Varun Varma Thozhiyoor

tvarun@iisc.ac.in

I am a Masters student at IISc Bangalore, advised by Prof. Venkatesh Babu in the Department of Computational and Data Sciences. I am fortunate to also collaborate with Prof. Anand Bhattad. I completed my bachelors in Computer Science and Engineering from Govt. Model Engineering College, Kochi.

I am currently interested in evaluating and improving the physics understanding of video genenerative models and exploring their use as world models for embodied tasks.

[Google Scholar] [Linkedin] [X]

Recent News

[Sep 26] Principia has been accepted to NeurIPS 2026!!
[Sep 26] Principia is out! Read the arXiv submission.
[Jun 26] Attended CVPR 2026 to present Gravity-Eval.
[Apr 26] Presented Gravity-Eval and Principia at the IISc EECS Students Research Symposium 2026.
[Feb 26] Our work on Benchmarking Gravitational Physics in Video Generators is accepted at Findings of CVPR 2026!!
[Aug 24] Joined VAL.

Publications

2026

Principia: Relational Physics Tests for Video Models

TL;DR: We introduce a scale- and frame-rate-invariant benchmark for measuring physics understanding in video models through a paired object protocol that isolates physics dependencies.

Varun Varma Thozhiyoor, Shivam Tripathi, R. Venkatesh Babu, Anand Bhattad
NeurIPS 2026 Evaluations and Datasets Track
[project page] [paper]

Objects in Generated Videos Are Slower Than They Appear: Models Suffer Sub-Earth Gravity and Don't Know Galileo's Principle...for now

TL;DR: We introduce a benchmark for evaluating physics understanding in video models by measuring the relative consistency between identical objects under different initial conditions. Specifically we check whether models understand the relationship of height with time of flight by having two objects drop from different heights.

Varun Varma Thozhiyoor, Shivam Tripathi, R. Venkatesh Babu, Anand Bhattad
Findings of CVPR 2026
[project page] [paper]

Teaching

[Aug 26] Teaching Assistant, DS265o: Deep Learning for Visual Analytics
[Jan 26] Teaching Assistant, DS265: Deep Learning for Computer Vision

This website template is taken from here .