Haotian Xu, Yue Hu, Zhengqiu Zhu, Chen Gao, Ziyou Wang, Junreng Rao, Wenhao Lu, Weishi Li, Quanjun Yin, Yong Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The demand for Low-earth-orbit (LEO) satellite networking and routing is continuously growing. Container-based satellite network emulation becomes an important networking evaluation choice for LEO satellite network. While many container-based satellite network emulators optimize virtual network efficiency, they often overlook route installation latency—the duration of installing computed routes into kernel routing table. This latency represents a significant bottleneck for emulating large-scale LEO constellations, which are characterized by dynamic topologies and frequent, massive routing re-convergences. High routing installation latency not only reduces emulation efficiency but also compromises evaluation reliability and renders subsequent data transmission experiments infeasible. In this paper, we present a Lock-Free and Filtered Route Installation (LFRI) mechanism. LFRI filters out non-effective route installations, thereby shortening the installation procedure without changing the routing results. LFRI employs a lock-free installation mechanism that enables parallelized and efficient route installation. Experimental results demonstrate that, compared to the legacy Netlink installation, LFRI reduces the route installation latency by up to 99.9% in emulations of typical LEO satellite constellations, bringing it down to under ten milliseconds, within a performance level comparable to routing installation on physical machines. Meanwhile, in the case study evaluating path availability ratio and restoration time, LFRI yields more consistent and credible experimental data.

The sigmoid function, as a widely used activation function in neural networks, has gained much attention for its approximation and associated usage in edge devices. A recent study applied the Gaussian cumulative function to approximate the sigmoid function. Although this probabilistic method simplifies hardware implementation through a low-complexity binary search, it requires intensive random access memory (RAM) storage, and the search process is time-consuming. Besides, it targets minimizing the maximum mapping error rather than ensuring accurate approximation across all inputs. To address these issues, this article proposes a hardware-friendly and high-accuracy probabilistic-based sigmoid approximator. We first present that given an input, the output of a sigmoid function is strictly equivalent to the probability of a logistic random variable less than or equal to this input. Then, an indirect random variable quantizing strategy is exhibited to reduce memory usage and concurrently minimize precision loss. The latency for the proposed scheme is also optimized. Afterward, a resource-efficient and low-latency sigmoid approximator is developed on digital circuits. Finally, we derive an upper bound on the absolute error between the approximator's output and the true value. Experiments verify the usefulness of our scheme and showcase superior performance in approximation accuracy and resource cost.

Quantization is widely adopted to accelerate inference and reduce memory consumption in large language models (LLMs).While activation-weight joint quantization enables efficient low-precision decoding, it suffers from substantial performance degradation on multistep reasoning tasks.We propose QSPEC, a novel quantization paradigm that decouples efficiency from quality by integrating two complementary schemes via speculative decoding: low-precision joint quantization for fast drafting and high-precision weight-only quantization for accurate verification.QSPEC reuses both weights and KV cache across stages, enabling near-zero-cost switching without retraining or auxiliary models.Compared to highprecision baselines, QSPEC achieves up to 1.64× speedup without quality degradation, and outperforms state-of-the-art speculative decoding methods by up to 1.55× in batched settings.Furthermore, QSPEC supports plug-andplay deployment and generalizes well across model scales, quantization methods, and workloads.These properties make QSPEC a practical and scalable solution for high-fidelity quantized LLM serving under memory-constrained scenarios.Our code is available at https: //github.com/hku-netexplo-lab/QSpec.

Low-earth-orbit (LEO) mega-constellations with inter-satellite links (ISLs) are becoming the Internet backbone in space. Satellites within LEO often need the capability to enforce data forwarding paths. For example, they may need to bypass the satellites over the untrusted areas for the data of mission-critical applications or minimize latency for the data of time-sensitive applications. However, typical source/segment routing techniques (e.g., SRv6) suffer from scalability issue, since they record source-route-style forwarding information via the list-based structure. This results in great payload and forwarding overhead. To overcome this drawback, we propose a source/segment routing architecture for LEO mega-constellations, which is named as Link-identified Routing (LiR). LiR leverages in-packet bloom filter (BF) to record source-route-style forwarding information. BF could efficiently record multiple elements via a probabilistic data structure, but overlooks the order of the encoded elements. To address this, LiR identifies each unidirectional ISL, and represents the path by encoding ISL identifiers into BF. We investigate how to optimize BF configuration and ISL encoding policy to address false positives caused by BF. We implement LiR in Linux kernel and develop a container-based emulator for performance evaluation. Results show that LiR significantly outperforms SRv6 in terms of packet forwarding and data delivery efficiency.

