Maximilian Augustin

Maximilian Augustin is an AI researcher at Mistral AI who co-authored Magistral, the company's first reasoning model, in 2025. With Matthias Hein at the University of Tübingen, he worked on making image classifiers robust and explainable, starting with the RATIO training procedure at ECCV 2020 and diffusion visual counterfactual explanations at NeurIPS 2022. His later papers with Hein built tools for finding model failures: the Spurious ImageNet dataset in 2023, DiG-IN at CVPR 2024 and DASH, a pipeline for systematic hallucinations in vision-language models, at ICCV 2025. He also co-authored the Voxtral audio chat models in 2025, the Ministral 3 language models in January 2026, and, in July 2026, the Robostral Navigate robot-navigation model and the Shieldstral safety classifier.

Now at: Mistral AI
Works: 11
Citations: 205
h-index: 5

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2020

RATIO at ECCV 2020

In 2020, with Alexander Meinke and Matthias Hein, he presents RATIO at ECCV, a training procedure that applies adversarial training to both in-distribution and out-of-distribution data. The resulting image classifiers give robust confidence estimates on out-of-distribution inputs and reach state-of-the-art l2-adversarial robustness on CIFAR10.
2022

Diffusion counterfactuals at NeurIPS

With Valentyn Boreiko, Francesco Croce and Matthias Hein, he co-authors Diffusion Visual Counterfactual Explanations at NeurIPS 2022. The method uses a diffusion process to make small, realistic image changes that flip the decision of arbitrary ImageNet classifiers.
2023

The Spurious ImageNet dataset

He co-authors Spurious Features Everywhere with Yannic Neuhaus, Valentyn Boreiko and Matthias Hein, presented at ICCV 2023. The paper builds a framework for finding harmful spurious features in ImageNet, introduces the Spurious ImageNet dataset and proposes SpuFix to reduce classifiers' reliance on those features.
2024

DiG-IN at CVPR 2024

With Yannic Neuhaus and Matthias Hein, he presents DiG-IN at CVPR 2024, a framework that generates images optimizing a classifier-derived objective. It produces visual counterfactual explanations, finds images where classifiers disagree, visualizes neurons and exposes systematic errors of zero-shot CLIP classifiers.
2025

DASH for VLM hallucinations

In 2025, he co-authors DASH with Yannic Neuhaus and Matthias Hein, an automatic, large-scale pipeline that finds systematic object hallucinations of vision-language models on real-world images. The team applies it to PaliGemma and two LLaVA-NeXT models across 380 object classes and presents it at ICCV 2025.
2025

Magistral at Mistral AI

He leaves Matthias Hein's machine learning group at the University of Tübingen in 2025 for Mistral AI. In June 2025, he co-authors Magistral, Mistral's first reasoning model, built on the company's own reinforcement learning pipeline; Magistral Medium is trained on top of Mistral Medium 3 with RL alone, and Magistral Small is open-sourced under Apache 2.0.
July 2025–January 2026

Voxtral and Ministral 3

In July 2025, he co-authors Voxtral, which presents the Voxtral Mini and Voxtral Small audio chat models that handle audio files up to 40 minutes long. In January 2026, he co-authors Ministral 3, a family of dense 3B, 8B and 14B language models derived through Cascade Distillation and released under Apache 2.0.
As of October 2026

Robostral Navigate and Shieldstral

As of October 2026, he works at Mistral AI. In July 2026, he co-authors Robostral Navigate, an 8B vision-language model that guides wheeled, legged and aerial robots from monocular RGB images, and Shieldstral, a 3B-parameter multimodal safety classifier that treats content moderation as a yes/no question-answering task.