessitate advanced built-in self-test (BIST) methodologies to ensure robust fault detection and localization while minimizing the testing overhead. In this paper, we introduce a novel BIST architecture tailored to efficiently detect ILV defects, particularly in irregularly positioned ILVs, and approximately localize them within clusters, using a walking pattern approach. In the proposed BIST framework, ILVs are grouped according to the probability of fault occurrence, enabling efficient detection of all SAFs and bridging faults (BFs) and most multiple faults within each cluster. This strategy empowers designers to fine-tune fault coverage, localization precision, and test duration to meet specific design requirements. The new BIST method addresses a critical shortcoming of existing solutions by significantly reducing the number of test configurations and overall test time using multiple ILV clusters. The method also enhances efficiency in terms of area and hardware utilization, particularly for larger circuit benchmarks. For instance, in the LU32PEENG benchmark, where ILVs are divided into 64 clusters, the power, area, and hardware overheads are minimized to 0.82%, 1.03%, and 1.14%, respectively.