Marcel Gröpl

Zürich, Switzerland

I’m a machine learning researcher working on multimodal and vision-language models and LLM agents. I’m also interested in efficient AI, such as pruning and making large models cheaper to run. I completed my MSc in Information Theory and Machine Learning at ETH Zürich in February 2026, after a BSc in Electrical Engineering at ETH.

My research interests include:

Multimodal LLMs (MLLMs)
LLM agents
Efficient AI: pruning (including token pruning for 2D/3D MLLMs) and faster, cheaper inference

My MSc thesis on Entropy-Gradient Grounding, a training-free way to retrieve visual evidence in VLMs, became a first-author paper at ECCV 2026. I’ve also worked as a research intern in industry: at the Bosch Computer Vision division in Yokohama, where I used SAM to improve pseudo-labels for panoptic segmentation, and at Sony in Tokyo, where I built NAS algorithms with LLMs as search agents. That work led to a co-authored paper at NeurIPS 2026.

Outside research, I co-founded Lexy AI, which builds LLM-based tools that generate legal documents, and I co-lead an LLM-bias study with ETH Juniors. I speak German, English, Japanese (JLPT N2), Polish, and some Korean. I’m always happy to chat about research or collaborations, so feel free to reach out.

news

Sep 24, 2026	Our paper Structuring Open-Ended NAS: Semi-Automated Design Knowledge Structuring with LLMs for Efficient Neural Architecture Search, from my internship at Sony, was accepted at NeurIPS 2026!
Jul 01, 2026	My first-author paper Entropy-Gradient Grounding: Training-Free Evidence Retrieval in Vision-Language Models was accepted at ECCV 2026!
Feb 28, 2026	I completed my MSc in Information Theory and Machine Learning at ETH Zürich.
May 01, 2025	I started a research internship at Sony in Tokyo, working on LLM-driven neural architecture search.
Mar 01, 2024	I joined the Computer Vision division at Bosch in Yokohama as a research intern.

selected publications

NeurIPS

Structuring Open-Ended NAS: Semi-Automated Design Knowledge Structuring with LLMs for Efficient Neural Architecture Search

Y. Sakuma, M. Yoshimura, Marcel Gröpl, and 4 more authors

In Advances in Neural Information Processing Systems (NeurIPS), 2026

arXiv Bib

@inproceedings{sakuma2026structuring,
  title = {Structuring Open-Ended NAS: Semi-Automated Design Knowledge Structuring with LLMs for Efficient Neural Architecture Search},
  author = {Sakuma, Y. and Yoshimura, M. and Gröpl, Marcel and Sun, Z. and Otsuka, J. and Irie, A. and Ohashi, T.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2026},
}

ECCV

Entropy-Gradient Grounding: Training-Free Evidence Retrieval in Vision-Language Models

Marcel Gröpl, Jaewoo Jung, Seungryong Kim, and 2 more authors

In European Conference on Computer Vision (ECCV), 2026

arXiv Bib Website

@inproceedings{groepl2026entropy,
  title = {Entropy-Gradient Grounding: Training-Free Evidence Retrieval in Vision-Language Models},
  author = {Gröpl, Marcel and Jung, Jaewoo and Kim, Seungryong and Pollefeys, Marc and Hong, Sunghwan},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year = {2026},
}