Some of my thoughts through NeurIPS 2025 trip

Intro

My trip to NeurIPS was the first conference trip of my life, something that I would never forget. First time in my life I felt so connected with the academia society. Those researchers are not just some names on the paper, but a person with vivid souls.

This was most clear in the poster section. Instead of reading papers (filled with cover ups and selective presentations that help get the paper published rather than deliver genuine technical insights), I get to talk with authors face-to-face and get instant answers to my questions. These talks at the conference are more sincere, since everyone's paper got accepted and cover ups are no longer needed.

It was exhilarating, and I have to reflect this to a blog.

The spectrum of fields in AI

Current research on AI is so rich. It can be pictured into a spectrum, with two end being: Applied AI and Theoretical AI. Some works can feel so disconnected they might as well be different fields.

On the Applied AI end, most of the papers focus on LLMs and RL and agents, often in combination. Another popular topic is the embodied, real-world scenario application, like autonomous driving or robotics. Then we have the above applied in more specific scenarios, like the usage of LLM for coding, for math, or in a social-fairness point of view.

On the Theoretical AI end, which, to my surprise, was also strongly represented. Here, papers are purely theoretical, with dizzying math equations, focusing on boundaries, optimization, generalization capabilities, distribution shift, and statistical machine learning.

Of course, another majority lies between these two ends. I would place post-hoc explanation and NeuroAI in this part of the spectrum. The position on the spectrum are often determined by the size of their experiments. Some play with toy models with less than 5 layers, then they would focus more on the theoretical insights, thus be categorized into a near-theoretical position. Some play with LLMs, explaining part of the behavior of LLMs and LRMs, which would lean toward application part.

Remember my goal

This clarity breaks down into two commitments:

1. From the micro perspective, I must be energetic and effective, with a good work-life balance. My health and sustained creativity are the foundation of everything.

2. From the macro perspective, I have to remember that I want to advance human knowledge, even to the least amount. This is definitely not a easy goal. Publishing papers is not enough. I try not to be cynicism, but current farce of ICLR rebuttal proved that the over-noisy papers and reviews are not outlier points. Paper is just an advertisement for work, and a credential only if you have no better proof of capability. But what's the point if the work itself is hollow?

Therefore, while my game remains on paper-centric, my vision must be higher. The goal is quality of insight, not quantity of publications. I will play the necessary game, but I will not confuse the game for the goal.

So I say to myself, DO NOT GET LOST.

Keep my head up

I need to keep my head up for latest papers and works. This means deliberately making time to read papers, not just frontier works, but also other valuable works that have published years ago.

Explore widely and Exploit wisely. Stand on the shoulder of giants.

Stop EGO and work together

This point was delivered in an invited talk by Yejin Choi, and it hit me instantly. The current structure in academia is problematic. We're still largely operating with feudal relations of production—small labs and individual mentorship. This comes with pros and cons, but from a productive point of view, this is limiting. Papers are usually published with fewer than 10 authors, with a first author doing most of the work. This leads to atomic research in academia, where researchers do not communicate or cooperate.

It's like training a chess engine using Monte Carlo Tree Search: if we only stochastically explore all possible moves independently, the agent learns agonizingly slow. We need a way to pool our efforts, to focus collective time and energy on the most promising directions, which is almost always more rewarding.

That being said, it's still a utopian vision. It is fundamentally difficult to form large-scale, universal cooperation in the academic world, but it is possible, and even happening in industry. It will be interesting to see if, in commercially viable fields, industry's model will fully outpace academia's.

About my ambitious

From the life philosophical perspective, eventually what we need is to be 'happy'. The only different is how do we get happy. Lots of people would prize those who say: 'I just want a normal, quiet life', because this seems to be a good virtuous path to live a happy life. Before this conference, I counted myself among them.

But, I ask myself, is this what I really wanted? Do I really have no dream of becoming a scientist invited for a talk in NeurIPS? Do I really have no dream of getting my own work distinguished and contributes to the community? Do I really have no dream of working with brilliant people and do something game changing?

No. I have such dreams. But it was so far away from me back in school. I have to deal with meaningless paperwork and homework, lost in trivialities that slowly eroded my ambition. It was easy to lie to myself, to adopt the persona of someone who didn’t crave those heights.

But, through this trip, I've felt the spark. Now I must tend to it.