Speech-language models (SLMs) offer a promising path toward unifying speech and text understanding and generation. However, challenges remain in achieving effective cross-modal alignment and high-quality speech generation. In this work, we systematically investigate the role of speech tokenizer designs in LLM-centric SLMs, augmented by speech heads and speaker modeling. We compare coupled, semi-decoupled, and fully decoupled speech tokenizers under a fair SLM framework and find that decoupled tokenization significantly improves alignment and synthesis quality. To address the information density mismatch between speech and text, we introduce multi-token prediction (MTP) into SLMs, enabling each hidden state to decode multiple speech tokens. This leads to up to 12× faster decoding and a substantial drop in word error rate (from 6.07 to 3.01). Furthermore, we propose a speaker-aware generation paradigm and introduce RoleTriviaQA, a large-scale role-playing knowledge QA benchmark with diverse speaker identities. Experiments demonstrate that our methods enhance both knowledge understanding and speaker consistency.

Dingwei Zhu, Shihan Dou, Zhiheng Xi, Senjie Jin, Guoqiang Zhang, Jiazheng Zhang, Junjie Ye, Mingxu Chai, Enyu Zhou, Ming Zhang, Yuhui Wang, Caishuang Huang, Chenhao Huang, Yunke Zhang, Yuran Wang, Tao Gui, Qi Zhang, Xipeng Qiu, Xuanjing Huang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Open Named Entity Recognition (NER), which involves identifying arbitrary types of entities from arbitrary domains, remains challenging for Large Language Models (LLMs). Recent studies suggest that fine-tuning LLMs on extensive NER data can boost their performance. However, training directly on existing datasets neglects their inconsistent entity definitions and redundant data, limiting LLMs to dataset-specific learning and hindering out-of-domain adaptation. To address this, we present B2NERD, a compact dataset designed to guide LLMs’ generalization in Open NER under a universal entity taxonomy. B2NERD is refined from 54 existing English and Chinese datasets using a two-step process. First, we detect inconsistent entity definitions across datasets and clarify them by distinguishable label names to construct a universal taxonomy of 400+ entity types. Second, we address redundancy using a data pruning strategy that selects fewer samples with greater category and semantic diversity. Comprehensive evaluation shows that B2NERD significantly enhances LLMs’ Open NER capabilities. Our B2NER models, trained on B2NERD, outperform GPT-4 by 6.8-12.0 F1 points and surpass previous methods in 3 out-of-domain benchmarks across 15 datasets and 6 languages. The data, models, and code are publicly available at https://github.com/UmeanNever/B2NER.

Daiye Miao, Yufang Liu, Jie Wang, Changzhi Sun, Yunke Zhang, Demei Yan, Shaokang Dong, Qi Zhang, Yuanbin Wu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Online-to-offline (O2O) food delivery platforms have greatly expanded urban residents' access to a wide range of food options by allowing convenient ordering from distant food outlets. However, concerns persist regarding the nutritional quality of delivered food, particularly as the impact of O2O food delivery platforms on users’ healthy food remains unclear. This study leverages large-scale empirical data from a leading O2O delivery platform to comprehensively analyze online food choice behaviors and how they are influenced by the online exposure to fast food restaurants, i.e., online food environment. Our analyses reveal significant variations in food preferences across demographic groups and city sizes, where male, low-income, and younger users are more likely to order fast food via O2O platforms. Besides, we also perform a comparative analysis on the food exposure differences in offline and online environments, confirming that the extended service ranges of O2O platforms can create larger "cyber food swamps". Furthermore, regression analysis highlights that a higher ratio of fast food orders is associated with "cyber food swamps", areas characterized by a higher proportion of accessible fast food restaurants. A 10% increase in this proportion raises the probability of ordering fast food by 22.0%. Moreover, a quasi-natural experiment substantiates the long-term causal effect of online food environment changes on healthy food choices. These findings underscore the need for O2O food delivery platforms to address the health implications of online food choice exposure, offering critical insights for stakeholders aiming to improve dietary health among urban populations.

The rapid pace of urbanization has led to unequal benefits for residents, creating significant inequality issues and discussions around Sustainable Development Goals 10 and 11. Accurate measurement of inequality within urban areas is essential for effective mitigation strategies. Traditional methods rely on survey-based census data, which are time-consuming and delayed, while some studies use coarse proxies like nighttime lights. However, these methods are limited by resolution and fail to capture fine-grained disparities within communities. To address this, we aim to leverage accessible urban imagery, which offers detailed visual features. Two key challenges must be addressed: 1) accurately perceiving micro-level inequalities within neighborhoods, and 2) ensuring that this perception is interpretable for policy guidance. To address these gaps, we propose UI-CoT, a framework that leverages the power of urban imagery-based visual language models in urban inequality perceiving, enhanced by Chain-of-Thought prompting to improve reasoning capabilities. We fine-tune a visual language model to predict three essential neighborhood inequality indicators: the income Gini coefficient, dominant race, and racial income ratio. Extensive experiments show that our model can effectively perceive micro-level inequalities, with the incorporation of Chain-of-Thought reasoning further improving the model's performance by 17.2%. This research offers valuable insights into addressing inequalities within urban environments and demonstrates the potential of web resources in empowering urban sustainable development. The code and data are available at https://github.com/tsinghua-fib-lab/UI-CoT.

