Yu Wu · 吴隅
Representation learning · Semantics · Intellectual history
I study how knowledge is represented, learned, and transformed across AI systems and historical intellectual practices.
My research combines representation learning, semantics, and diagnostic evaluation with computational approaches to intellectual history.
I am particularly interested in how learned representations capture (or fail to capture) meaningful concepts, relations, and distinctions, and in how ideas are interpreted, contested, and transformed over historical time. More broadly, I am interested in what the comparison between artificial learning and historically situated practices of knowing can teach us about the nature of knowledge and ideas.
I study how AI systems represent complex concepts and relations, what semantic structure emerges in learned representations, and how it supports semantic comparison, retrieval, reasoning, and judgment. I am particularly interested in what these representations reveal about the structure and organization of learned knowledge.
Through diagnostic evaluation, I systematically analyze model capabilities and limitations, use these findings to probe learned representations, and investigate how models can be adapted to become more robust and responsive to different tasks and contexts. This work also contributes to broader questions of responsible AI and alignment, particularly the design of systems that are sensitive to cultural and historical context.
I study how ideas are interpreted, taken up, contested, adapted, and transformed over time. My current work develops computational approaches to identify intellectual engagement beyond generic semantic similarity, supporting close reading and substantive interpretation. This also raises broader questions about how computational methods can work with complex, heterogeneous, and context-dependent humanistic data.
Adapting computational methods to challenges in historical research at scaleBuilding on CASCADE's expertise in semantic change, PhD10 adapts computational methods for historical research at scale, tests their robustness against data and methodological confounders, and develops adaptable environments for new data and questions.
University of Helsinki
Provisional thesis topic: Search as Archival Discovery: Language Models for Eighteenth-Century Intellectual History
ShanghaiTech University
Thesis: Relational Matching in Vision and Language Based on Linguistic Structure
ShanghaiTech University
Thesis: Top-down Reasoning Policy for Referring Expression Grounding
A large collaborative benchmark for evaluating AI agents on long-horizon professional tasks, has been accepted to the Evaluations and Datasets track at NeurIPS 2026. I am glad to contribute to testing agent capability limits in using digital interfaces for historical knowledge work.
* Equal contribution
ContributionA large collaborative benchmark for evaluating AI agents on long-horizon, economically valuable tasks with verifiable outcomes.
ContributionA position paper on tracing increasingly indirect forms of text reuse and the reciprocal shaping of AI methods and historical research.
ContributionFinds substantially more implicit receptions than lexical baselines, while diagnosing a lexical gatekeeping effect that still constrains dense retrieval.
ContributionEstablishes a 724-page multimodal benchmark and shows that zero-shot models can detect Latin reliably without functionally understanding the language.
ContributionIntroduces image-conditioned caption correction, improving zero-shot vision-language reasoning without task-specific labelled data.
ContributionIntroduces a grounded matching task and a relation-sensitive reasoning network with stronger data efficiency and length generalisation.
Co-supervision of a master's thesis at the University of Helsinki.
Teaching assistance for the University of Helsinki course.
Co-leader of the Oral History team.
Across five semesters in Linear Algebra, Artificial Intelligence, Information Science and Technology, and Modern Poetry.