Xiaochuan Li
- Affiliation:
- M.Sc. student
Human and Artificial Intelligence
University of Technology Nuremberg - E-mail:
- xiaochuan.li@utn.de
- Homepage:
- xchuan-li.github.io
- GitHub:
- github.com/xchuan-li
What is there in language itself that we have not yet discovered? Linguistics and cognitive science offer precise questions about meaning and decades of measured human behaviour. In language models, those questions become controlled experiments: training shows what can be learned from language alone, behavioural tests show what a model does, and interpretability shows where in the model it happens.
Education
Oct. 2025 – exp. 2027
University of Technology Nuremberg, Germany
M.Sc. Human and Artificial Intelligence
Designing and analysing behavioural experiments; cognitive psychology; training neural networks; interpretability methods for language models.
Oct. 2024 – Jul. 2025
University of Marburg, Germany
German for graduate study
Sep. 2020 – Jun. 2024
Shenzhen University, China
B.A. Philosophy
Formal logic and proof, including modal logic and Gödel’s incompleteness theorems; philosophy of language; psychology of reading.
Research
Apr. 2026 – Mar. 2027
Supervised at the German Research Center for Artificial Intelligence (DFKI)
Investigated how moral wording affects dilemma choices. Designed and ran controlled experiments and activation interventions in 7B and 14B models for the utilitarian-framing branch, with checks of answer labels, option positions, and task understanding; manuscript in preparation.
Sep. 2026 –
Independent project
Contrasted properties that depend on an added feature with properties that only co-occur with it (13 feature groups, 477 items, three instruction-tuned models), using predictors fixed outside the models: a dependency table, human feature norms and word-association norms; exploratory activation patching. Manuscript in preparation.
Oct. 2026 –
Independent project
A proposition on when speakers’ disagreement about word boundaries, a small cost to a listener, gives a text-only learner evidence about the scale, tested with language models trained from scratch on naming text (artificial language; World Color Survey). Preprint.
Sep. 2026 –
M.Sc. thesis, design stage · part of a PhD project at the Max Planck Institute for Human Cognitive and Brain Sciences, Leipzig
Asks how preschool children come to distinguish possibility from necessity in language, and how this relates to their nonlinguistic reasoning about possibilities; a behavioural study; data collection has not started. It lays the groundwork for later work on modal language in language models.
Projects
Sep. 2026
Independent theoretical study. Distinguishes three conditions for typicality in referent choice and the limits of what a forced-choice task can establish. Li, X. (2026). Working paper. lingbuzz/010343 · doi:10.5281/zenodo.22855036
May – Aug. 2026
May – Jul. 2026
Team course project, Deep Learning. My part: CNN training, ablations, and performance analysis.
Nov. 2025 – Feb. 2026
Course study, Foundations in Psychology and Empirical Study Design. Between-subjects experiment on optional checking after ChatGPT or Wikipedia answers (49 participants, 47 analysed); co-designed the study, ran the data analysis and wrote the report.
Dec. 2023 – May 2024
Leibniz’s logical system and characteristica universalis
Undergraduate thesis · Shenzhen University. Analysed Leibniz’s metaphysics, logic and proposal for a universal symbolic language.
Employment
Apr. 2026 – present
University of Technology Nuremberg, Germany
IT Service Desk (part-time)
Delivered student IT onboarding and individual support; handled system installation, service reporting, and inventory.
Sep. – Dec. 2021
Shenzhen Learning Disorders Association, China
Internship · Learning assessment at Yucai No. 2 Primary School
Prepared character-reading materials, administered timed assessments, and analysed response data to support the assessment of learning difficulties; contributed to teaching items and individual learning plans.
Skills and languages
Language models
Training small models from scratch; controlled behavioural evaluation, minimal pairs; activation patching, low-rank analysis, probing, steering
Human data
Experimental design, PsychoPy/Pavlovia, JASP; testing children
Programming
Python (PyTorch, Hugging Face Transformers, pandas, scikit-learn), Git, LaTeX, Slurm on an HPC cluster
Languages
Mandarin Chinese (native), German (C1), English (C1), Japanese (learning)