To improve Multi-step Mathematical Reasoning (MsMR) of Large Language Models (LLMs), it is crucial to obtain scalable supervision from the corpus by automatically critiquing mistakes in the reasoning process of MsMR and rendering a final verdict of the problem-solution. Most existing methods rely on crafting high-quality supervised fine-tuning demonstrations for critiquing capability enhancement and pay little attention to delving into the underlying reason for the poor critiquing performance of LLMs. In this paper, we orthogonally quantify and investigate the potential reason — imbalanced evaluation preference, and conduct a statistical preference analysis. Motivated by the analysis of the reason, a novel perplexity-aware reinforcement learning algorithm is proposed to rectify the evaluation preference, elevating the critiquing capability. Specifically, to probe into LLMs' critiquing characteristics, a One-to-many Problem-Solution (OPS) benchmark is meticulously constructed to quantify the behavior difference of LLMs when evaluating the problem solutions generated by itself and others. Then, to investigate the behavior difference in depth, we conduct a statistical preference analysis oriented on perplexity and find an intriguing phenomenon — "LLMs incline to judge solutions with lower perplexity as correct", which is dubbed as imbalanced evaluation preference. To rectify this preference, we regard perplexity as the baton in the algorithm of Group Relative Policy Optimization, supporting the LLMs to explore trajectories that judge lower perplexity as wrong and higher perplexity as correct. Extensive experimental results on our built OPS and existing available critic benchmarks demonstrate the validity of our method.

Yuchun Fan, Bei Li, Peiguang Li, Yilin Wang, Yongyu Mu, Jian Yang, Xin Chen, Rongxiang Weng, Jingang Wang, Xunliang Cai, JingBo Zhu, Tong Xiao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Retrieval-Augmented Generation (RAG) demonstrates great value in alleviating outdated knowledge or hallucination by supplying LLMs with updated and relevant knowledge. However, RAG still faces several challenges in tackling complex multi-hop queries, which require LLMs to perform accurate reasoning and retrieval at each step. Inspired by the human reasoning process, where we progressively search for missing information after acquiring useful clues, it is natural to question whether LLMs have similar capabilities. In this work, we first experimentally verified the ability of LLMs to extract information from the retrieved knowledge as well as to know what is still missing. Based on the above discovery, we propose a Missing Information Guided Retrieve-Extraction-Solving paradigm (MIGRES), where we leverage the identification of missing information to generate a targeted query that steers the subsequent knowledge retrieval. Besides, we design a sentence-level re-ranking filtering approach to filter the irrelevant content from the document, along with the information extraction capability of LLMs to extract useful information from denoised documents. Extensive experiments conducted on multiple public datasets reveal the superiority of the proposed MIGRES method, and analytical experiments demonstrate the effectiveness of our proposed modules. Code and data are released in https://github.com/AdelWang/MIGRES.

In vehicle-road cooperation, the advancement of vehicle-side autonomous driving is hindered by bottlenecks, such as limited sensing range, computational power, and environmental adaptability. Collaboration with roadside units is essential for achieving more accurate and complex spatial perception. This study presents a new prototype to enhancing spatial perception in road environments within Augmented Intelligence of Things (AIoT) systems using light detection and ranging (LiDAR) technology. Unlike traditional AIoT systems, which rely on cameras and struggle in complex conditions, the proposed prototype uses an edge-cloud–terminal collaborative sensing model to enhance 3-D spatial perception. A notable feature of this prototype is the integration of the distance and density adaptive filtering (DDAF) method, which ensures efficient point cloud filtering at the edge, with an average F1-score of 96.03% and an average latency of 11.78 ms, demonstrating strong accuracy and low latency across various scenarios. The incorporation of DDAF further improves the mean average precision (mAP) of deep learning-based 3-D object detection on the cloud by 2.34%, reduces processing time by 83.54%, and decreases peak memory usage by 90.18%, facilitating precise 3-D spatial analysis. The final results are displayed in real-time on the terminal for visualization and interaction. The efficacy of this prototype is demonstrated through a real-world case study. This research highlights the role of LiDAR and AIoT in overcoming spatial perception challenges in vehicle-road cooperation, leading to safer, more efficient transportation solutions.

