Rohan Asthana

I am a PhD researcher at FAU Erlangen-Nürnberg, working on structural and geometric signals for designing, evaluating, and diagnosing neural models. My research is unified by a single question: what do geometric and structural properties of neural networks reveal about how they learn and generalize?

Concretely, I have shown that anisotropy of the log-probability can detect and mitigate memorization in diffusion models 5–8× faster than prior art (ICLR 2026), that extrinsic curvature and SVD enable zero-shot neural architecture search without any labels (TMLR 2025), and that graph diffusion over architecture space generates valid neural architectures in under 0.2 seconds (TMLR 2024). Current work extends these geometric diagnostics to Vision Language Models (VLMs) and Vision Language Action (VLA) models.

I also collaborate with Nokia and Astrum IT on remaining useful life prediction of hardware components using pretrained time series foundation models, and have industry experience at BMW Group and Fraunhofer IKS. I am an official reviewer at TMLR and a sub-reviewer for NeurIPS, ICLR, and CVPR.

news

May 01, 2026	Paper on memorization detection in diffusion models via log-probability anisotropy accepted at ICLR 2026. 5–8× faster than prior methods with only 2 forward passes.
Aug 01, 2025	Dextr: Zero-shot NAS using SVD and extrinsic curvature, accepted at TMLR. No labels needed, only one unlabeled sample.
Mar 01, 2024	DiNAS: Conditional graph diffusion for neural architecture search, accepted at TMLR. Generates valid architectures in <0.2s across 6 benchmarks.

selected publications

ICLR

Detecting and Mitigating Memorization in Diffusion Models through Anisotropy of the Log-Probability

Rohan Asthana and Vasileios Belagiannis

In International Conference on Learning Representations, 2026

Abs HTML Code Website

We propose an anisotropy-based metric for detecting and mitigating memorization in diffusion models (e.g., Stable Diffusion). Our method is 5-8x faster than prior approaches and requires only 2 forward passes.
TMLR

Dextr: Zero-Shot Neural Architecture Search with Singular Value Decomposition and Extrinsic Curvature

Rohan Asthana, Joschua Conrad, Maurits Ortmanns, and 1 more author

Transactions on Machine Learning Research, Aug 2025

Abs HTML Code

We introduce Dextr, a zero-shot NAS proxy that uses singular value decomposition and extrinsic curvature to rank neural architectures without any labeled data, requiring only a single unlabeled sample.
TMLR

Multi-conditioned Graph Diffusion for Neural Architecture Search

Rohan Asthana, Joschua Conrad, Youssef Dawoud, and 2 more authors

Transactions on Machine Learning Research, Mar 2024

Abs HTML Code

We propose DiNAS, a conditional graph diffusion model capable of generating valid neural architectures in under 0.2 seconds, evaluated across 6 NAS benchmarks.