Low- earth-orbit (LEO) satellite constellations (e.g., Starlink) are becoming a necessary component of future Internet. There have been increasing studies on LEO satellite networking. It is a crucial problem how to evaluate these studies in a systematic and reproducible manner. In this paper, we present OpenSN, i.e., an open source library for emulating large-scale satellite network (SN). Different from Mininet-based SN emulators (e.g., LeoEM), OpenSN adopts container-based virtualization, thus allows for running distributed routing software on each node, and can achieve horizontal scalability via flexible multi-machine extension. Compared to other container-based SN emulators (e.g., StarryNet), OpenSN streamlines the interaction with Docker command line interface and significantly reduces unnecessary operations of creating virtual links. These modifications improve emulation efficiency and vertical scalability on a single machine. Furthermore, OpenSN separates user-defined configuration from container network management via a Key-Value Database that records the necessary information for SN emulation. Such a separation architecture enhances the function extensibility. To sum up, OpenSN exhibits advantages in efficiency, scalability, and extensibility, thus is a valuable open source library that empowers research on LEO satellite networking. Experiment results show that OpenSN constructs mega-constellations 5X-10X faster than StarryNet, and updates link state 2X-4X faster than LeoEM. We also verify the scalability of OpenSN by successfully emulating the five-shell Starlink constellation with a total of 4408 satellites.

Low-earth-orbit (LEO) satellite constellations (e.g., Starlink) are becoming the necessary component of future Internet. There have been increasing studies on LEO satellite networking. It is a crucial problem how to evaluate these studies in a systematic and reproducible manner. In this paper, we present OpenSN, i.e., an open-source library for emulating large-scale satellite network (SN). Different from Mininet-based SN emulators (e.g., LeoEM), OpenSN adopts container-based virtualization, thus allows for running distributed routing software on each node, and can achieve horizontal scalability via flexible multi-machine extension. Compared to other container-based SN emulators (e.g., StarryNet), OpenSN streamlines the interaction with Docker command line interface and significantly reduces unnecessary operation of creating virtual links. These modifications improve emulation efficiency and vertical scalability on a single machine. Furthermore, OpenSN separates user-defined configuration from container network management via a key-value database that records the necessary information of SN emulation. Such a separation architecture enhances the function extensibility. To sum up, OpenSN exhibits advantages in efficiency, scalability, and extensibility, thus is a valuable open-source library that empowers research on satellite networking. Experimental results show that OpenSN can construct mage-constellations 5X-10X faster than StarryNet, and update link state 2X-4X faster than Mininet.

Recent advancements in large language models have showcased their remarkable generalizability across various domains. However, their reasoning abilities still have significant room for improvement, especially when confronted with scenarios requiring multi-step reasoning. Although large language models possess extensive knowledge, their reasoning often fails to effectively utilize this knowledge to establish a coherent thinking paradigm. These models sometimes show hallucinations as their reasoning procedures are unconstrained by logical principles. Aiming at improving the zero-shot chain-of-thought reasoning ability of large language models, we propose LoT (Logical Thoughts), a self-improvement prompting framework that leverages principles rooted in symbolic logic, particularly Reductio ad Absurdum, to systematically verify and rectify the reasoning processes step by step. Experimental evaluations conducted on language tasks in diverse domains, including arithmetic, commonsense, symbolic, causal inference, and social problems, demonstrate the efficacy of enhanced reasoning by logic. The implementation code for LoT can be accessed at: https://github.com/xf-zhao/LoT.

