Babak Ehteshami Bejnordi

Also known as B. Ehteshami Bejnordi, Babak Bejnordi, Babak E. Bejnordi

Babak Ehteshami Bejnordi is a machine-learning researcher whose work connects breast-cancer imaging with efficient AI models. He completed a PhD at Radboud University, organized the CAMELYON16 challenge, and studied breast-cancer biomarkers at Harvard. At Qualcomm AI Research, he led an Amsterdam team focused on efficient large language models and coauthored work on conditional computation, mobile mixture-of-experts inference, and on-device reasoning. In 2026, he coauthored Dirichlet-Prior Shaping, a method for specializing mixture-of-experts models that was accepted at ICML. Since September 2026, he is a Principal Research Scientist at NVIDIA, where he works on the Nemotron team and efficient language-model architectures.

Now at: NVIDIA
Born: 1986
Works: 55
Citations: 21,974
h-index: 23
Badges: Bridge

Years link to their sources. Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2013–2017

PhD work at Radboud

He completes a PhD in machine learning for breast-cancer diagnosis at Radboud University's Diagnostic Image Analysis Group in 2017. He develops machine-learning algorithms for breast-cancer diagnostics and organizes the CAMELYON16 metastasis-detection challenge. The challenge demonstrates that AI algorithms can outperform expert pathologists on a medical diagnostic task.
June 2016–November 2016

Visiting researcher at Harvard

From June to November 2016, he is a visiting researcher at Harvard University, where he studies tumor-associated stroma as a prognostic biomarker in breast cancer.
2018–2026

Leading Qualcomm AI research

From 2018 to 2026, he works at Qualcomm AI Research in Amsterdam and leads a team focused on efficient LLM architectures. He organizes the Qualcomm Innovation Fellowship Program in Europe from 2019 to 2023.
2019–2020

Batch-Shaping conditional networks

In 2019, he coauthors Batch-Shaping for Learning Conditional Channel Gated Networks, accepted at ICLR 2020. The method gates convolutional maps on or off according to network features, lowering dynamic computational cost.
2025

Cache-MoE for mobile inference

In 2025, he coauthors Mixture of Cache-Conditional Experts for Efficient Mobile Device Inference, which TMLR accepts in May 2025. Its cache-aware routing reuses experts during token generation to improve cache locality on memory-constrained devices.
March 2026

Reasoning on the edge

In March 2026, he coauthors Efficient Reasoning on the Edge, which trains small language models with LoRA adapters and supervised fine-tuning and applies reinforcement-learning budget forcing. It uses parallel test-time scaling and dynamic adapter switching for on-device reasoning.
May 2026

Dirichlet-Prior Shaping at ICML

In May 2026, he coauthors Dirichlet-Prior Shaping: Guiding Expert Specialization in Upcycled MoEs, accepted at ICML 2026. Its router regularization matches expert assignments to a target Dirichlet prior to improve expert specialization.
Since September 2026

Principal scientist at NVIDIA

Since September 2026, he is a Principal Research Scientist at NVIDIA on the Nemotron team in Zurich, where he develops efficient architectures for large language models.