Tharun Kumar
Princeton University '26
I'm a researcher and engineer interested in training large models to reason about the world. I graduated from Princeton University in 2026, where I studied CS and math, and researched mech-interp, SSMs, and LLM pre-training under Professor Elad Hazan in the Google DeepMind @ Princeton lab.
Previously, I worked on state-of-the-art AI agents for Fortune 500 companies as a founding engineer at Universal AGI, increased availability of distributed systems at AWS, and researched generative video models at Hooglee under mentorship from Sebastian Thrun.
In my spare time, I'm usually drumming, playing volleyball, lifting weights, or making new things (ask me about how I hem my clothes).
I love meeting new people. Feel free to reach out if you have questions or just want to chat :)
Work
Working on something new.
Projects
- Emergent Compositional Skills in Mixture-of-Experts VLAsICML CompLearn '26A VLA with a simple Mixture-of-Experts action head learns to break manipulation tasks into reusable skills like grasp, transport, and release, with no skill labels or hierarchy. Accepted at the CompLearn Workshop at ICML 2026.
- Mechanistic Capability Probes as a Cheap Screen for New ArchitecturesICML MechInterp '26Can tiny 1M-parameter models trained on synthetic probes predict which sequence-mixer architecture wins at 1B scale? Our paper at the ICML 2026 Mechanistic Interpretability Workshop says yes.
- Learned Diffusion Priors to Optimize Gaussian SplattingNovel view synthesis in 3D reconstruction with sparse-view inputs, using learned priors from image generation models.