If you are looking for a tenacious individual who
- Dives deep into complexity & never cuts corners
- Navigates difficult/untrodden paths with surgical precision
- Does all of the above with as little ego as possible
You do not want to miss out on working with Ishaan Shrivastava.
Read my CV here - Curriculum Vitae
Read my blog posts here - My Posts
Work
I was part of the computer vision research team at Metafusion. The coolest thing I did here - dropped 3k lines of code over a weekend to remove a data pipeline bottleneck, which increased the throughput of my training runs 15x.
Interests
I love all things rationalism, AI Safety and mech interp. I also post on Lesswrong!
I am rated 1705 on codeforces, which puts me in the top 7% of codeforces competitive programmers, as well as being ranked #1 on TensorTonic, a website where you earn points by implementing machine learning algorithms.
Outside of work, I mess around with vision-language models to make them reason faster and better. I also read blogs on lesswrong dot com, meditate, go to the gym, and fold origami.
Publications
Learning the Power of “No”: Foundation Models with Negations [WACV 2025 | Citations: 51]
Jaisidh Singh*, Ishaan Shrivastava*, Richa Singh, Mayank Vatsa, Aparna Bharati
[arXiv] [GitHub] [Google Scholar]
Improved compositional reasoning in VLMs via CoN-CLIP, a proposed modification to CLIP’s contrastive learning objective. Also open-sourced CC-Neg, a novel unbiased dataset to benchmark whether models understand the absence of concepts implied by negation semantics such as no, not, and without in CLIP-like VLMs. I worked on this during my undergraduate in Artificial Intelligence and Data Science @Indian Institute of Technology, Jodhpur
Pet-Projects
Experiments with hybrid post-pre-norm and Per-Layer-Embeddings - I am on a mission to figure out the reason behind all the innovations jam-packed into the Gemma-3n (and, more recently, Gemma-4 series) models - Read my Gemma Blog!
Grokking, in JAX - I spend some time running experiments for this every day! GitHub - shrivastava95/jax-moe-ladder
Recurrent Transformer Architectures - check out my implementation of Hierarchical Reasoning Models (HRM) on GitHub.
SparseRNN - shrivastava95/sparsernn Implemented SparseRNN, an RNN with exponentially dilated temporal connectivity with subquadratic complexity, as an alternative to dense transformer decoder, beating RNN, LSTM, and naive transformer baseline on sequence classification and tagging tasks. Demo and Report linked in GitHub repo!
Heavy-Tailed Self-Regularisation theory - discovered a data-free method that prunes entire layers from VLMs, with minimal loss in performance. Works by finding the best power-law coefficient that fits the eigenspectrum of the weight matrices in a layer. Work in progress with Jaisidh Singh.
