Huangsheng Gan
Undergraduate, UNSW Sydney. Independent researcher
I do independent research on how we measure language models: what they remember and actually use, how pretraining recipes shape what they learn and forget, and how faithfully they simulate people.
I’m a third-year undergraduate at UNSW Sydney. Much of my work looks for hidden flaws in standard evaluations and designs simple checks that remove them. I like careful studies whose analysis is planned before the results come in.
I’m open to collaborations and research opportunities. Feel free to reach out.
- GitHub
- forrest-gan2005
Research interests
-
Memory & personalization
Do assistants act on what they know about someone when it matters, and how can we tell that apart from advice that would suit anyone?
-
Training dynamics
How pretraining choices shape what a model can still learn and what it forgets, and how to compare models fairly.
-
LLMs as simulated people
When language models stand in for survey respondents, which human differences do they keep, flatten or reverse?
Research
All papers are sole-authored.
-
Effective learning rates and forgetting in fine-tuning
Pretraining choices such as weight decay change the overall scale of a model’s parameters, so the same fine-tuning learning rate amounts to a different effective step for each model. That can make one starting model look better than another for reasons unrelated to what it learned, and it tracks how much each one forgets.
Same update, different angle. A matrix followed by a normalization layer is scale-invariant, so an update of length η rotates it by an angle θ ≈ η/‖W‖. -
Group differences in LLM survey simulations
When language models answer surveys in place of people, they tend to erase differences in how groups respond to the same intervention. That matters because simulations are often used to see who will push back against a policy or message, not just how people react on average. Comparing models before and after post-training suggests that this stage does not create the pattern and does not reliably undo it.
Simulated respondents miss a reversal. Red frowns: a vaccine mandate raised reactance; blue smiles: it lowered it. -
Do LLM assistants act on what they remember about users?
Models that remember details about their users can repeat them back when asked, yet seldom act on them when a new request quietly calls for it. I suggest simple checks for whether a model is truly drawing on what it remembers, rather than giving advice that would suit anyone or simply reacting to a longer prompt.
Projects
-
Transformer language model from scratch
A byte-level BPE tokenizer, Transformer language model, AdamW optimizer and training loop implemented from the ground up in PyTorch.
-
DeepFM recommendation model
A DeepFM click-through-rate model (factorization machine plus deep MLP) trained in PyTorch on MovieLens-1M, exported with TorchScript and run by a C++ inference program built on LibTorch.
Education
-
UNSW Sydney
Bachelor of Computer Science, third year