There is a growing consensus that, in order to serve as models of human language processing, language models (LMs) need to be constrained in their use of memory for context, the analogue to human working memory (WM).Here we take a novel yet simple approach to constraining WM in language models, in a way that reflects models of human cognition where memory is treated as a limited resource and deployed strategically.In order to capture this constraint on memory encoding, we inject noise into the hidden representations of Transformerbased LMs at tunable rates.Then we train the models with a hybrid objective, such that they learn to maximize the performance of nextword prediction subject to explicit constraints on the total encoding precision.We find that explicit WM constraints improve the model's alignment with human reading times.More importantly, we find that the need to manage encoding precision reshapes the nature of the models' context representations, making them more compressed and categorical.Our results show how resource-rational models of WM allocation can be implemented in neural models simply and successfully, and point to a dissociation between WM retrieval mechanisms and the underlying memory representations in models of human sentence processing.

Large language models (LLMs) often exhibit systematic preferences for certain answer choices when responding to multiple-choice questions—a behavior known as selection bias. This bias reduces the accuracy and reliability of LLM outputs, limiting their usefulness in decision-critical applications. While prior work has focused on adjusting model inputs or outputs to mitigate this issue, our work takes a fundamentally different approach by identifying and removing the internal sources of bias. We introduce two methods: Bias Node Pruning (BNP), which prunes parameters that contribute to selection bias, and Auxiliary Option Injection (AOI), which introduces an additional answer choice to reduce bias in both white-box and black-box settings. To address the shortcomings of existing evaluation metrics, we propose Choice Kullback-Leibler Divergence (CKLD), a new metric that captures distributional imbalances in model predictions. Experiments on three LLMs across multiple datasets demonstrate that our methods consistently improve answer accuracy while reducing selection bias, providing a robust solution for both open- and closed-source models.

Topic modeling is a fundamental task in natural language processing, allowing the discovery of latent thematic structures in text corpora.While Large Language Models (LLMs) have demonstrated promising capabilities in topic discovery, their direct application to topic modeling suffers from issues such as incomplete topic coverage, misalignment of topics, and inefficiency.To address these limitations, we propose LLM-ITL, a novel LLM-in-theloop framework that integrates LLMs with Neural Topic Models (NTMs).In LLM-ITL, global topics and document representations are learned through the NTM.Meanwhile, an LLM refines these topics using an Optimal Transport (OT)-based alignment objective, where the refinement is dynamically adjusted based on the LLM's confidence in suggesting topical words for each set of input words.With the flexibility of being integrated into many existing NTMs, the proposed approach enhances the interpretability of topics while preserving the efficiency of NTMs in learning topics and document representations.Extensive experiments demonstrate that LLM-ITL helps NTMs significantly improve their topic interpretability while maintaining the quality of document representation.

The use of LLMs in natural language reasoning has shown mixed results, sometimes rivaling or even surpassing human performance in simpler classification tasks while struggling with social-cognitive reasoning, a domain where humans naturally excel. These differences have been attributed to many factors, such as variations in prompting and the specific LLMs used. However, no reasons appear conclusive, and no clear mechanisms have been established in prior work. In this study, we empirically evaluate how role-playing persona-based prompting influences Theory-of-Mind (ToM) reasoning capabilities. Grounding our research in psychological theory, we found that, beyond the inherent variance in the complexity of reasoning tasks, ToM performance differences arise because of socially-motivated prompting differences. In an era where prompt engineering with role-play is a typical approach to adapt LLMs to new contexts, our research advocates caution as models that adopt specific personas might potentially result in errors in social-cognitive reasoning.

Speaker verification is a core component of embodied robotic systems because it enables fast and secure user authentication. Mandarin numerical pass-phrases offer a compact lexical scope and high entropy, thus combining convenience with strong security and extending the same technology to voice-based financial payments. The community, however, lacks a public corpus of Chinese numerical strings. We address this gap by releasing SHALCAS22A, an 18.3-hour studio-quality corpus of Mandarin numerical utterances from 60 balanced speakers across nine rhythm-controlled templates, now hosted on OpenSLR1. Building on this resource, we present DE-CNSV, a dual-ended network that explicitly separates text and speaker embeddings. The text branch employs an enhanced Transformer trained with a composite loss that merges text classification, connectionist temporal classification, and sequence-to-sequence decoding objectives. The speaker branch adopts a sliding-window attentive statistics pooling mechanism to better capture temporal speaker characteristics from short utterances, thereby enhancing discriminability in low-duration scenarios. The proposed system attains an equal error rate of 0.32 % on the Hi-Mia dataset and 0.12 % on SHALCAS22A, establishing a new benchmark for Mandarin text-dependent speaker verification. These results demonstrate that numerical TD-SV can support both secure robot–human interaction and practical voice payment services.

