Marco De Negri

Zürich, Switzerland

I’m an MSc student in Computer Science, majoring in Machine Intelligence, at ETH Zürich. Before that, I completed my Bachelor’s in Engineering of Computer Systems at Politecnico di Milano (110/110 with honors), with an exchange year at Universidad de Málaga.

I’m currently turning the protein language model ESM2 into a generative model for antibody design via Direct Preference Optimization, with the Learning & Adaptive Systems Group (Prof. Andreas Krause). With the ETH AI Center Apertus group, I trained linear probes on LLM hidden states that warn, token by token, before a model falls into a repetitive loop; this work, Degeneration Detection Is Not Early Warning, was accepted at the NeurIPS 2026 Workshop on Interpretability as a Science, and I also contributed to the Apertus v1.5 technical report as part of the Interpretability team. Earlier this year, with the Computer Vision and Geometry Lab (Daniel Barath), I trained a reinforcement learning agent to dynamically allocate resolution in 3D maps under a tight memory budget.

I’m broadly interested in machine learning for the natural sciences, language model interpretability and safety, and reinforcement learning. You can find more details on my publications, projects and resume pages.

Feel free to reach out if you’d like to chat about research, collaborate, or just say hi.

Contacts: mdenegri@ethz.ch | GitHub | LinkedIn

selected publications

NeurIPS Workshop
Degeneration Detection Is Not Early Warning

Marco De Negri, Luca Sartori, Lorenzo Baggi, and 5 more authors

In NeurIPS 2026 Workshop: Interpretability as a Science, 2026

Abs Bib

We train linear probes on the hidden states of Apertus-8B, Llama-3.1-8B and Mistral-7B to warn, token by token, that a model is about to fall into a repetitive loop, before it shows in the text. We propose ONWARD, an evaluation that fixes the false-alarm rate on healthy responses and scores alarms against a human-verified failure onset. We show that AUROC saturates on this task and cannot tell an early warning from a late detection, and that LoRA adaptation below the probe more than doubles warning coverage on all three models.
@inproceedings{denegri2026degeneration, title = {Degeneration Detection Is Not Early Warning}, author = {De Negri, Marco and Sartori, Luca and Baggi, Lorenzo and Reihs, Moritz and {\v{D}}urech, Eduard F. and El-Assady, Mennatallah and Pyatkin, Valentina and Hedstr{\"o}m, Anna}, booktitle = {NeurIPS 2026 Workshop: Interpretability as a Science}, year = {2026}, }

arXiv

From Open Text Models to Open Multimodal Reasoners: Apertus v1.5 Technical Report

Project Apertus

arXiv preprint arXiv:2610.15542. Contributor, Interpretability team , 2026

arXiv Bib

@article{apertus2026v15,
  title = {From Open Text Models to Open Multimodal Reasoners: Apertus v1.5 Technical Report},
  author = {{Project Apertus}},
  journal = {arXiv preprint arXiv:2610.15542},
  year = {2026},
}