In this paper, we demonstrate an energy-efficient real-time object detection and classification system featuring a hybrid event-based frame generation pipeline and a background-removal region proposal algorithm. The event-based frame is generated by aggregating active events within a programmable time interval, generating an event-based binary image (EBBI). This approach enables the utilization of low-complexity algorithms for denoising and object detection. The background-removal region proposal algorithm reduces memory requirements and removes dynamic backgrounds, leading to better detection performance. The proposed system is demonstrated on Zynq-7000 FPGA device with a DAVIS346 sensor. Experimental results show that the proposed system achieves comparable detection accuracy while requiring significantly less computation than existing event-based trackers.

Neuromorphic vision sensors (NVSs) mimic the function of the human visual system, with significant energy-saving potential in IoT-based object detection systems. Unlike conventional sensors, NVSs only generate asynchronous spiking events in response to changes in light intensity. However, the inherent noise generated by NVSs causes a degradation of detection performance. Moreover, an interested object usually occupies only a portion of the entire image frame. Therefore, a real-time, accurate event-based object detection system is needed to identify the region of interest (Rol) and leverage this spatial redundancy to reduce computational load in subsequent recognition modules. In this article, we present an energy-efficient real-time object detection system featuring a hybrid event-based frame generation pipeline and a background-removal region proposal algorithm. The event-based frame is generated by aggregating active events within a programmable time interval, generating an event-based binary image (EBBI). This approach enables the utilization of low-complexity algorithms for denoising and object detection. The background-removal region proposal algorithm reduces memory requirements and removes dynamic backgrounds, leading to better detection performance. The proposed system is demonstrated on Zynq-7000 FPGA device with a DAVIS346 sensor. Experimental results show that the proposed system achieves comparable accuracy while requiring significantly less computation than existing event-based trackers.

Dynamic vision sensors (DVS) have become prevalent in edge vision applications due to their low power and short latency attributes. However, current DVS-based object tracking systems suffer from high power consumption or long processing latency due to high computing intensity of the object detection algorithms. This paper proposes an energy-efficient object detection system through algorithm and hardware co-optimization. We design hardware-efficient denoising and region proposal (RP) algorithms to reduce on-chip memory usage and power consumption. Besides, the processing latency is dramatically reduced thanks to the less computing complexity. The devised algorithm is executed on a heterogeneous platform, with segments particularly sensitive to latency being accelerated via FPGA. An RP processor, supporting both parallel and systolic computing modes, is developed to facilitate the computing-intensive RP generation. Remarkably, the proposed system reduces the on-chip memory by 95.3% in contrast to traditional methods that employ connected component labeling. Moreover, the processing time per frame stands at 92.2 ms, marking a reduction of 82.4% compared to CPU-only operations.

Landform elements such as ridgeline and valleyline play a crucial role in understanding the Earth’s surface, its composition, geomorphic processes, and the changes over time. With the rapid advancement of remote sensing technologies, high density and very high-density point cloud data have enabled fine-scaled landform characterization and terrain analysis. However, extracting ridgeline and valleyline at fine scales remains a challenge for conventional approaches, which are often difficult to address inherent errors and uncertainties in the data, and distinguish different landform elements that exhibit similar characteristics. To address these challenges, we generate a large-scale dataset involving the labels for two major landform elements (ridge lines and valley lines) at different scales. We then propose a framework that combines spatial-contextual approach and neural networks for multiscale ridgeline and valleyline extraction. In the experimental section, we evaluated the performance of our proposed approach by comparing the accuracy and recall of ridgeline and valleyline extraction using a variety of machine learning-based and deep learning-base approaches. The results demonstrate that our method effectively leverages the advantages of feature learning and enhances the robustness of multiscale terrain morphology extraction and recognition. We hope our work can provide a explainable deep learning solution for multiscale landform element extraction in the community of geomorphology.

