We introduce FinMMDocR, a novel bilingual multimodal benchmark for evaluating multimodal large language models (MLLMs) on real-world financial numerical reasoning. Compared to existing benchmarks, our work delivers three major advancements. (1) Scenario Awareness: 57.9% of 1,200 expert-annotated problems incorporate 12 types of implicit financial scenarios (e.g., Portfolio Management), challenging models to perform expert-level reasoning based on assumptions; (2) Document Understanding: 837 Chinese/English documents spanning 9 types (e.g., Company Research) average 50.8 pages with rich visual elements, significantly surpassing existing benchmarks in both breadth and depth of financial documents; (3) Multi-Step Computation: Problems demand 11-step reasoning on average (5.3 extraction + 5.7 calculation steps), with 65.0% requiring cross-page evidence (2.4 pages average). The best-performing MLLM achieves only 58.0% accuracy, and different retrieval-augmented generation (RAG) methods show significant performance variations on this task. We expect FinMMDocR to drive improvements in MLLMs and reasoning-enhanced methods on complex multimodal reasoning tasks in real-world scenarios.

Junpeng Ding, Zichen Tang, Haihong E, Mengyuan Ji, Yang Liu, Haolin Tian, Haiyang Sun, Pengqi Sun, Yang Xu, Yichen Liu, Haocheng Gao, Zijie Xi, Ruomeng Jiang, Peizhi Zhao, Rongjin Li, Yuanze Li, Jiacheng Liu, Zhongjun Yang, Jintong Chen, Siying Lin. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Bo Zhang, Tzu-Yen Ma, Zichen Tang, Junpeng Ding, Zirui Wang, Yizhuo Zhao, Peilin Gao, Zijie Xi, Zixin Ding, Haiyang Sun, Haocheng Gao, Yuan Liu, Liangjia Wang, Yiling Huang, Yujie Wang, Yuyue Zhang, Ronghui Xi, Yuanze Li, Jiacheng Liu, Zhongjun Yang, Haihong E. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

For polar codes, perturbing received information can enhance the successive cancellation (SC) decoding performance. This is an effective approach for realizing low-latency yet high performance for long polar codes, since all the perturbation-enhanced SC (PSC) decoding can be performed in parallel. This paper provides theoretical insights into soft information perturbation, revealing that the PSC decoding can be equivalently interpreted as perturbing thea posteriorilog-likelihood ratios (LLRs) of information bits. Such a revelation leads to the design of an improved PSC (IPSC) decoding that yields a lower perturbation complexity. By better utilizing the decodinga posterioriLLRs, a set of possibly erroneous estimations can be formed and further perturbed, resulting in the proposed hybrid PSC (HPSC) decoding. During each new SC decoding attempt, it takes turns to flip the first erroneous bit by introducing a biased perturbation, while the subsequent erroneous estimations are corrected through random perturbations. Our simulation results validate that, for various codeword lengths and rates, the proposed IPSC decoding can achieve a similar performance as the conventional PSC decoding, but yield a significantly reduced perturbation complexity. With the same number of decoding attempts, the proposed HPSC decoding outperforms several state-of-the-art SC-based decoding, such as the thresholded SC-flip (TSCF) decoding and the dynamic SC-flip (DSCF) decoding.

Zichen Tang, Haihong E, Ziyan Ma, Haoyang He, Jiacheng Liu, Zhongjun Yang, Zihua Rong, Rongjin Li, Kun Ji, Qing Huang, Xinyang Hu, Yang Liu, Qianhe Zheng. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

For polar codes, the bit-flipping strategy can significantly improve performance of its successive cancellation (SC) decoding. However, the gain derived from SC-flip (SCF) decoding diminishes as the codeword length increases. Addressing this issue, this paper proposes a novel hybrid perturbation-based SC (HPSC) decoding. If the initial SC decoding fails, the algorithm will generate multiple SC decoding attempts, each of which introduces stochastic perturbations to the received symbols. By soft information perturbations, the SC decoding can divert from the initial erroneous estimation and converge to the intended one. Our simulation results show that the proposed HPSC decoding consistently yields stable coding gains over various codeword lengths and rates. With the same number of decoding attempts, the HPSC decoding outperforms the thresholded SCF (TSCF) decoding. Moreover, it can achieve a similar performance as the cyclic redundancy check (CRC) aided SC list (CA-SCL) decoding, without any path sorting and expansion requirements.

