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.

Email
ForrestGan1103@outlook.com
UNSW
z5616333@ad.unsw.edu.au
GitHub
forrest-gan2005

Research interests

Research

All papers are sole-authored.

  1. Effective learning rates and forgetting in fine-tuning

    In preparation · 2026

    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.

    η η θ 3θ ‖W‖/3 ‖W‖
    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‖.
  2. Group differences in LLM survey simulations

    In preparation · 2026

    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.

    People German respondents +4.7 +2.1 +1.0 −0.1 −1.2 −2.6 Tülu 3 8B final r = 0.29 +2.3 +1.7 +1.6 +1.2 +0.7 +0.2 OLMo 2 7B final r = 0.20 +2.5 +1.7 +1.6 +1.4 +1.4 +1.0 OLMo 3 7B DPO r = 0.00 −0.3 −0.2 −0.1 −0.2 −0.2 −0.3 opposed in favour of mandates
    Simulated respondents miss a reversal. Red frowns: a vaccine mandate raised reactance; blue smiles: it lowered it.
  3. Do LLM assistants act on what they remember about users?

    Manuscript · 2026

    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

Education