Sparse MLP Neurons Shape Emergent Misalignment and Moral Choices
Editing small groups of native MLP neurons changes emergent misalignment and explicit moral choices.
Undergraduate Student
Li Yuejia
Hongyi Honor Class of Computer Science, Wuhan University
Sep. 2024 - Jun. 2028
Supervised by Prof. Mang Ye
I am an undergraduate student in the Hongyi Honor Class of Computer Science at Wuhan University. My research interests are representation learning, explainable AI (XAI), responsible AI, and scalable AI oversight. I study the internal mechanisms behind model behavior and how they can help identify safety-relevant failures.
About
I am interested in understanding AI systems: how internal representations shape behavior, why safety failures emerge, and how internal signals can support scalable oversight.
- Representation Learning
- Understanding AI Systems
- Explainable AI (XAI)
- Scalable AI Oversight
- Responsible AI
News
- Our paper “Threshold-Stable Selective Classification for Freedom of Information Act Privilege Review” was selected for an oral presentation at NLPCC 2026. 🎉
- Started a research internship in the Microsoft Research Asia Star of Tomorrow Internship Program.
- Our paper on cross-species animal re-identification was accepted to ECCV 2026. 🎉
- Released our preprint on spatially aware educational animation generation on arXiv.
Blog
Publications
Selected publications and preprints. View all publications →
Radioactive Feedback: Tracing Models Trained by Label-Only LLM Supervision
Encodes behavioral preferences in evaluator scores to trace models trained through label-only LLM supervision, enabling black-box detection without teacher-generated demonstrations.
Cross-Species Animal Re-Identification with Semantic Consistency Learning
Proposes Semantic Consistency Learning for animal re-identification across species, improving cross-species generalization under diverse morphologies and ecological domains.
See Before You Code: Learning Visual Priors for Spatially Aware Educational Animation Generation
Introduces OmniManim, a render-feedback-aware framework for generating educational animations with explicit visual planning, structured diagnostics, and localized repair.
Projects
Synesis
Co-building an AI learning workspace that decomposes problems, explains reasoning paths, and helps students connect worked examples to notes, plans, and follow-up practice.
Experience
Research Intern, Microsoft Research Asia (MSRA)
Completed the Microsoft Research Asia Star of Tomorrow Internship Program, working on language models and model alignment under the mentorship of Fangzhao Wu.
Undergraduate Student, Wuhan University
Coursework and project work across language models, educational AI, multimodal systems, and agentic tools.
Co-Founder, Synesis AI
Co-leading product design and prototyping for AI-assisted learning workflows and educational content generation.
Honors & Programs
NLPCC 2026 Conference Oral
Threshold-Stable Selective Classification for Freedom of Information Act Privilege Review
Microsoft Research Asia Star of Tomorrow Internship Program
Hongyi Honor Class of Computer Science, Wuhan University
Contact
I am open to research collaboration and academic conversations in understanding AI systems, representation learning, responsible AI, and scalable AI oversight.
Email: liyuejia@whu.edu.cn