Arnav Dhiman

Research

Papers and patents across machine learning, biology, mathematics, and quantum computing.

Research Papers

Saddle Escape Efficiency: Introducing a Novel Metric and Revealing Its Criterion-Dependence in Non-Convex Landscapes

Accepted · Optimization

Flagship paper on a novel metric for how efficiently optimizers escape saddle points in non-convex landscapes, and why reported values depend on the chosen criterion.. Accepted ICCAI 2026, Okinawa, Japan. ICYS 2026 Bronze (oral presentation) and 2nd place (poster) · Indian Science and Engineering Fair Gold (Nationals and Regionals) · St. Yau Award shortlist (second round, in progress) · Genius Olympiad Finalist · CREST Gold Award · International Greenwich Olympiad Finalist

Interpretable SELFormer: Biologically Validated Gene Importance for Spatial Transcriptomics Reconstruction from scRNA-seq

Accepted · Computational Biology

Solo-authored work on interpretable spatial transcriptomics reconstruction from scRNA-seq with biologically validated gene importance.. Accepted IEEE iCGNIS 2026 (full paper), IEEE CIBCB (short paper). Solo-authored.

Why Do Single-Cell Foundation Models Lose to the Mean? A Diagnostic Study of Perturbation Prediction Failure Modes

Accepted · Computational Biology

Solo-authored diagnostic study of why single-cell foundation models fail to beat mean baselines on perturbation prediction.. Accepted IEEE CIBCB (short paper). Solo-authored.

Detecting Distributional Shift via Local Neuron Activation Rates

Under Review · Machine Learning Theory

Introduces Metabolic Neural Networks. Presented at Young Asians Conference 2026 (3rd place STEM research track).. Presented at Young Asians Conference 2026, 3rd place STEM research track. Under review at IEEE conference.

Do Foundation Models Outperform Lightweight CNNs in Breast Cancer Histopathology Classification?

Published · Artificial Intelligence

Comparison of foundation models and lightweight CNNs for breast cancer histopathology classification.. Published in IJFMR

Multimodal Deep Learning for Breast Cancer Subtype Classification Using Histopathology and Biomarker Data

Under Review · Artificial Intelligence

Multimodal deep learning for breast cancer subtype classification using histopathology and biomarker data.. Under review, IEEE. Preprint on Zenodo.

Analysing the Riemann Hypothesis and the Approach to Solve It Through Zeta Function Analysis and Quantum-Inspired Random Matrix Theory

Published · Mathematics

White paper on the Riemann Hypothesis through zeta function analysis and quantum-inspired random matrix theory.. White paper, published by One Young India

How Close Can a Chromatic Root Get to One?

Working Paper · Mathematics

Euler Circle Independent Research and Paper Writing (IRPW), under Dr. Simon Rubinstein-Salzedo. Graph theory, graph colourings.. Euler Circle Independent Research and Paper Writing (IRPW), under Dr. Simon Rubinstein-Salzedo

Entanglement Entropy Scaling and MPS Simulability of Quantum Algorithms

In Progress · Quantum Computing

BeyondQuantum research programme, mentored by David Chudozilov (Technical University of Munich). Presented at BeyondQuantum Research Fair.. BeyondQuantum research programme, mentored by David Chudozilov (Technical University of Munich). Presented at BeyondQuantum Research Fair.

Patents

Computer-Implemented System and Method for Efficient Machine Learning Training via Runtime Optimizer Switching

Filed

Indian Patent Application No. 202621067166, filed May 29, 2026

Indian Patent Application No. 202621067166, filed May 29, 2026. Extension of Saddle Escape Efficiency research.

Early Parkinson's Detection Device

Under filing

Currently under filing process for patentability

Low-cost, multimodal diagnostic device combining motion and olfactory sensing with on-device ML to detect Parkinson's years before motor symptoms. Backed by Emergent Ventures ($1,500 grant, Mercatus Center). Currently under filing process for patentability.