The sigmoid function is a representative activation function in shallow neural networks. Its hardware realization is challenging due to the complex exponential and reciprocal operations. Existing studies applied piecewise models to approximate sigmoid function and employed numerical methods or non-uniform input segmentations to mitigate fitting inaccuracies. However, the breakpoints introduce inevitable approximation precision loss. Besides, additional fitting processes greatly increase hardware complexity and power consumption. This paper presents a hardware-friendly sigmoidal approximation from the perspective of probability theory. We find that for a given input, the output of a sigmoid function can be approximated by the probability that the sum of this input and a Gaussian random variable is greater than or equal to zero. As the derived theorem does not involve piecewise expressions, the precision loss caused by the breakpoint issue is avoided. A low-complexity binary-search-based address localization method is proposed to optimize our theorem for hardware implementation. For the optimized scheme, an efficient implemented circuit is also presented. Our scheme’s approximation ability and hardware efficiency are validated through software modeling and FPGA- and ASIC-based experiments. Feedforward neural network-based classification applications demonstrate that building networks with the proposed sigmoid approximator has only a tiny recognition rate loss.

The dual neural network (DNN)-based k -winner-take-all (WTA) model is able to identify the k largest numbers from its m input numbers. When there are imperfections, such as non-ideal step function and Gaussian input noise, in the realization, the model may not output the correct result. This brief analyzes the influence of the imperfections on the operational correctness of the model. Due to the imperfections, it is not efficient to use the original DNN- k WTA dynamics for analyzing the influence. In this regard, this brief first derives an equivalent model to describe the dynamics of the model under the imperfections. From the equivalent model, we derive a sufficient condition for which the model outputs the correct result. Thus, we apply the sufficient condition to design an efficiently estimation method for the probability of the model outputting the correct result. Furthermore, for the inputs with uniform distribution, a closed form expression for the probability value is derived. Finally, we extend our analysis for handling non-Gaussian input noise. Simulation results are provided to validate our theoretical results.

Among many -winners-take-all ( WTA) models, the dual-neural network (DNN- WTA) model is with significantly less number of connections. However, for analog realization, noise is inevitable and affects the operational correctness of the WTA process. Most existing results focus on the effect of additive noise. This brief studies the effect of time-varying multiplicative input noise. Two scenarios are considered. The first one is the bounded noise case, in which only the noise range is known. Another one is for the general noise distribution case, in which we either know the noise distribution or have noise samples. For each scenario, we first prove the convergence property of the DNN- WTA model under multiplicative input noise and then provide an efficient method to determine whether a noise-affected DNN- WTA network performs the correct WTA process for a given set of inputs. With the two methods, we can efficiently measure the probability of the network performing the correct WTA process. In addition, for the case of the inputs being uniformly distributed, we derive two closed-form expressions, one for each scenario, for estimating the probability of the model having correct operation. Finally, we conduct simulations to verify our theoretical results.

Personalized retrieval seeks to retrieve items relevant to a user event (e.g. a page visit or a query) that are adapted to the user's personal preferences. For example, two users who happen to perform the same event such as visiting the same product page or asking the same query should receive potentially distinct recommendations adapted to their individual tastes. Personalization is seldom attempted over catalogs of millions of items since the cost of existing personalization routines scale linearly in the number of candidate items. For example, performing two-sided personalized retrieval (with both event and item embeddings personalized to the user) incurs prohibitive storage and compute costs. Instead, it is common to use non-personalized retrieval to obtain a small shortlist of items over which personalized re-ranking can be done quickly. Despite being scalable, this strategy risks losing items uniquely relevant to a user that fail to get shortlisted during non-personalized retrieval. This paper bridges this gap by developing the XPERT algorithm that identifies a form of two-sided personalization that can be scalably implemented over millions of items and hundreds of millions of users. Key to overcoming the computational challenges of personalized retrieval is a novel concept of morph operators that can be used with arbitrary encoder architectures, completely avoids the steep memory overheads of two-sided personalization, provides millisecond-time inference and offers multi-intent retrieval. On multiple public and proprietary datasets, XPERT offered upto 5% superior recall and AUC than state-of-the-art techniques. Code for XPERT is available at https://github.com/personalizedretrieval/xpert.

