PeopleofSuperintelligence

Researcher

Roman Soletskyi

Roman Soletskyi is an AI researcher whose work spans mathematical reasoning, neural-network verification, multimodal language models and speech generation. He undertook an ENS physics internship in 2022–2023, presented MATS research on neural-network verification in April 2024, and coauthored NuminaMath and Pixtral in 2024. In June 2025, he published research on variational-inference mode collapse and coauthored Magistral, Mistral AI's first reasoning model; in March 2026, he coauthored Voxtral TTS. As of October 2026, he is at Mistral AI.

Works
12
Citations
18
h-index
3
Badges

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

  1. 2022–2023

    A component-separation project

    In 2022–2023, at ENS's Department of Physics, he undertakes an internship with Erwan Allys on “From Point to Interval Estimates of Scattering Statistics in the Component Separation Task.”

  2. April 2024

    Verifying neural networks

    In April 2024, he presents his MATS project, “Verifying Global Properties of Neural Networks,” which explores model checking and control-barrier functions for safer reinforcement-learning agents.

  3. 2024

    The AIMO progress prize

    In 2024, he coauthors NuminaMath 7B TIR, a tool-integrated reasoning model that wins the first AI Math Olympiad Progress Prize with a score of 29/50 on the public and private test sets.

  4. 2024

    Pixtral 12B

    In 2024, he coauthors Pixtral 12B, a 12-billion-parameter multimodal language model trained to understand natural images and documents.

  5. June 2025

    Mode collapse in variational inference

    In June 2025, he publishes a theoretical study of mode collapse in variational inference with Marylou Gabrié and Bruno Loureiro; it analyzes gradient flow on Gaussian-mixture models and identifies mean alignment and vanishing weight as drivers.

  6. June 2025

    Magistral reasoning model

    In June 2025, he coauthors Magistral, Mistral AI's first reasoning model, whose paper details a scalable reinforcement-learning pipeline for large language models.

  7. March 2026

    Voxtral TTS

    In March 2026, he coauthors Voxtral TTS, an expressive multilingual text-to-speech model that generates natural speech from as little as three seconds of reference audio.

  8. As of October 2026

    At Mistral AI

    He completes an M2 in 2024, co-supervised with Marylou Gabrié, and is at Mistral AI.

Career

  1. NowMistral AI
  2. 2025–2026CNRS
  3. 2025–2026Université Paris Cité
  4. 2025–2026Sorbonne University
  5. 2025–2026Université Paris 8
  6. 2024–2026Université Paris Sciences et Lettres
  7. 2022–2023École Normale SupérieureIntern

From the biography and OpenAlex affiliations (the years of papers that list the organization), so some years are approximate.

Most-cited works

  1. Pixtral 12B2024 · 8 citationsWith Alexandre Sablayrolles, Amélie Héliou, Diego de las Casas, Emma Bou Hanna, Nikhil Raghuraman and Pavankumar Reddy Muddireddy
  2. Ministral 32026 · 4 citationsWith Abhinav Rastogi, Alexander H. Liu, Alexandre Sablayrolles, Amélie Héliou, Andy Ehrenberg and Andy Lo
  3. Magistral2025 · 3 citationsWith Alexander H. Liu, Alexandre Sablayrolles, Amélie Héliou, Andy Ehrenberg, Baptiste Bout and Baudouin De Monicault
  4. Bayesian imaging inverse problem with scattering transform2026 · Astronomy and Astrophysics · 1 citations
  5. A theoretical perspective on mode collapse in variational inference2025 · Machine Learning Science and Technology · 0 citations

From OpenAlex, as of 21 October 2026.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Roman Soletskyi. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes Mistral AI. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  3. Of 943 people shaping AI who worked at Google, 446 work there nowIncludes Mistral AI. Our estimates of where the 943 people shaping AI who worked at Google and Google DeepMind work now: 446 at Google and 452 across 278 other organizations.
  4. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staffIncludes Mistral AI. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staff. We test the estimate against 3,913 named model contributors and dated all-staff records across Google DeepMind, OpenAI and Anthropic, plus six additional frontier-grade labs.

Colleagues

People with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Roman Soletskyi yet. Send us a free-licensed photo.

  1. Internship 2022-2023 phys.ens.psl.eu. Checked 20 October 2026.
  2. Provable AI Safety seminars, April 2024 luma.com. Checked 20 October 2026.
  3. AI-MO/NuminaMath-7B-TIR · Hugging Face huggingface.co. Checked 20 October 2026.
  4. How NuminaMath Won the 1st AIMO Progress Prize huggingface.co. Checked 20 October 2026.
  5. Pixtral 12B arxiv.org. Checked 20 October 2026.
  6. A theoretical perspective on mode collapse in variational inference researchportal.ip-paris.fr. Checked 20 October 2026.
  7. Magistral arxiv.org. Checked 20 October 2026.
  8. Voxtral TTS arxiv.org. Checked 20 October 2026.
  9. group | Bruno Loureiro brloureiro.github.io. Checked 20 October 2026.