Xiaocheng Zhang, Xi Wang, Yifei Lu, Jianing Wang, Zhuangzhuang Ye, Mengjiao Bao, Peng Yan, Xiaohong Su. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Distributed Acoustic Sensing (DAS) systems provide rich vibration information for industrial monitoring. However, raw one-dimensional signals are often noisy and exhibit complex and heterogeneous temporal characteristics, making it difficult for traditional models to extract discriminative features and limiting classification performance and generalization. To address these challenges, we propose a DAS signal classification frame-work that integrates dynamic time-frequency representation with a learnable multi-source feature fusion network. First, a dynamic Short-Time Fourier Transform (STFT) method is introduced, which adaptively adjusts the window size and overlap ratio according to the input signal length. This design ensures consistent time-frequency resolution while avoiding distortions caused by padding or resampling, thereby enhancing spectral discriminability. Second, we design MFOFUNet, a dual-branch UNet-based architecture that separately encodes the dynamic spectrograms of intensity and phase channels and integrates them through a multi-stage fusion mechanism. This structure preserves modality-specific complementary information while enabling effective cross-modal interaction. Furthermore, a lightweight learnable fusion module is proposed to adaptively balance the contributions of different modalities at each encoding stage, achieving improved accuracy and stability with minimal additional complexity. In addition, a dual-stage data augmentation strategy is developed, combining 1D signal-level and 2D spectrogram-level enhancements to improve robustness. Extensive experiments on a challenging 10-class DAS dataset demonstrate that the proposed method achieves superior performance, reaching 89.33±4.27% accuracy, 89.02±4.50% F1-score, and 95.57±2.40% mAP under 10-fold cross-validation, outperforming existing mainstream approaches and showing strong potential for industrial sensing applications.

Iterative preference optimization has recently become one of the de-facto training paradigms for large language models (LLMs), but the performance is still underwhelming due to too much noisy preference data yielded in the loop. To combat this issue, we present an Uncertainty-enhanced Preference Optimization (UPO) framework to make the LLM self-evolve with reliable feedback. The key idea is mitigating the noisy preference pairs derived from the current policy and reward models by performing pair-wise uncertainty estimation and judiciously reliable feedback sampling. To reach this goal, we thus introduce an estimator model, which incorporates Monte Carlo (MC) dropout in Bayesian neural network (BNN) to perform uncertainty estimation for the batch of preference pairs. Compared to the existing methods that directly filter generated responses based on the reward score, the estimator focuses on the model uncertainty in a pair-wise manner and effectively bypasses the confirmation bias problem of the reward model. Additionally, we also propose an uncertainty-enhanced self-evolution algorithm to better improve the LLM robustly align with these reliable feedback data. Extensive experiments over multiple benchmarks demonstrate our framework substantially improves the performance of iterative preference optimization.

Intent-oriented controlled video captioning aims to generate targeted descriptions for specific targets in a video based on customized user intent. Current Large Visual Language Models (LVLMs) have gained strong instruction following and visual comprehension capabilities. Although the LVLMs demonstrated proficiency in spatial and temporal understanding respectively, it was not able to perform fine-grained spatial control in time sequences in direct response to instructions. This substantial spatio-temporal gap complicates efforts to achieve fine-grained intention-oriented control in video. Towards this end, we propose a novel IntentVCNet that unifies the temporal and spatial understanding knowledge inherent in LVLMs to bridge the spatio-temporal gap from both prompting and model perspectives. Specifically, we first propose a prompt combination strategy designed to enable LLM to model the implicit relationship between prompts that characterize user intent and video sequences. We then propose a parameter efficient box adapter that augments the object semantic information in the global visual context so that the visual token has a priori information about the user intent. The final experiment proves that the combination of the two strategies can further enhance the LVLM's ability to model spatial details in video sequences, and facilitate the LVLMs to accurately generate controlled intent-oriented captions. Our proposed method achieved state-of-the-art results in several open source LVLMs and was the runner-up in the IntentVC challenge. Our code is available on https://github.com/thqiu0419/IntentVCNet.