Zhicong Lu, Li Jin, Peiguang Li, Yu Tian, Linhao Zhang, Sirui Wang, Guangluan Xu, Changyuan Tian, Xunliang Cai. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

The sequential recommendation task aims to predict the item that user is interested in according to his/her historical action sequence. However, inevitable random action, i.e. user randomly accesses an item among multiple candidates or clicks several items at random order, cause the sequence fails to provide stable and high-quality signals. To alleviate the issue, we propose the StatisTics-Driven Pre-traing framework (called STDP briefly). The main idea of the work lies in the exploration of utilizing the statistics information along with the pre-training paradigm to stabilize the optimization of recommendation model. Specifically, we derive two types of statistical information: item co-occurrence across sequence and attribute frequency within the sequence. And we design the following pre-training tasks: 1) The co-occurred items prediction task, which encourages the model to distribute its attention on multiple suitable targets instead of just focusing on the next item that may be unstable. 2) We generate a paired sequence by replacing items with their co-occurred items and enforce its representation close with the original one, thus enhancing the model’s robustness to the random noise. 3) To reduce the impact of random on user’s long-term preferences, we encourage the model to capture sequence-level frequent attributes. The significant improvement over six datasets demonstrates the effectiveness and superiority of the proposal, and further analysis verified the generalization of the STDP framework on other models.

With the development of remote sensing (RS) acquisition technology, a mass of RS images have been produced, which brings challenges to the traditional manual retrieval methods and gives birth to the automatic RS image retrieval methods. Cross-modal RS image retrieval allows the usage of text and other modalities to retrieve RS images. For its flexible and convenient advantages, it has become a research hotspot. However, cross-modal RS image retrieval encounters the information asymmetry between modalities, i.e., RS images possess multi-scale, multi-objective properties and own rich information. At the same time, the query text is usually short and with less information. To solve the issues above, a cross-modal feature matching network is proposed to learn the feature fusion intra-modalities and the feature association inter-modalities to avoid the poor retrieval performance caused by the information asymmetry. Specifically, for the feature fusion intra-modalities, relying on the powerful feature representation ability of graph network, text and RS image graph modules are designed to fuse the intra-modal features. In terms of the feature correlation between modalities, RS image-text association module is created to attend the parts in text related to RS images and vice versa. Extended experiments on two public standard datasets verify the effectiveness of the proposed model.

The development of social media enables fake news to be expressed in a multi-modal form, which is disseminated on various social platforms and brings harmful social impacts. To handle this challenge, the fake news detection task was proposed to examine whether false information is contained in multi-modal news. Existing methods exploit various approaches with cross-modal interaction and fusion, which have proven to be effective in detecting common fake news. However, although the description of multi-modal news is narrated around entities, the previously developed methods pay less attention to this characteristic. They do not explore its benefits to the detection task and underperform with respect to the detection of fake news that requires entity-centric comparisons. To make up for this omission, we explore a novel paradigm to detect fake news by aligning and fusing multi-modal entities and propose the Entity-oriented Multi-modal Alignment and Fusion network (EMAF). Our work adopts entity-centric cross-modal interaction, which can reserve semantic integrity and capture the details of multi-modal entities. Specifically, we design an Alignment module with the improved dynamic routing algorithm and introduce a Fusion module based on the comparison, the former aligns and captures the important entities and the latter compares and aggregates entity-centric features. Comparative experiments conducted on multiple public datasets, including Weibo, Twitter, and Reddit, reveal the superiority of the proposed EMAF method, and extensive analytical experiments demonstrate the effectiveness of our proposed modules.

Cloud removal is an essential step to enhance the quality of cloud-covered remote sensing image. In recent years, conditional Generative Adversarial Network (cGAN) yields promising improvement in plentiful image-to-image translation tasks. In this paper, we propose a novel objective function to upgrade the structural similarity index based on cGAN. We discover that ImageGAN is effective to focus on global information for thick cloud-covered images and Patch-GAN has fewer parameters while maintaining outstanding performance for thin cloud-covered remote sensing images in the experiments. Experimental results demonstrate that our method achieves remarkable performance in both PSNR, SSIM and visual effect on cloud-covered remote sensing images especially thin cloud-covered images.