Why I Struggle to Practice Good Research?
Published:
Nowadays, as AI grows increasingly capable and can write most research code, I have been pondering what kind of work I would truly enjoy doing.
My ideal research workflow would start with observations. From these observations and intuition, one forms conjectures. Building upon these conjectures and existing theory, one derives a method that actually works. Crucially, the method should deliver results not merely on toy models, but for more general cases. Through iterating on conjectures across this whole process, one’s intuition keeps getting refined.
In practice, however, I have found this ideal state extremely hard to attain. My research centers on large‑language models, a field largely driven by experimental science. This forces constant firefighting on countless engineering details. Very often, a method only works under special cases. Extending it requires massive debugging, and it is difficult to encapsulate everything within one coherent high‑level framework. Debugging itself is highly engineering‑heavy: much of the time it boils down to iterating with agents — you give instructions for revisions, run experiments, inspect outputs, and iterate further.
Given these realities, I often feel the work I produce is inherently imperfect, yet I cannot see a way around such imperfections. This explains my strong preference for parsimonious designs: my hypothesis is that simpler formulations are more likely to generalize.
On the other hand, I understand that academic work does not need to be perfectly polished. Academia’s greatest value lies in proposing novel ideas and sketching promising directions, while full‑fledged validation can be left to industry. Still, impact and community feedback matter greatly. Without any empirical evidence to back a proposal, there is little incentive for industry to adopt and test it.
This leaves me in frequent confusion: what constitutes good research, and how can I produce work that feels complete to my own standards?
For this reason, I am deeply grateful to Seed for giving me exposure to industry and the opportunity to empirically validate my proposed methods firsthand. Joining industry is the only viable path I can currently envision for this goal. If anyone has alternative perspectives, feel free to reach out via email.