The dual neural network-based$k$-winner-take-all (DNN-$k$WTA) is an analog neural model that is used to identify the$k$largest numbers from$n$inputs. Since threshold logic units (TLUs) are key elements in the model, offset voltage drifts in TLUs may affect the operational correctness of a DNN-$k$WTA network. Previous studies assume that drifts in TLUs follow some particular distributions. This brief considers that only the drift range, given by$[-\Delta, \Delta]$, is available. We consider two drift cases: time-invariant and time-varying. For the time-invariant case, we show that the state of a DNN-$k$WTA network converges. The sufficient condition to make a network with the correct operation is given. Furthermore, for uniformly distributed inputs, we prove that the probability that a DNN-$k$WTA network operates properly is greater than$(1-2\Delta)^{n}$. The aforementioned results are generalized for the time-varying case. In addition, for the time-invariant case, we derive a method to compute the exact convergence time for a given data set. For uniformly distributed inputs, we further derive the mean and variance of the convergence time. The convergence time results give us an idea about the operational speed of the DNN-$k$WTA model. Finally, simulation experiments have been conducted to validate those theoretical results.

Pre-trained language models have achieved great success in a wide variety of natural language processing (NLP) tasks, while the superior performance comes with high demand in computational resources, which hinders the application in low-latency information retrieval (IR) systems. To address the problem, we present TwinBERT model, which has two improvements: 1) represent query and document separately using twin-structured encoders and 2) each encoder is a highly compressed BERT-like model with less than one third of the parameters. The former allows document embeddings to be pre-computed offline and cached in memory, which is different from BERT, where the two input sentences are concatenated and encoded together. The change saves large amount of computation time, however, it is still not sufficient for real-time retrieval considering the complexity of BERT model itself. To further reduce computational cost, a compressed multi-layer transformer encoder is proposed with special training strategies as a substitution of the original complex BERT encoder. Lastly, two versions of TwinBERT are developed to combine the query and keyword embeddings for retrieval and relevance tasks correspondingly. Both of them have met the real-time latency requirement and achieve close or on-par performance to BERT-Base model.

In this paper, we propose a novel image search scheme, contextual image search. Different from conventional image search schemes that present a separate interface (e.g., text input box) to allow users to submit a query, the new search scheme enables users to search images by only masking a few words when they are reading through Web pages or other documents. Rather than merely making use of the explicit query input that is often not sufficient to express user's search intent, our approach explores the context information to better understand the search intent with two key steps: query augmenting and search results reranking using context, and expects to obtain better search results. Beyond contextual Web search, the context in our case is much richer and includes images besides texts. In addition to this type of search scheme, called contextual image search with text input, we also present another type of scheme, called contextual image search with image input, to allow users to select an image as the search query from Web pages or other documents they are reading. The key idea is to use the search-to-annotation technique and the contextual textual query mining scheme to determine the corresponding textual query, to finally get semantically similar search results. Experiments show that the proposed schemes make image search more convenient and the search results are more relevant to user intention.

In this paper, the problem of part detection, description and selection is discussed. This problem is crucial in the learning algorithms of part-based models, but can't be solved well when some candidate parts are extracted from background. This paper studies this problem and introduces a new algorithm, HCRF-PS (Hidden Conditional Random Fields for Part Selection), for part detection, description, especially selection. Our algorithm is distinguished for its power to optimize multiple kinds of information at the same time, including texture, color, location and part label. Finally, we did some experiments with HCRF-PS algorithm which give good results on both virtual and real data.

Aiming at vehicle detection and tracking problems in video monitoring and controlling system, this paper mainly studies vehicle detection and tracking problems in conditions of high traffic density in daytime. This paper is distinguished by two key contributions. First, we develop an improvement — SEAP(Simple but Efficient After Process) which checks the detection results in an accurate way and is an after process of Adaboost [1] detector which used to detect car in every frame. Second, we propose a tracking algorithm named 4-states tracking algorithm based on Kalman[5] linear filter. Tracking results turn unsteady as traffic density grows higher because of much more false positives and false negatives appear. However, 4-states tracking algorithm can solve this problem in an easy way by introducing FSM (Finite State Machine) into tracking algorithm. Finally, we implement a real-time vehicle detection and tracking system with the upper methods. Experiments give good results in relative crowded Conditions.