Explorer first, researcher second, engineer third.

I’m deeply fascinated by intelligence and reasoning, both in narrow and general contexts. My core interests lie in reverse-engineering systems like Large Language Models (LLMs) and diffusion models, particularly exploring their non-trivial capabilities.

My research in Mech Interp follows a simple moto - “If a neural network can consistently show nontrivial abilites, I should be able to reverse engineer it”. If a network that satisfies the above two criteria is reading this, be wary! I am coming for you.

I iterate quickly with the goal to fail fast and adjust. I love learning things that fascinate me from ground up – currently focusing on proteins and robotics. I love hiking, writing, and playing tech mods on minecraft.

Currently, I’m pursuing an MS in Computer Science at UMass Amherst, working on my thesis in Mechanistic Interpretability under the guidance of Prof. David Jensen and Prof. Anna Green (Preliminary report coming soon!).

You can contact me at: nainani.jatin.0@gmail.com

What’s new about me?

• 10/01/24 ⮕ Got into Neel Nanda’s MATs Training Phase
• 09/05/24 ⮕ Completed summer internship at NVIDIA, working on using LLM agents for Chip Design and VLSI

Publications and In-Progress Work

Adaptive Circuit Behavior and Generalization in Mechanistic Interpretability
Jatin Nainani*, Sankaran Vaidyanathan*, AJ Yeung, Kartik Gupta, David Jensen
In preparation | 📄 arXiv

CS4: Measuring the Creativity of Large Language Models Automatically by Controlling the Number of Story-Writing Constraints
Anirudh Atmakuru*, Jatin Nainani*, Rohith Siddhartha Reddy Bheemreddy, Anirudh Lakkaraju, Zonghai Yao, Hamed Zamani, Haw-Shiuan Chang
In preparation | 📄 arXiv

Evaluating Brain-Inspired Modular Training in Automated Circuit Discovery for Mechanistic Interpretability
Jatin Nainani
📄 arXiv

Smartphone based tactile feedback system providing navigation and obstacle avoidance to the blind and visually impaired
Anish Pawar, Jatin Nainani, Priyanka Hotchandani, Gayatri Patil
Publised at IEEE ICAST 2022 | 📄 IEEE

* Denotes co-first authorship.