Data Science & AI · IIIT Naya Raipur

What resembles understanding in AI?

I'm Navyansh Singh, an undergraduate in Data Science & AI at IIIT Naya Raipur. My research is on language models and NLP, much of it with peer collaborators and some under faculty supervision. Most of it is self-directed: I choose the question, build the benchmark or the probe, and run the experiments.

I'll be applying to research-track MS/PhD programs for Fall 2028. Papers and preprints are below, and my CV and email are at the bottom; if something here overlaps with what you work on, I'd like to hear from you.

Research

What I work on

What I find interesting is which capabilities current models actually possess, especially on tasks that in humans require understanding and reasoning: what exactly the models do in those cases, and where and why they are limited.

I go after specific behaviours where a model drifts from what I assumed, or does something I don't have a clean explanation for, and then want the what and the why. I'm also getting into mechanistic interpretability, to check what is happening internally against the behaviour.

• Open questions about well-documented behaviours.

  A behaviour that is definitely there and well documented, but with a specific question about it I can't answer from what's published. Especially where behavioural similarity makes a clean separation difficult.

• Construct validity.

  What limits current benchmarks in measuring capability, and whether a precise issue is being hidden by the way they measure it. Fallacy detection was where I first found one.

• Where, exactly, the "understanding" runs out.

  When a model performs a task that needs something at least similar to reasoning or understanding: where does it fail, why, and can that limit be removed.

Publications

Papers & preprints

Published  ·  Accepted  ·  Preprint — status updated as papers move through review.

Get in touch

CV & contact

Navyansh Singh