For polar codes, perturbing received information can enhance the error-correction performance of successive cancellation (SC) decoding. This is an effective approach for realizing low-latency yet high decoding performance for long polar codes, since each perturbation-enhanced SC (PSC) decoding can be performed in parallel. This paper provides theoretical insights into the soft information perturbation. It first reveals that the PSC decoding can be equivalently viewed as perturbing the SC decoding a posteriori log-likelihood ratio (LLR) of the information bits. Such a revelation enables us to reduce the perturbation complexity by only targeting the information bits, resulting in an improved PSC (IPSC) decoding. By better utilizing the a posteriori LLRs, a set of possibly erroneous estimations can be formed to be perturbed, further reducing the perturbation complexity. Our simulation results show that, for various codeword lengths, the proposed IPSC decoding can achieve a similar performance as the PSC decoding, while yielding significant perturbation complexity reduction.

MOTIVATION: With the rapid advancements in Cryo-Electron Microscopy (Cryo-EM), an increasing number of high-resolution 3D density maps are being made publicly available, highlighting the urgent need for efficient structure similarity retrieval. Exploring map similarity at various levels is critical for fully utilizing these valuable resources. Our previously proposed CryoAlign can provide more accurate density map alignment while maintaining a low failure rate. However, CryoAlign only offers a method for aligning density maps, with low efficiency in local alignment, and has not yet been applied to the retrieval of Cryo-EM density maps. RESULTS: We have developed an alignment-based retrieval tool to perform both global and local retrieval. Our approach adopts parallel-accelerated CryoAlign for high-precision 3D alignment and transforms density maps into point clouds for efficient retrieval and storage. Additionally, a multi-dimension scoring function is introduced to accurately assess structural similarities between superimposed density maps. To demonstrate its applicability, we conducted thorough testing across different retrieval tasks, such as global, local or hybrid similarity retrieval. Our tool achieves up to a 7-fold speedup while supporting precise local alignments. Comprehensive experiments demonstrate that even when one density map is entirely contained within another, our tool performs exceptionally well in high-resolution density map retrieval. It provides researchers with an efficient and accurate solution for density map similarity search. AVAILABILITY AND IMPLEMENTATION: The source code, documentation, and sample data can be downloaded at https://github.com/JokerL2/CryoAlign2.

Abstract With the rapid spread of the Internet of Vehicles (IoV) technology, vehicle network security is facing increasingly severe challenges. Intrusion detection technology has become a crucial tool for ensuring the information security of IoV. Since the traffic data of the IoV is large and has spatio-temporal characteristics, most previous studies are based on a single deep learning method to extract temporal or spatial features, which does not fully extract features of IoV data. To address the above issues, a spatio-temporal feature extraction model with feature selection is proposed. First, to solve the problem of long detection time with huge data traffic, a new feature selection method is proposed to screen the optimal feature subset by combining the correlation-based feature selection method with the crayfish optimization algorithm (CFS-COA). Second, the selected optimal features are used in a spatio-temporal feature extraction model that combines a Temporal Convolutional Network and a Bidirectional Gated Recurrent Unit (TCN-BiGRU) for classification. Finally, the performance of the model is evaluated using two types of datasets: the NSL-KDD and UNSW-NB15 datasets for external communications, and the Car-Hacking dataset for in-vehicle networks. The experimental results indicate that the proposed model demonstrates high classification performance and lightweight characteristics, achieving 100% accuracy on the Car-Hacking dataset.

