Utkarsh Tiwari (he/him)

Hi! I'm a first-year PhD student at Toyota Technological Institute, Chicago in the University of Chicago. I'm advised by Prof. Will Merrill (check out our lab website). I graduated with a B.E. in Computer Science from BITS Pilani where I spent the best 4 years of my life (so far). I'm interested in ML theory, theoretical CS and their connections. In the past I've been at the STARS team at INRIA Sophia Antipolis, and at Prof. Manik Varma's lab at Microsoft Research, India where I also worked with Dr. Amit Deshpande .

I can be reached at utkarsh [at] ttic [dot] edu. All spelling errors are left in as proofs of human-generated text :)

Email  /  Resume  /  Blogs  /  Scholar  /  Github

profile photo

Research
Image description On the Hardness of “Infinite Energy” Puzzles
Utkarsh Tiwari , Utkarsh Thanvi, Armaan Sethi, Aalok Thakkar, Neeldhara Misra.
Under Review.
(Early results accepted at the JCDCG^3 conference)
Image description Modeling Tool Use in Transformers via Computation Oracles
Utkarsh Tiwari , Sai Soumya Nalli, Amit Deshpande, Michael Hahn.
Under Review.
(Early results accepted at the FLaNN workshop at Yale University, and Workshop on Latent & Implicit Thinking ICLR 2026)
Slides from presentation at LaCoCo lab,
Poster from FLaNN workshop, Yale University

Image description Improving Reachability on Reasoning Puzzles
Sukruta Prakash Midigeshi, Sai Soumya Nalli, Utkarsh Tiwari, Amit Deshpande, Nagarajan Natarajan, Vineeth N. Balasubramanian, Amit Sharma, Gaurav Sinha.
(Accepted at the Workshop on LLM Reasoning, ICLR 2026) OpenReview

Image description Emergent Stack Representations in Modeling Counter Languages Using Transformers
Utkarsh Tiwari, Aviral Gupta, Michael Hahn
ICLR, 2025 (Accepted at the World Models Workshop!)
NAACL, 2025 (Accepted at the Student Research Workshop!)
arXiv

Image description DebateBench: A Challenging Long Context Reasoning Benchmark For Large Language Models
Utkarsh Tiwari, Aryan Seth, Adi Mukherjee, Kaavya Mera, Kavish, Dhruv Kumar
NAACL, 2025 (Accepted at the Student Research Workshop!)

Image description What Matters in Autonomous Driving Anomaly Detection: A Weakly Supervised Horizon
Utkarsh Tiwari, Snehashis Majhi, Michal Balazia, François Brémond
ECCV, 2024 (Accepted with an Oral presentation at the ROAM workshop!)
arXiv / code

Results Under Preparation

1. Finding perfect matching cuts on induced subgraphs of the integer grid is NP Hard via reduction from Barnette (3-connected, cubic, planar, bipartite) graphs

2. Finding perfect matching cuts on hole-free induced subgraphs of the integer grid has a poly-time algorithm via maximum weight independent sets in outerstring graphs

3. Kingdomino style domino tiling is NP Hard for unbounded number of colors via reduction from graph decomposition


Website template from Jon Barron