On the afternoon of October 23, 2026, the School of Foreign Languages at Shanghai Jiao Tong University hosted the first lecture of its "Language Intelligence Interdisciplinary Salon" series in Room 224 of the Yang Yongman Building. Yindong Wang, a doctoral researcher in the Artificial Intelligence and Intelligent Systems research group at the Hasso Plattner Institute (HPI) in Germany, delivered a talk titled "The Reliability of Large Language Models: From Hallucination Detection to the Mathematical Foundations of Generative Modeling." HPI, founded in 1998 by Hasso Plattner, co-founder of SAP, is a leading research institution in digital engineering and computer science based in Potsdam, Germany. The lecture was moderated by Assistant Professor Ercong Nie of the School of Foreign Languages and drew a large audience of faculty and students from within and beyond the School.
Nie Ercong opened the lecture by noting that the School has launched the "Language Intelligence Interdisciplinary Salon" series to create an interdisciplinary platform for exchange among academia, industry, and the international community—one that takes language as its core thread, bridging the language humanities and computational intelligence. Through this initiative, the School hopes to foster cross-fertilization between language science and frontier fields such as artificial intelligence.
Drawing on her own experience of learning multiple languages, Wang Yindong took the relationship between human intelligence and linguistic communication as her point of departure, organizing her presentation around three dimensions of language models—effectiveness, efficiency, and explainability. She began with few-shot learning to contrast generative and non-generative language models, and demonstrated how natural language inference (NLI) tasks can bring diverse NLP tasks, such as sentiment classification and semantic similarity, under a unified reasoning framework, thereby improving models' efficiency and generalization on new tasks.
Turning to the widely debated issue of "hallucination" in large language models, Wang Yindong carefully delineated how the term is defined differently in computer vision and in natural language processing. Pointing out that existing evaluations fall short in error localization and correction, she introduced the ReFACT dataset, constructed by her team from authentic online conversations, which enables fine-grained evaluation of large language models. She then presented her research on inference-time alignment: directly introducing natural-language suffixes during the generation process allows generation and safety judgment to proceed in tandem, without requiring a separate classifier. The method is expected to reduce extra computational overhead and unnecessary token generation, offering a new avenue for real-time, efficient safety alignment.
During the Q&A session, faculty and students delved into the underlying principles of large language models, the generalization of the suffix method, and how to surmount the mathematical hurdles of language intelligence research. The audience further debated whether model hallucination should be treated as a "bug" to be fixed or a "feature" resembling human memory mechanisms. Wang Yindong encouraged students from foreign-language backgrounds to keep an open perspective, harness the unique strengths of linguistic thinking, and seek out questions at interdisciplinary intersections.
As the inaugural event of the salon series, the lecture not only offered faculty and students of the School a window onto the latest advances in large language model reliability, but also further sparked their interest in interdisciplinary research on language intelligence. The School will continue to invite interdisciplinary researchers and industry experts to future sessions, sustaining the dialogue on frontier research and application scenarios at the intersection of large models and the language humanities.