Chiwei Zhu, Benfeng Xu, Mingxuan Du, Shaohan Wang, Xiaorui Wang, Zhendong Mao, Yongdong Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Web browsing—widely used for information retrieval and fact verification—has become a fundamental capability of recently emerged large language model (LLM) agents, which is often elicited by training on complex questions requiring web search. However, this task faces challenges with respect to data and training: existing QA datasets are mostly 1-3 hop over closed corpora (e.g., Wikipedia); meanwhile, outcome-based on-policy RL that used by recent works is inefficient and brittle in long-horizon, tool-heavy browsing environments. To address these challenges, we introduce GraphSynthQA, a knowledge-graph (KG)—guided synthesis framework in an open-web setting. Starting from Wikidata seed entities, GraphSynthQA iteratively retrieves and verifies evidence from the internet to expand a KG, then synthesizes complex, answer-verifiable queries grounded in multi-evidence dependencies. Building on the synthesized data, we train web-browsing agents with a compute-efficient two-stage recipe: (i) cold-start supervised fine-tuning on ReAct-style trajectories, and (ii) step-level Direct Preference Optimization (DPO), where preferences are constructed offline via single-step branched rollouts that contrast candidate actions by their downstream success rates, providing dense process supervision without expensive on-policy exploration. Experiments show that our approach consistently improves performance on challenging web-browsing benchmarks and remains competitive among models of similar size.

This paper introduces a novel approach to Generalized Category Discovery (GCD) by leveraging the concept of contextuality to enhance the identification and classification of categories in unlabeled datasets. Drawing inspiration from human cognition’s ability to recognize objects within their context, we propose a dual-context based method. Our model integrates two levels of contextuality: instance-level, where nearest-neighbor contexts are utilized for contrastive learning, and cluster-level, employing prototypical contrastive learning based on category prototypes. The integration of the contextual information effectively improves the feature learning and thereby the classification accuracy of all categories, which better deals with the real-world datasets. Different from the traditional semi-supervised and novel category discovery techniques, our model focuses on a more realistic and challenging scenario where both known and novel categories are present in the unlabeled data. Extensive experimental results on several benchmark data sets demonstrate that the proposed model outperforms the state-of-the-art. Code is availale at: https://github.com/Clarence-CV/Contexuality-GCD

Multi-source unsupervised domain adaptation (MUDA) has received increasing attention that leverages the knowledge from multiple relevant source domains with different distributions to improve the learning performance of the target domain. The most common approach for MUDA is to perform pairwise distribution alignment between the target and each source domain. However,existing methods usually treat each source domain identically in source-source and source-target alignment, which ignores the difference of multiple source domains and may lead to imperfect alignment. In addition, these methods often neglect the samples near the classification boundaries during adaptation process, resulting in misalignment of these samples. In this paper, we propose a new framework for MUDA, named Joint Alignment and Compactness Learning (JACL). We design an adaptive weighting network to automatically adjust the importance of marginal and conditional distribution alignment, and such weights are adopted to adaptively align each pair of source-target domains. We further propose to learn intra-class compact features for some target samples that lie in boundaries to reduce the domain shift. Extensive experiments demonstrate that our method can achieve remarkable results in three datasets (Digit-five, Office-31, and Office-Home) compared to recently strong baselines.