Fact verification constitutes a pivotal application in the effort to combat the dissemination of disinformation, a concern that has recently garnered considerable attention. However, previous studies in the field of fact verification, particularly those focused on question-answering dialogue, have exhibited limitations, such as failing to fully exploit the potential of question structures and ignoring relevant label information during the verification process. In this paper, we introduce Label-Infused Iterative Information Interacting (LI4), a novel approach designed for the task of question-answering dialogue based fact verification. LI4 consists of two meticulously designed components, namely the Iterative Information Refining and Filtering Module (IIRF) and the Fact Label Embedding Module (FLEM). The IIRF uses the Interactive Gating Mechanism to iteratively filter out the noise of question and evidence, concurrently refining the claim information. The FLEM is conceived to strengthen the understanding ability of the model towards labels by injecting label knowledge. We evaluate the performance of the proposed LI4 on HEALTHVER, FAVIQ, and COLLOQUIAL. The experimental results confirm that our LI4 model attains remarkable progress, manifesting as a new state-of-the-art performance.

The automatic intelligent identification of open-pit mines could improve the efficiency in the supervision of mines. In this paper, a massive sample database of open-pit mines was firstly established with remote sensing resolution of 0.5 m, 2 m from 2018 to 2021. Then, an intelligent identification algorithm for damaged land of open-pit mines was constructed with space pyramid, cavity convolution and other structures considered in the design part. Taking Zhongliang mountain of Chongqing as a study area, the manual visual interpretation results were applied to validate the cavity convolution model. The result showed that total precision of algorithm was more than 85% which improved the identification and extraction ability of open-pit mines. The precision with remote sensing image resolution of 0.5m were higher than that of 2m. The efficiency was 18 times compared with manual visual interpretation.

A gradual refinement strategy method to build a large-scale landcover semantic segmentation dataset of remote sensing image based on the existing monitoring data and historical orthophotos was proposed in this paper. Firstly, a fixed-size sliding window was used to slide on the remote sensing image and the coverage mask obtained from the monitoring vector data at the same time to extract the image slice and label slice, and the original dataset was composed of a large number of slices. Secondly, the original dataset was divided into several subsets by DBSCAN, and the error samples covered by clouds or shadows were effectively classified and deleted from the dataset. Then, the low-precision samples were detected and deleted from the dataset by an optimization and purification method based on iterative optimization intelligent model. Experimental results show that the method proposed in this paper could build a large-scale remote sensing image semantic segmentation dataset effectively based on historical orthophoto images and the monitoring vectors which were not strictly matching the images.

The monitoring of vegetation recovery was an important factor to evaluate the effectiveness of ecological restoration of open-pit mines. Lidar data can obtain the horizontal and vertical structures of trees that was wildly used to characterize the 3-dimensional (3D) structure of trees. In this paper, taking Ba’nan District of Chongqing as a study area, high density lidar data was applied to obtain the vegetation vertical structure parameters based on marker-controlled watershed segmentation method with IPTD filtering algorithm. The result showed, average tree height, diameter at breast height (DBH), crown diameter of open-pit mines was relatively lower compared with other areas outside as the trees of open-pit mines were mainly artificial planting and the growth time of trees was about 3-5 years. The trees had recovered well in the process of restoration. It indicated that the ecological restoration and planting of trees in open-pit mines were still in the process of recovery.

This paper studies the distributed state estimation problem for a class of discrete-time stochastic systems with nonlinear uncertain dynamics over time-varying topologies of sensor networks. An extended state vector consisting of the original state and the nonlinear dynamics is constructed. By analyzing the extended system, we provide a design method for the filtering gain and fusion matrices, leading to the extended state distributed Kalman filter. It is shown that the proposed filter can provide the upper bound of estimation covariance in real time, which means the estimation accuracy can be evaluated online. It is proven that the estimation covariance of the filter is bounded under rather mild assumptions, i.e., collective observability of the system and jointly strong connectedness of network topologies. Numerical simulation shows the effectiveness of the proposed filter.