Jiwei Tang, Shilei Liu, Zhicheng Zhang, Qingsong Lv, Runsong Zhao, Tingwei Lu, Langming Liu, Haibin Chen, Yujin Yuan, Hai-Tao Zheng, Wenbo Su, Bo Zheng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Entity Set Expansion (ESE) aims to identify new entities belonging to the same semantic class as the given set of seed entities. Traditional methods solely relied on positive seed entities to represent the target fine-grained semantic class, rendering them tough to represent ultra-fine-grained semantic classes. Specifically, merely relying on positive seed entities leads to two inherent shortcomings: (i) Ambiguity among ultra-fine-grained semantic classes. (ii) Inability to define “unwanted” semantics. Hence, previous ESE methods struggle to address the ultra-fine-grained ESE (Ultra-ESE) task. To solve this issue, we first introduce negative seed entities in the inputs, which jointly describe the ultra-fine-grained semantic class with positive seed entities. Negative seed entities eliminate the semantic ambiguity by providing a contrast between positive and negative attributes. Meanwhile, it provides a straightforward way to express “unwanted”. To assess model performance in Ultra-ESE and facilitate further research, we also constructed UltraWiki, the first large-scale dataset tailored for Ultra-ESE. UltraWiki encompasses 50,973 entities and 394,097 sentences, alongside 236 ultra-fine-grained semantic classes, where each class is represented with 3–5 positive and negative seed entities. Moreover, a retrieval-based framework RetExpan and a generation-based framework GenExpan are proposed to provide powerful baselines for Ultra-ESE. Additionally, we devised two strategies to enhance models' comprehension of ultra-fine-grained entities' semantics: contrastive learning and chain-of-thought reasoning. Extensive experiments confirm the effectiveness of our proposed strategies and also reveal that there remains a large space for improvement in Ultra-ESE. All the codes, dataset, and supplementary notes are available at https://github.com/THUKElab/UltraWiki.

Vision-Language Models (VLMs) have shown broad effectiveness due to extensive training that aligns visual inputs with corresponding language responses. However, this conclusive alignment training causes models to overlook essential visual reasoning, leading to failures in handling detailed visual tasks and producing unfaithful responses. Drawing inspiration from human cognition in solving visual problems (e.g., marking, zoom in), this paper introduces Chain of Manipulations, a mechanism that enables VLMs to tackle problems step-by-step with evidence. After training, models can solve various visual problems by eliciting intrinsic manipulations (e.g., grounding, zoom in) with results (e.g., boxes, image) actively without relying external tools, while also allowing users to trace error causes. In this paper, we study the comprehensive methodology that includes: (1) a flexible design of manipulations based on extensive analysis, (2) an efficient automated data generation pipeline, (3) a compatible VLM architecture capable of multi-turn, multi-image, and (4) a model training process for versatile capabilities. With the design, we also manually annotate **6K** high-quality samples for challenging graphical mathematical problems. Our trained model, CogCoM, equipped with this mechanism and 17B parameters, achieves SOTA performance across **9** benchmarks in **4** categories, demonstrating its effectiveness while maintaining interpretability. Code, model, and data are available at https://github.com/THUDM/CogCoM.

People are spending an enormous amount of time on dig-ital devices through graphical user interfaces (GUIs), e.g., computer or smartphone screens. Large language models (LLMs) such as ChatGPT can assist people in tasks like writing emails, but struggle to understand and interact with GUIs, thus limiting their potential to increase automation levels. In this paper, we introduce CogAgent, an 18-billion-parameter visual language model (VLM) specializing in GUI understanding and navigation. By utilizing both low-resolution and high-resolution image encoders, CogA-gent supports input at a resolution of1120 × 1120, enabling it to recognize tiny page elements and text. As a general-ist visual language model, CogAgent achieves the state of the art on five text-rich and four general VQA benchmarks, including VQAv2, OK- VQA, Text- Vqa, St- Vqa, ChartQA, infoVQA, DocVQA, MM-Vet, and POPE. CogAgent, using only screenshots as input, outperforms LLM-based methods that consume extracted HTML text on both PC and Android GUI navigation tasks-Mind2Web and AITW, ad-vancing the state of the art. The model and codes are available at https://github.com/THUDM/CogVLM.

