Professor Bae Hong-kyun of the Department of Software Engineering at Kookmin University (President Jeong Seung Ryul) will present a paper at EMNLP 2026 (Conference on Empirical Methods in Natural Language Processing), the world’s largest conference in the field of natural language processing. A total of 17,669 papers were submitted to this conference from around the world, and only 2,719 of them—15.4%—were selected for the Main Conference. This research was conducted in collaboration with a research team led by Professor Kim Sang-wook of Hanyang University.

SAND, proposed by the research team, overturns a long-standing assumption of news recommendation systems. While existing technologies analyze the actual body of an article to match it with a user’s preferences, users have not yet read the body when they actually click on a news article. The starting point of this research is the observation that what drives a user’s decision to click is not the actual content of the article, but rather their own expectations based on the headline. To generate this “anticipated body,” the research team designed a structure in which three large language model (LLM) agents collaborate by dividing roles among themselves. The first agent summarizes the user’s interests based on past click history; the second agent uses this information to write out what that user might expect to read; and the third agent monitors the results and instructs the second agent to modify its writing style based on instances where the predictions were inaccurate. Since the text generated in this way differs in style from human-written articles, it could confuse recommendation models; however, the research team resolved this issue using a “dual-pathway distillation” technique, which transfers knowledge from a model trained on actual news articles through two separate pathways: one for user expressions and one for news expressions. In experiments using the English-language dataset MIND and the Norwegian dataset Adressa, SAND outperformed all 11 existing state-of-the-art methods. This research, titled “SAND: Subjective-Anticipation-Augmented News Recommendation via Dual-Pathway Distillation,” will be presented at EMNLP 2026, to be held in Budapest, Hungary, this November.

Professor Bae Hong-kyun, the principal investigator, stated, “While recommendation system research has traditionally focused on ‘what users have read,’ this study models the psychological state just before a click—specifically, ‘what users expected to read.’” He added, “This approach can be extended beyond news to various fields, such as e-commerce and video recommendations, where users must make choices without having fully verified the information.”