Baixuan Li, Jialong Wu, Wenbiao Yin, Kuan Li, Zhongwang Zhang, Huifeng Yin, Zhengwei Tao, Liwen Zhang, Pengjun Xie, Jingren Zhou, Yong Jiang, Wentao Zhang, Zhiqiang Gao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Retrieval-augmented generation (RAG) demonstrates remarkable performance across tasks in open-domain question-answering. However, traditional search engines may retrieve shallow content, limiting the ability of LLMs to handle complex, multi-layered information. To address this, we introduce WebWalkerQA, a benchmark designed to assess the ability of LLMs to perform web traversal. It evaluates the capacity of LLMs to traverse a website’s subpages to extract high-quality data systematically. We propose WebWalker, which is a multi-agent framework that mimics human-like web navigation through an explore-critic paradigm. Extensive experimental results show that WebWalkerQA is challenging and demonstrates the effectiveness of RAG combined with WebWalker, through this horizontal and vertical integration in real-world scenarios.

Zekun Xi, Wenbiao Yin, Jizhan Fang, Jialong Wu, Runnan Fang, Yong Jiang, Pengjun Xie, Fei Huang, Huajun Chen, Ningyu Zhang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Ding-Chu Zhang, Yida Zhao, Jialong Wu, Liwen Zhang, Baixuan Li, Wenbiao Yin, Yong Jiang, Yu-Feng Li, Kewei Tu, Pengjun Xie, Fei Huang. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Addressing intricate real-world problems necessitates in-depth information seeking and multi-step reasoning. Recent progress in agentic systems, exemplified by Deep Research, underscores the potential for autonomous multi-step research. In this work, we present a cohesive paradigm for building end-to-end agentic information seeking agents from a data-centric and training-stage perspective. Our approach consists of four key stages: (1) browsing data construction, (2) trajectories sampling, (3) supervised fine-tuning for effective cold start, and (4) reinforcement learning for enhanced generalisation. We instantiate this framework in a web agent based on the ReAct format, WebDancer. Empirical evaluations on the challenging GAIA and WebWalkerQA benchmarks demonstrate the strong performance of WebDancer, achieving considerable results and highlighting the efficacy of our training paradigm. Further analysis of agent training provides valuable insights and actionable, systematic pathways for developing more capable agentic models.

Chinese Spelling Check (CSC) aims to detect and correct Chinese spelling errors. Most Chinese spelling errors are the misuse of semantically, phonetically or graphically similar characters. Previous state-of-the-art works on the CSC task pursue transitions from misspelled sentences to correct sentences directly. However, it is difficult to force the current CSC methods to find the correct answer at one run. Thus, we propose a simple and effective method for CSC task by making fully use of the trained model to generate multiple candidate sentences and simply ranking to select the best, in which no additional training and parameters are required. The experimental results show that our approach outperforms previous methods and achieves the state-of-the-art performances.

Retrieval-based methods have been proven effective in many NLP tasks.Previous methods use representations from the pre-trained model for similarity search directly.However, the sentence representations from the pre-trained model like BERT perform poorly in retrieving semantically similar sentences, resulting in poor performance of the retrieval-based methods.In this paper, we propose KNN-EC, a simple and efficient non-parametric emotion classification (EC) method using nearest neighbor retrieval.We use BERT-whitening to get better sentence semantics, ensuring that nearest neighbor retrieval works.Meanwhile, BERTwhitening can also reduce memory storage of datastore and accelerate retrieval speed, solving the efficiency problem of the previous methods.KNN-EC average improves the pre-trained model by 1.17 F1-macro on two emotion classification datasets.

In sentiment analysis, it is concerned recently that a significant portion of subjective sentences across different domains do not contain explicit sentiment words but still convey clear subjective sentiment, which is known as implicit sentiment. However, currently available methods do not perform well on implicit sentiment analysis. In most cases implicit sentiment can be inferred from the context. To address this issue, we propose the ContextBert model, which employs implicit-sentiment-query attention to find the contextual information related to the target sentence to enhance implicit sentiment analysis. Experimental results show that our method achieves state-of-the-art performance on SMP2019-ECISA and EmoContext-Implicit (reconstructed based on EmoContext, to enrich the implicit sentiment analysis dataset).