Harsh Singh

Hi! I’m an undergrad at Stanford studying CS (AI & Systems). I’ve worked on robotics, distributed training, and GPU infrastructure @ NVIDIA and edge inference @ AMD. I was also the first undergrad intern at Khosla Ventures.

Email  /  LinkedIn  /  GitHub  /  X

Harsh Singh by the water at sunset

Research

I'm motivated by Moravec’s paradox and the possibility of making physical intelligence as general and adaptable as we’re learning to make language models.

Figure 1: A donor rollout’s imagined future is inserted into a recipient rollout to test how it changes the robot’s action. How Do World Action Models Use Their Imagined Futures? Evidence from Mechanistic Interventions
Nicole H. Ma*, Harsh Singh*
* Equal contribution
NeurIPS Robot Learning Workshop, 2026 (Oral Presentation)

We use activation patching to reveal how robot policies use imagined futures for manipulation. By intervening on these future representations, we redirect robot actions and identify an attention pathway carrying their causal influence.

Figure 1: FlashAttention-2 and Delta Attention compared across two forward and five backward attention products; green cells show the sparse computation. Efficient Long-Context Continued Pretraining with Delta Attention
Harsh Singh, Jeffrey Willette, Krishna C Puvvada
Work done at NVIDIA.
NeurIPS Pre-to-Post & LCFM Workshops, 2026

We introduce a sparse attention method based on Delta Attention that controls how gradients flow through reused corrections, cutting update time by 23% while retaining long-context comprehension and near-ceiling retrieval accuracy.

Writing

Boothe Prize for Excellence in Writing
Stanford University, Program in Writing and Rhetoric · October 2025

3/1800 selected. Recognized for "Cognitive Cyborgism as a Zero-Sum Game," a 26-page paper that dissects human-AI collaboration through a game-theoretic lens.

Honors & Awards

1st Place Overall @ Act-athon (Physical Intelligence & Actor Labs) · Sep 2026

2nd Place ($5,000) @ Cognition x Etched x Mercor Inference-Time Compute Hackathon · Jun 2026

1st Place Scrapybara Award ($16,000) @ TreeHacks · Feb 2025

Regeneron STS Scholar · Jan 2024

ISEF Grand Award (Cellular & Molecular Biology) · May 2023

Outside research, I also enjoy modeling with Stanford FashionX.


Template from Jon Barron.