Artificial Scientific Intelligence Lab - Science of AI Team

Active

Research focus

We consider AI systems as scientific objects and try to develop theories and algorithms to enhance our understanding of AI systems, especially the representation of multimodal models.

Open to collaboration

Yes

Looking for

We are looking for collaborators in general mechanistic interpretability.

Contact

yiming@nus.edu.sg

Join the Damaqu community to see this team's full profile.

Papers on Damaqu (6)

When Language Overwrites Vision: Over-Alignment and Geometric Debiasing in Vision-Language Models
arxiv· 7 May 2026
Bridging Mechanistic Interpretability and Prompt Engineering with Gradient Ascent for Interpretable Persona Control
arxiv· 6 Jan 2026
A Unified Theory of Sparse Dictionary Learning in Mechanistic Interpretability: Piecewise Biconvexity and Spurious Minima
arxiv· 5 Dec 2025
How does My Model Fail? Automatic Identification and Interpretation of Physical Plausibility Failure Modes with Matryoshka Transcoders
arxiv· 13 Nov 2025
Rep2Text: Decoding Full Text from a Single LLM Token Representation
arxiv· 9 Nov 2025
Human-like Content Analysis for Generative AI with Language-Grounded Sparse Encoders
arxiv· 20 Aug 2025