We introduce CogVLM, a powerful open-source visual language foundation model. Different from the popular \emph{shallow alignment} method which maps image features into the input space of language model, CogVLM bridges the gap between the frozen pretrained language model and image encoder by a trainable visual expert module in the attention and FFN layers. As a result, CogVLM enables a deep fusion of vision language features without sacrificing any performance on NLP tasks. CogVLM-17B achieves state-of-the-art performance on 17 classic cross-modal benchmarks, including 1) image captioning datasets: NoCaps, Flicker30k, 2) VQA datasets: OKVQA, TextVQA, OCRVQA, ScienceQA, 3) LVLM benchmarks: MM-Vet, MMBench, SEED-Bench, LLaVABench, POPE, MMMU, MathVista, 4) visual grounding datasets: RefCOCO, RefCOCO+, RefCOCOg, Visual7W. Codes and checkpoints are available at Github.

Interpretation of predictions made by Convolutional Neural Networks (CNNs) is a rapidly growing field of research. A common approach involves enhancing semantic segmentation predictions through the generation of heatmaps that illustrate the significance of individual pixels in the segmentation. Nevertheless, the selection of beneficial features from these heatmaps remains a challenge. This is because the introduced information often contains interfering factors such as mutual features between different objects, background, and insufficient heat map resolution which often diminish its effectiveness. To overcome these limitations, we introduce Refined Weak Slices (RWS). Our main idea is to identify low attention regions in heat maps i.e.weak slices, in conjunction with segmentation accuracy, and utilize them to select effective features across different DNN layers, to enhance segmentation. We then seamlessly integrate these features back into the CNN, thusrefiningand enhancing the semantic segmentation result with selected features. Through extensive experiments, we demonstrate that incorporating the RWS module into state-of-the-art methods yields a notable improvement in the average mIoU by 2.84% on benchmark datasets (VOC 2012, COCOStuff, ADE20K, Cityscapes) for both ResNet-101 and ResNet-50 architectures. Furthermore, we achieve a maximum improvement of 5.8% with a single CNN. Overall, the combination of RWS and CNNs exhibits excellent performance in image segmentation tasks.

In recent years, pretrained models revolutionized the paradigm of natural language understanding (NLU), where we append a randomly initialized classification head after the pretrained backbone, e.g.BERT, and finetune the whole model.As the pretrained backbone makes a major contribution to the improvement, we naturally expect a good pretrained classification head can also benefit the training.However, the final-layer output of the backbone, i.e. the input of the classification head, will change greatly during finetuning, making the usual head-only pretraining (LP-FT) ineffective.In this paper, we find that parameter-efficient tuning makes a good classification head, with which we can simply replace the randomly initialized heads for a stable performance gain.Our experiments demonstrate that the classification head jointly pretrained with parameter-efficient tuning consistently improves the performance on 9 tasks in GLUE and SuperGLUE.* Equal contribution.Codes are at https://github. com/THUDM/Efficient-Head-Finetuning.

Heterogeneous graph neural networks (HGNNs) have been blossoming in recent years, but the unique data processing and evaluation setups used by each work obstruct a full understanding of their advancements. In this work, we present a systematical reproduction of 12 recent HGNNs by using their official codes, datasets, settings, and hyperparameters, revealing surprising findings about the progress of HGNNs. We find that the simple homogeneous GNNs, e.g., GCN and GAT, are largely underestimated due to improper settings. GAT with proper inputs can generally match or outperform all existing HGNNs across various scenarios. To facilitate robust and reproducible HGNN research, we construct the Heterogeneous Graph Benchmark (HGB) , consisting of 11 diverse datasets with three tasks. HGB standardizes the process of heterogeneous graph data splits, feature processing, and performance evaluation. Finally, we introduce a simple but very strong baseline Simple-HGN-which significantly outperforms all previous models on HGB-to accelerate the advancement of HGNNs in the future.

Multilingual knowledge graphs (KGs) such as DBpedia and YAGO contain structured knowledge of entities in several distinct languages, and they are useful resources for cross-lingual AI and NLP applications.Cross-lingual KG alignment is the task of matching entities with their counterparts in different languages, which is an important way to enrich the crosslingual links in multilingual KGs.In this paper, we propose a novel approach for crosslingual KG alignment via graph convolutional networks (GCNs).Given a set of pre-aligned entities, our approach trains GCNs to embed entities of each language into a unified vector space.Entity alignments are discovered based on the distances between entities in the embedding space.Embeddings can be learned from both the structural and attribute information of entities, and the results of structure embedding and attribute embedding are combined to get accurate alignments.In the experiments on aligning real multilingual KGs, our approach gets the best performance compared with other embedding-based KG alignment approaches.