| Home | Resume | Research | Interesting Reads | Old Books | Random |
|---|
You can find all my research papers on Google Scholar, and all of my AI safety work on the Alignment Forum. Read on here for a bit more commentary.
Hua, T.T., Qin, A., Marks, S. and Nanda, N., 2025. Steering Evaluation-Aware Language Models to Act Like They Are Deployed. arXiv link Alignment Forum link
My paper from MATS. I provide a pretty detailed breakdown on (what I believe to be) the strengths and weaknesses of the paper in the Alignment Forum post. I think more people should write critical self-reviews of their papers.
Qin, A., Hua, T.T., Marks, S. and Nanda, N., 2025. Discovering Backdoor Triggers. Alignment Forum Link.
Another project we did during MATS. We tried to answer this question: suppose we're worried about some specific behavior from a language model, can we find semantically coherent concepts that trigger the behavior? For example, if we're worried about a model launching rogue deployments, can we pre-fill text of the model doing this action (similar to coup probes) and then check if concepts like "not monitored" is upstream of this action? We had some promising results on toy models, but our methods were not successful in more realistic settings.
Hua, T.T., Baskerville, J., Lemoine, H., Hopman, M., Bhatt, A. and Tracy, T., 2025. Combining Cost-Constrained Runtime Monitors for AI Safety. arXiv link
A technical note on how to combine monitors with different costs and performance profiles. Check out the alignment forum post for a guide on how to use our method to construct a monitoring protocol from a probe monitor and a black box monitor. This was the first AI safety paper I wrote. It ended up being a somewhat econ-flavored paper, but I think the structure is appropriate for the problem we're trying to solve.
Thesis: Fox News's Effect on Social and Moral Preferences. Winner of the D.K. Smith Prize in Economics for best thesis.
Abstract: This paper examines how Fox News influences social and moral preferences: two crucial inputs in people's decision-making process. I conduct a survey among Americans aged 45 or older and use the variation in the channel positions of Fox News and MSNBC across different towns and cable providers as instruments. After confirming that these channel positions do not predict voting patterns before Fox started broadcasting, I find evidence that Fox shifted moral values to be more communal, some suggestive evidence that it decreased altruism and trust, and that Fox does not appear to affect negative reciprocity. In addition, these treatment effects are concentrated among those who did not vote for Bill Clinton or Bob Dole in the 1996 election.
I think this is a good paper, and I'm still proud of it. In retrospect though, I could've gone for something much more ambitious. I had this whole research agenda in mind exploring how media affects culture, which I define as a set of social preferences (e.g., how much to trust others, how much to value family members over strangers) and social constraints (I think economists don't seriously model social constraints and norms). I wrote this Fox News paper because it was a concrete example of how media could've changed people's preferences, there was an existing instrument from AER paper, and it's a very standard economics paper where I could showcase my economics skills for grad school.
However, instead of trying to write "a great economics paper," I should've just tried to do the real thing and come up with, maybe not a grand theory, but at the minimum a set of interesting observations on how culture interacts with media. The final paper will probably be a bunch of vignettes, with lots of qualitative stories, maybe a few quantitative regressions, and some theory. It would've been a weird thesis, but I would get to study the real, core thing I actually really wanted to study, and that would've been better for my soul. Alas, it is not so, but I'm still quite happy with getting a citation from Ben Enke :)
Axioms and Theorems in Voting Theory with a Brief Biography of Kenneth May.
I wrote this for the final paper of ECON 1080: Great Theorems of Economics taught by Professor Jerry Green. The class is about 2/3 microeconomic theory topics and 1/3 history of microeconomic theory. In this paper, I first give an overview of constructions in social choice theory, and then provide a proof of May's theorem. I then dive into a biography of Kenneth May, together with pictures of artifacts of his era (such as the fourth pamphlets from the Worker University of Paris, which he attended.) I also did a presentation and the slides can be found here.
Low Wage Gig Sector Increase Wages in Indivisible Labor Monopsony Labor Market
This is a short and sweet theory paper. The intuition is that the introduction of gig work offers flexibility on the number of hours people could work, which is something that people demand. In equilibrium, the arrival of gig opportunities should increase the wages of people who work in roles without a lot of flexibility. Specifically, their wages could increase even if (1) they don't do any gig work and (2) the per-hour wage in the gig industry is lower than their original wage, as long as their employers hold market power.
I wanted to add some data component to the paper to test my model in practice, but now I fear I might never have the time for it. The gist is that you can look at the professions where people are most likely to drive for a gig company, and then see if their wages increased proportional to the local market concentration (i.e., monopsony power of their employers, which people have measured using proprietary data here, although this article is supposed to have some market level measures available to the public). Then the specification uses data on uber/lyft entry. You could even run a triple differences by comparing those who had a 40 hour work week (i.e., probably not a lot of hours flexibility) with those who didn't.
Dobson, Emily, Carol Graham, Tim Hua, and Sergio Pinto. 2022. “Despair and Resilience in the US: Did the COVID Pandemic Worsen Mental Health Outcomes?” Working Paper 171. Brookings Global Working Paper Series. Brookings Institution. Brookings WP link
Started this while interning at Brookings. My first big and professional paper! Talks about (among other things) how the household pulse survey is a terrible survey (probably due to survey nonresponse bias) and you probably shouldn't use its data. I'm now much more skeptical of government surveys given the results from this paper.
Hua, Tian., Kim, Chris Chankyo, Zhang, Zihan., & Lyford, Alex. 2021. "COVID-19 Tweets of Governors and Health Experts: Deaths, Masks, and the Economy"
Journal of Student Research 10 (1). https://doi.org/10.47611/jsr.v10i1.1171
PDF Dataset
This is a descriptive study on Twitter discourse during the pandemic. We scrapped tweets from governors and health experts, and then compared the prevalence of certain words in them. We found that Democratic governors tweeted about deaths from COVID four times as often as Republican governors.
Me and my friends from Bellevue High School wrote a paper together during COVID about COVID because that's what friends are for (Professor Lyford helped us through the review process).