Welcome!

Hi! My name is Xin Liang (梁馨). I recently graduated from Tongji University with a Bachelor of Engineering in Computer Science. I’m deeply engaged in generative modeling and multimodal learning, with an emphasis on reliable, scalable methods for large multimodal models and LLM-based agents. More specifically, I focus on:

Representation learning and interpretability in VLMs

I am very interested in making adapters transparent so that we can see which features drive task-relevant behavior. This led to STAN, an input-adaptive sparse adapter that selects a small set of latent units via top-k regularization and exposes unit specialization through activation statistics and visualization.

Simulation of synthetic-data training cycles and bias evaluation

I am interested in how synthetic data reshapes future training, bias, and robustness across generations. We designed a multi-generation simulation with subgroup construction and iterative retraining. We find that with cleaner data, stronger models, and balanced generation, bias recedes; with noisier data, weaker models, or imbalanced sampling, it stacks up and stability suffers.

Modular LLM agents

I work on LLM agents, with interests in planning, tool use, memory, and multi-agent coordination. We built SlideGen, a multi-agent visual-in-the-loop pipline that converts research papers into well-structured slide decks.

You can browse my projects and CV. Feel free to reach out via email.