Infrared small target detection (IRSTD) remains challenging due to the extremely low signal-to-noise ratio (SNR). Existing methods struggle to balance accuracy and speed, especially under limited computational resources. To address these issues, we propose the frequency-spatial contextual fusion network (FSCFNet) based on You Only Look Once (YOLO) v10n architecture. Particularly, the novel frequency-spatial convolution (FSConv) is designed that decomposes input features via Haar Wavelet Transform. High-frequency cues focus on local details to highlight small targets, while low-frequency cues provide global information to complement spatial features. Subsequently, the asymmetric cross-domain attention (ACA) is developed to enhance the local central feature extraction, which reflects the typical spatial Gaussian pattern of small targets. Furthermore, we introduce the customized multi-scale receptive contextual Block (MRCB) to capture the long-range information by leveraging diverse dilated convolutions. In addition, the Wasserstein Distance Loss (WDL) is utilized to improve bounding box quality. Extensive experiments on three public datasets including IRSTD-1k, NUDT-SIRST, and NUAA-SIRST confirm the effectiveness of FSCFNet. Notably, FSCFNet surpasses the baseline by 4.7% in precision, 3.3% in recall, and 3.9% in AP@50 on IRSTD-1k, with only a 3.6% increase in parameters. FSCFNet provides a robust solution for real-time infrared surveillance systems under resource-constrained environments. More comparisons are shown in Fig. 1.

Severe Coronavirus Disease 2019 (COVID-19) has been a global pandemic which provokes massive devastation to the society, economy, and culture since January 2020. The pandemic demonstrates the inefficiency of superannuated manual detection approaches and inspires novel approaches that detect COVID-19 by classifying chest x-ray (CXR) images with deep learning technology. Although a wide range of researches about bran-new COVID-19 detection methods that classify CXR images with centralized convolutional neural network (CNN) models have been proposed, the latency, privacy, and cost of information transmission between the data resources and the centralized data center will make the detection inefficient. Hence, in this article, a COVID-19 detection scheme via CXR images classification with a lightweight CNN model called MobileNet in edge computing is proposed to alleviate the computing pressure of centralized data center and ameliorate detection efficiency. Specifically, the general framework is introduced first to manifest the overall arrangement of the computing and information services ecosystem. Then, an unsupervised model DCGAN is employed to make up for the small scale of data set. Moreover, the implementation of the MobileNet for CXR images classification is presented at great length. The specific distribution strategy of MobileNet models is followed. The extensive evaluations of the experiments demonstrate the efficiency and accuracy of the proposed scheme for detecting COVID-19 over CXR images in edge computing.

More and more applications under Internet of Things have strong need for more dedicated localization techniques. As a wireless signal strength measurement standard, received signal strength indicator (RSSI) nowadays is widely utilized as a quantity to build advanced fingerprint indoor localization techniques. However, the mixed noise such as Gaussian noise together with the abrupt noise always causes the deviation of the RSSI value and the mismatched fingerprints in the fingerprint-based method, which results in the deterioration of positioning accuracy. In this paper, we propose an online risk-sensitive localization technique named compositional online kernel indoor localization (COKIL), which further improves the performance and reduces the prediction variance under multi-path effects. Meanwhile, the Student’s t kernel is firstly employed in COKIL to fight against RSSI instability, which leads to the great performance improvement compared with the Gaussian kernel . Moreover, surprise criterion, novelty criterion and kernel orthogonal matching pursuit are embedded into COKIL to reduce the size of the neural networks . Comparing their performances in experiments, surprise criterion is the optimal sparse method in practice. Finally, a new model-based technique, RSSIq, is proposed to deal with the missing fingerprints, which significantly improves the performance in indoor environment compared to traditional path-loss model.

Entering text is very challenging on smartwatches, especially on non-touch smartwatches where virtual keyboards are unavailable. In this paper, we designed and implemented COMPASS, a non-touch bezel-based text entry technique. COMPASS positions multiple cursors on a circular keyboard, with the location of each cursor dynamically optimized during typing to minimize rotational distance. To enter text, a user rotates the bezel to select keys with any nearby cursors. The design of COMPASS was justified by an iterative design process and user studies. Our evaluation showed that participants achieved a pick-up speed around 10 WPM and reached 12.5 WPM after 90-minute practice. COMPASS allows users to enter text on non-touch smartwatches, and also serves as an alternative for entering text on touch smartwatches when touch is unavailable (e.g., wearing gloves).