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.

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.

Recent advances in diffusion models have demonstrated their remarkable potential in text generation, achieving performance that rivals or even exceeds autoregressive language models. However, existing diffusion language models face two critical challenges. First, they lack explicit mechanisms for semantic constraints, which is crucial for maintaining semantic consistency. Second, their non-autoregressive generation process at each timestep leads to insufficient token dependencies, resulting in semantic incompleteness or repetition in the generated text. To address these issues, we propose DiffuER (Diffusion with Embedding and Reconstruction loss). By utilizing word embeddings from a pre-trained masked language model as auxiliary regularization and leveraging an encoder-decoder module to reconstruct the original text from the generated sequence, DiffuER strengthens the model’s ability to capture semantics and maintain semantic consistency. Comprehensive experiments on four text generation tasks demonstrate DiffuER’s superiority over autoregressive, non-autoregressive models and diffusion-based baselines in terms of generation quality, diversity, and semantic consistency. Notably, our method achieves a 52.21 increase in generation diversity compared to autoregressive models, and a 38.7 decrease in perplexity compared to diffusion models. Our code is available at https://github.com/miaoGao/DiffuER.

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.

Code style transformation models built on code Language Models (code LMs) have achieved remarkable success. However, they typically focus on basis style transformations, where the target style follows a single criterion, and often struggle with combination styles, where the target style involves multiple criteria. In practice, style guides encompass multiple criteria, making the lack of effective combination style transformation a major limitation to their real-world applicability. In this paper, we propose Absent-Basis-Combination (abbreviated as ABC), a novel framework for code style transformation that significantly improves combination style transformation and overcomes the limitations of existing approaches. We implement four variants of ABC with parameter sizes of 0.5B, 1.3B, 1.5B, and 3B, demonstrating consistent superiority over existing approaches across all model sizes in both basis and combination style transformations. Specifically, ABC achieves performance gains of up to 86.70%, and remains superior even when baseline approaches use three times the parameters. Furthermore, to address the lack of high-quality datasets and evaluation metrics, we construct and release a new style transformation dataset, Basis & Combination Code Style (abbreviated as BCC Style ), and introduce Code Sequence, Syntactic, Semantic and Stylistic BLEU (abbreviated as CS4BLEU), a novel code similarity metric that surpasses existing metrics in accuracy and consistency.

Large language models (LLMs) achieve advanced capabilities in natural language processing (NLP) through pre-training on massive data, suggesting that LLMs have great potential to enhance human language learning. In second language (L2) acquisition, complex sentences pose significant challenges for learners due to limited working memory, and mastering grammar is crucial for overcoming these difficulties. However, syntactic parsing datasets utilized in NLP tasks are typically designed for computational purposes and are not ideal for pedagogical use. In this work, we propose a pipeline leveraging LLMs to construct a human-centered dataset for L2 acquisition, where complex sentences are visualized into a hierarchical tree diagram built from decomposed short sentences (i.e., clauses) within an interactive interface. Specifically, we employ a Least-to-Most (LtM) prompting strategy with seven mainstream LLMs to convert complex sentences into clause hierarchy with an F1 score of 0.8–0.9. Using this pipeline, we have constructed a "sentence-to-clausal-tree" dataset comprising 39,422 complex sentences from newswires and weblogs. Additionally, we offer a web-based application for visualization and user interaction. In a trial involving ten L2 learners, our web application has been highly acknowledged for enhancing comprehension and memory retention of complex English sentences.