Predicting citizens’ visiting behaviors to urban facilities is instrumental for city governors and planners to detect inequalities in urban opportunities and optimize the distribution of facilities and resources. Previous works predict facility visits simply using observed visit behavior, yet citizens’ intrinsic demands for facilities are not characterized explicitly, causing potential incorrect learned relations in the prediction results. In this article, to make up for this deficiency, we present a demand-driven urban facility visit prediction method that decomposes citizens’ visits to facilities into their unobservable demands and their capability to fulfill them. Demands are expressed as the function of regional demographic attributes by a neural network, and the fulfillment capability is determined by the urban region’s spatial accessibility to facilities. Extensive evaluations of datasets of three large cities confirm the efficiency and rationality of our model. Our method outperforms the best state-of-the-art model by 8.28% on average in facility visit prediction tasks. Further analyses demonstrate the reasonableness of recovered facility demands and their relationship with citizen demographics. For instance, senior citizens tend to have higher medical demands but lower shopping demands. Meanwhile, estimated capabilities and accessibilities provide deeper insights into the decaying accessibility with respect to spatial distance and facilities’ diverse functions in the urban environment. Our findings shed light on demand-driven urban data mining and demand-based urban facility planning.

This paper proposes a novel active boundary loss for semantic segmentation. It can progressively encourage the alignment between predicted boundaries and ground-truth boundaries during end-to-end training, which is not explicitly enforced in commonly used cross-entropy loss. Based on the predicted boundaries detected from the segmentation results using current network parameters, we formulate the boundary alignment problem as a differentiable direction vector prediction problem to guide the movement of predicted boundaries in each iteration. Our loss is model-agnostic and can be plugged in to the training of segmentation networks to improve the boundary details. Experimental results show that training with the active boundary loss can effectively improve the boundary F-score and mean Intersection-over-Union on challenging image and video object segmentation datasets.

We consider semantic image segmentation. Our method is inspired by Bayesian deep learning which improves image segmentation accuracy by modeling the uncertainty of the network output. In contrast to uncertainty, our method directly learns to predict the erroneous pixels of a segmentation network, which is modeled as a binary classification problem. It can speed up training comparing to the Monte Carlo integration often used in Bayesian deep learning. It also allows us to train a branch to correct the labels of erroneous pixels. Our method consists of three stages: (i) predict pixel-wise error probability of the initial result, (ii) redetermine new labels for pixels with high error probability, and (iii) fuse the initial result and the redetermined result with respect to the error probability. We formulate the error-pixel prediction problem as a classification task and employ an error-prediction branch in the network to predict pixel-wise error probabilities. We also introduce a detail branch to focus the training process on the erroneous pixels. We have experimentally validated our method on the Cityscapes and ADE20K datasets. Our model can be easily added to various advanced segmentation networks to improve their performance. Taking DeepLabv3+ as an example, our network can achieve 82.88% of mIoU on Cityscapes testing dataset and 45.73% on ADE20K validation dataset, improving corresponding DeepLabv3+ results by 0.74% and 0.13% respectively.

In recent years, China has witnessed the proliferation and success of the online food delivery industry, an emerging type of the gig economy. Online food deliverers who deliver the food from restaurants to customers play a critical role in enabling this industry. Mediated by algorithms and coupled with interactions with multiple stakeholders, this emerging kind of labor has been taken by millions of people. In this paper, we present a mixed-methods analysis to investigate this labor of online food deliverers and uncover how the mediation of algorithms shapes it. Combining large-scale quantitative data-driven investigations of 100,000 deliverers' behavioral data with in-depth qualitative interviews with 15 online food deliverers, we demonstrate their working activities, identify how algorithms mediate their delivery procedures, and reveal how they perceive their relationships with different stakeholders as a result of their algorithm-mediated labor. Our findings provide important implications for enabling better experiences and more humanized labor of deliverers as well as workers in gig economies of similar kinds.

This paper proposes a novel deep learning-based video object matting method that can achieve temporally coherent matting results. Its key component is an attention-based temporal aggregation module that maximizes image matting networks' strength for video matting networks. This module computes temporal correlations for pixels adjacent to each other along the time axis in feature space, which is robust against motion noises. We also design a novel loss term to train the attention weights, which drastically boosts the video matting performance. Besides, we show how to effectively solve the trimap generation problem by fine-tuning a state-of-the-art video object segmentation network with a sparse set of user-annotated keyframes. To facilitate video matting and trimap generation networks' training, we construct a large-scale video matting dataset with 80 training and 28 validation foreground video clips with ground-truth alpha mattes. Experimental results show that our method can generate high-quality alpha mattes for various videos featuring appearance change, occlusion, and fast motion. Our code and dataset can be found at: https://github.com/yunkezhang/TCVOM

This paper studies the structure of a deep convolutional neural network to predict the foreground alpha matte by taking a single RGB image as input. Our network is fully convolutional with two decoder branches for the foreground and background classification respectively. Then a fusion branch is used to integrate the two classification results which gives rise to alpha values as the soft segmentation result. This design provides more degrees of freedom than a single decoder branch for the network to obtain better alpha values during training. The network can implicitly produce trimaps without user interaction, which is easy to use for novices without expertise in digital matting. Experimental results demonstrate that our network can achieve high-quality alpha mattes for various types of objects and outperform the state-of-the-art CNN-based image matting methods on the human image matting task.