I'm an Assistant Professor at Institute of Science Tokyo (since October 2024).
I work on bridging the gap between symbolic AI and neural networks.
I'd like the knowledge AI learns to be something people can read, check, and correct. Right now, I'm building foundation models that find logical rules in new datasets without retraining, and studying how neural networks can learn when to backtrack during reasoning.
Classic rule learners start from scratch on every new dataset. I'm building models that instead learn how to find the rules behind the data.
NRI is trained once on randomly generated logic problems, then finds human-readable rules in new datasets without any retraining. G-NRI builds the symmetries of logic into the model, which lets it take on problems with over a thousand variables.
NRI · IJCAI 2026
G-NRI · NeSy 2026 (accepted)
Reasoning often means searching: you hit a dead end, go back, and try another path.
With my collaborators, I'm studying how neural networks can learn when to do this. We found that showing the model its current search state, separated from all the earlier steps, makes backtracking easier to learn.
Can Transformers Learn to Verify During Backtracking Search? · 2026 preprint
Read more »More research Selected publications
If you're working on something related, or just want to chat about an idea, feel free to email me at phua [at] comp.isct.ac.jp.