MB

Research

What else can
a model do?

I enjoy discovering capabilities hidden inside different models, then finding creative ways to use them to solve new problems. My work has explored pretrained tabular models, 3D reconstruction, and graph neural networks.

01 / First paperDeep learning / Tabular anomaly detection
Manuscript under review at ICLR 2027

Adapting prior-data fitted networks for tabular anomaly detection

Maximilian Bershtman & Niv Cohen

Faculty of Electrical and Computer Engineering
Technion - Israel Institute of Technology

Can a model pretrained for tabular prediction also learn to detect anomalies?

Niv Cohen and I developed ZEN and FOCUS, two methods that achieve state-of-the-art performance in our tabular anomaly detection benchmark evaluation. This paper studies how representations from TabPFN, a deep-learning prior-data fitted network, can be adapted for unsupervised tabular anomaly detection. The task is particularly challenging when the reference set contains the very anomalies the model needs to identify.

Read the paper Code & experiments
Results in our paper

Across 47 ADBench datasets and five random seeds, both methods outperform all 17 compared baselines in mean AUROC in both clean and contaminated reference-set settings.

Frozen representations

ZEN

Zero-training Embedding Neighbors uses features from a frozen model and nearest-neighbor distances to score anomalies. Trust weights reduce the influence of unusual reference samples.

Adapted representations

FOCUS

Fine-tuned One-Class Unsupervised Scoring adapts the model’s representations with an unsupervised compactness objective, then applies the same scoring approach.

The public repository includes the methods, benchmark results, and scripts for reproducing the reported tables and statistics.

02 / Research projectComputer vision / 3D understanding

Semantic probing

Looking inside
3D reconstruction.

With Yoni Kasten (NVIDIA)

Does a model that reconstructs a scene also learn what the objects in that scene are?

I investigated object-level semantics in the frozen latent representations of an iterative 3D reconstruction model. A lightweight probing head tested whether these features could support object segmentation across new viewpoints and unseen object instances.

The results indicated that the model learns a 3D-consistent semantic representation, extending what its reconstruction objective explicitly asks it to do.

Technical reportSemantic Probing of Iterative 3D Reconstruction Models: Extending Cut3r

03 / Earlier researchGraph neural networks / Scientific modeling

Learning from
molecular graphs.

I developed graph neural network models to predict micro-level geometric changes in molecular chemistry simulations. This work introduced me to research at the intersection of deep learning and a scientific domain.