Rishabh Sharma

About

Hi! I'm Rishabh Sharma.

I'm a Simons Center Independent Postdoctoral Fellow at New York University, working at the intersection of AI and physics. My current research focuses on the physics of memorization and privacy in diffusion models, world models and their emergent neural representations, and mechanistic interpretability.

Previously, I completed my Ph.D. in Physics at the Tata Institute of Fundamental Research, Hyderabad, where my work on activity-induced transitions in amorphous solids was published in Nature Physics. I was awarded the 2025 Ramakrishna Cowsik Medal for this work.

Featured Research

Cyclic Denoising Reveals Ultrastable Memories in Diffusion Models

arXiv preprint (2026)

A fully unsupervised training-data extraction attack: repeated noising–denoising cycles drive generation into ultrastable attractors that decode to memorized training images — no training data, no prompts, a single GPU. The thumbnail is one such memorized image, recovered from Stable Diffusion.

movies arxiv

Cross-View World Models

ICLR 2026 Workshop on World Models

World models that predict future frames from any viewpoint given context from another, and in doing so develop a view-invariant spatial code — including the ring manifold for heading in the thumbnail, found by probing the internal representation layer by layer.

poster & animations paper arxiv

Activity-induced annealing leads to a ductile-to-brittle transition in amorphous solids

Nature Physics (2025)

Part of my Ph.D. work, where I explore how activity anneals glasses and discover a novel correspondence between active glasses and ordinary glasses under oscillatory shear. Featured in News & Views.

paper news & views

Positions

News

Publications

* denotes co-first authors

Awards & Fellowships

Professional Service

Selected Media