Today is marking the fourth week into the Spring 2025 semester at UTD. But I have not had a chance to share about my experience last semester. It’s better late than never, so let me tell you my journey last term.

The premise of Fall 2024 was my wholesome summer trip back home. From HCMC, I flew back to Dallas with a heart full of love.

Research

Research played a big part in my last semester. I had two separate projects going. One was led by my former labmate Jeongsik, another one by me. When Jeongisk was still in our lab, he used to be in charge of the computational experiments while I took care of the writing. However, now without any technical partner, I had to do the programming myself. Even though I was sad to say bye to Jeongsik, I was very happy to start doing something more technical.

I was basically writing training loops to train vision-language models. I started with LLaVA and then expanded on a few other SOTA VLMs. While digging through the code, I learned pretty deeply about HuggingFace library, to the point that I ran the Python debugger thought HF’s individual lines of code. Thanks to that, I got to know the best practices in the field on LLM training and inference.

I should mention that it was also very stressful, because I did not have all the time in the world to learn that. There were always a deadline approaching me in a few weeks. GPU time was so precious during that time that I had to stay up to around 3 AM to implement and test my model training code so that it could run while I went to bed. That is because model training takes hours to finish, and during that time, I can’t do anything useful about the model. By training it at night, when I wake up, I already have the results to analyze and write about.

During that time, especially in the August ARR submission, I learned a lot about paper writing because I wrote a new one. I learned to adjust my mindset to not fear of rejections, and to be really honest in reporting things.

Next, a fruit of my research work last school year has come when I was sponsored to attend EMNLP in Miami. To my surprise, it was not the type of conference experience I had expected. The NLP field currently has some problems and they manifest pretty obviously in the conference. Even though I tried to prepare as much as I could (even asking my advisor for conferencing advices), I still felt in shock with how things were there. I expected a bit more, partly because my previous conference experiences were quite good (SIGCSE24 and ICAPS24). But as my advisor said, it was my first NLP conference, so I should have just do whatever I want. The thing I am most grateful for during the trip was meeting many Vietnamese NLP researchers. Most of them are PhD students, some are industry researchers, while others are already professors. I felt like being home admist the LLM hurricane in academia.

Now I am about to take the QE to advance to PhD candidacy. This is also the time I start reflecting about my current reearch topic. I have been working on automated meme processing in the past 1.5 years. It started as a fun project that already existed in the lab. Now it is the only thing I am doing and may probably become my entire PhD topic. Memes are funny, but meme processing is not serious enough to be a research end goal in itself. I met meme people at EMNLP and I found that no one is really interested in memes alone. They all have some bigger problems in mind, such as multimodal hate speech, fact checking, or reasoning with LMMs. To me, vision-language reasoning feel most exciting, so I am probably look more into that.

I am so curious what would myself 5 years later from now think about my thoughts at this point in time. But regardless, I know everything is going to be fine.

Coursework

I took two organized courses last sem – ML and a seminar.

ML course by Vibhav Gogate was not too bad for a course I have taken for the fourth time in my life. He is quite technical, so his lecture has a lot of maths. Here are the related posts: NB vs LR, regularization priors, Bayes nets. I enjoyed asking him questions about unifying views of ML algorithms. Regarding the assignments, it feels quite unique, as if the prof created all the assignmens on his own, even for the coding templates.

It was also the first time that I took a seminar course. It was called something about advances in LLMs. We presented papers about LLM hot topics such as hallucination detection, submodular optimization, Reinforment Learning with Human Feedback, etc. This is probably my first time having substantial, collective, and intellectual discussions between CS PhD students at UTD.1

Lastly, I am still a TA. That’s a hard job for me, probably because I don’t like what it entails. But I am grateful for the assistantship because it is keeping financially afloat.

Life

About life, the highlight was my 3-week trip to California. I couch-surfed on my acquaintances’ couches the entire time. I met my high-school best friend who is now in UCLA and has had dinner with Terrence Tao. I met two of my former professors at Fulbright, who are called its founding faculty. They reminded me of the wholesome, liberal, optimistic, and self-motivating time in the early time of Fulbright. I also got to meet two others Fulbright alums in UCI and UCB. Hah, I collected quite a bit of experiences walking on UC-system campuses.

The most transforming experience during the trip was a one-week retreat at Deer Park Monastery, located on the hidden mountain of Escondido (near San Diego). It gave me perhaps the most peaceful feelings I have been having for years. I still remember the atmosphere where people were walking slowly and mindful of their breaths. No one was feeling the urge to talk even though they sit in the same table and look at each other’s eyes. There was nowhere to go, nothing to do. I can just suddenly pause my walk while people are around, close my eyes, and enjoy the breeze over my face. It was magical. It helps me to stop and look deeply into why I am here on Earth.

Anyway, back to the city of Dallas. It was not that bad, actually. I have figured out the way to live with the infrequent bus system. I went to downtown Dallas about twice last term just to stroll around and enjoy some urban vibe. Recently, I found out Downtown Plano also. It is the closest town to UTD that still gives a downtown vibe. I am happy that I got the opportunity to be in this peaceful city to focus on my work, while still have fun options available around me.

What’s next?

A recent Doist newsletter gave advice for new year planning. It basically says that one should spend the entire January to plan for the year and experiment with ideas. I have been doing that. I am actually not done yet, and I don’t mind spending until June to “figure out” that I want to do in this year.

My plan right now? Just to have lots of deep work, regular exercises, regular meditation, and a bit more reading and crafting.

To you, my readers: I am grateful for your readership to this unpolished blog. Sometimes you show up to me and let me know your comments on my writing, which I hugely appreciate. That is what keeps me writing. This blog is intellectually and spritually important to me. Thanks to your readership, I know that I have a place to share some cool things I learn during my study. Also thanks to you, I know the frustrations from my soul will be heard. Thank you.

May you be blooming beautifully this Spring season,

Khoi

  1. For some reason, the department is very hands-off about community building among PhD students. The people in charge are pretty much delegating the engagement responsibility to students ourselves. ↩