Due to the high parallelism of belief propagation (BP) decoding, it is considered as a promising solution for the decoding latency challenge of long polar codes. However, the error-correction performance of the classical BP decoding is inferior to that of the successive cancellation (SC) and the SC list (SCL) decoding. In this paper, an adaptive BP (ABP) decoding algorithm is proposed to bridge this performance discrepancy. It iteratively adjusts the a priori log-likelihood ratios (LLRs) of error-prone bits, which can be efficiently detected using the frozen and information processing elements (FIPEs). Moreover, a novel low-complexity FIPE-based early termination criterion (ETC) is proposed to further reduce the decoding complexity. It functions when all the frozen bits in the FIPEs are successfully decoded with stable LLR magnitudes. Our numerical results show that for the (1024,512) polar code, the ABP decoding outperforms the classical BP decoding by 0.3 dB at the frame error rate (FER) of 10–4over the additive white Gaussian noise (AWGN) channel. It can also achieve up to 78.5% latency reduction over the fast simplified SC (FSSC) decoding, while maintaining the same performance. The proposed ETC also exhibits a lower hardware complexity over the existing G-matrix criterion.

Abstract Intrusion detection has become a research focus in internet information security, with deep learning algorithms playing a crucial role in its development. Typically, intrusion detection data are transformed into a two-dimensional matrix by segmenting, stacking and padding them with zeros for input into deep learning models. However, this method consumes computational resources and fails to consider the correlation between features. In this paper, we transform the data into images through visualization operations and propose an information entropy weighted scheme to optimize the collision element problem during the transformation process. This method enhances the correlation between pixel frame features, leading to approximately 2% improvement in accuracy of the classification model when using the generated image samples for detection in experiments. To address the issues of insensitivity to target feature locations and incomplete feature extraction in traditional neural networks, this paper introduces a new network model called CBAM-CapsNet, which combines the advantages of the lightweight Convolutional Block Attention Module and capsule networks. Experimental results on the UNSW-NB15 and IDS-2017 datasets demonstrate that the proposed model achieves accuracies of 92.94% and 99.72%, respectively. The F1 scores obtained are 91.83% and 99.56%, indicating a high level of detection.

Summary Network intrusion detection system (NIDS) is a key component to identify abnormal behavior of network systems and plays an important role in preventing the occurrence of network attacks. Although a considerable number of machine learning methods have been applied in the field of intrusion detection, it is still a challenge for existing solutions to achieve a good classification performance. The existing traffic datasets generally have redundant and irrelevant features, which hinder classifiers from making more accurate predictions. Furthermore, a single classifier has limited classification performance and may not be able to achieve a better detection performance overall in the face of unbalanced multi‐category traffic data. Therefore, in order to improve the classification performance of intrusion detection models, this paper proposes an adaptive ensemble model by combining feature selection techniques and effective ensemble methods. Firstly, a heuristic feature selection algorithm (NRS‐SSA) is proposed by introducing the neighborhood dependency degree of the neighborhood rough set (NRS) into the salp swarm algorithm (SSA). Then, an improved adaptive weighted voting algorithm is designed. The SSA is introduced to optimize the weight matrix when setting the voting weight. Finally, we use the designed voting algorithm to combine the classification advantages of homogeneous classifiers and heterogeneous classifiers, respectively, and propose an M‐Tree algorithm and an adaptive ensemble model. The experimental results on multiple intrusion detection datasets show that the proposed adaptive ensemble model achieves an advanced detection level.

coauthors Rajiv Kumar has no contribution and was added incorrectly. e correct author list is shown above.

Offloading to fog servers makes it possible to process heavy computational load tasks in local devices. However, since the generation problem of offloading decisions is an N-P problem, it cannot be solved optimally or traditionally, especially in multitask offloading scenarios. Hence, this paper has proposed a randomization-based dynamic programming offloading algorithm, based on genetic optimization theory, to solve the offloading decision generation problem in mobile fog computing. The algorithm innovatively designs a dynamic programming table-filling approach, i.e., iteratively generates a set of randomized offloading decisions. If some in these sets improve the decisions in the DP table, then they will be merged into the table. The iterated DP table is also used to improve the set of decisions generated in the iteration to obtain the optimal offloading approximate solution. Extensive simulations show that the proposed DPOA can generate decisions within 3 ms and the benefit is especially significant when users are in multitask offloading scenarios.