PeopleofSuperintelligence

Researcher

Josselin Somerville Roberts

Josselin Somerville Roberts is an AI researcher whose work spans climbing-robot perception, reinforcement learning, foundation-model evaluation, reasoning models and speech systems. In 2023, he coauthored separate work on crack detection for climbing robots and task-specific layers for multi-task reinforcement learning; his 2024 collaborations produced HELM Instruct, VHELM and Image2Struct. His 2025 papers covered reasoning and multimodal audio chat, and his 2026 research covered efficient dense language models, streaming speech recognition and multilingual speech generation.

Timeline

  1. 2023

    Rock-crack detection for climbing robots

    In 2023, he coauthors a method that detects rock cracks and edges to identify grasp sites for climbing robots. The paper introduces SKIL, a thin-object segmentation loss based on skeletons.

  2. 2023

    Projected Task-Specific Layers

    In 2023, he and Julia Di introduce Projected Task-Specific Layers (PTSL), a multi-task reinforcement-learning architecture that uses a common policy with dense task-specific corrections to represent shared and task-specific information.

  3. February 2024

    HELM Instruct evaluation

    In February 2024, he coauthors HELM Instruct, a multidimensional framework that evaluates instruction-following models with absolute ratings.

  4. May 2024

    VHELM vision-language evaluation

    In May 2024, he coauthors the first VHELM version, which extends HELM to evaluate vision-language models with new scenarios and models.

  5. 2024

    Image2Struct benchmark

    In 2024, he coauthors Image2Struct, a benchmark that evaluates how well vision-language models extract structure from images.

  6. June 2025

    Magistral reasoning models

    In June 2025, he coauthors Magistral, Mistral's first reasoning model. The paper presents Magistral Medium, trained for reasoning with reinforcement learning, and releases open-source Magistral Small.

  7. July 2025

    Voxtral multimodal audio

    In July 2025, he coauthors Voxtral Mini and Voxtral Small, multimodal audio chat models that process spoken audio and text documents.

  8. As of October 2026

    Ministral 3 and Voxtral systems

    As of October 2026, he is an AI researcher whose 2026 coauthored papers introduce the Ministral 3 series of parameter-efficient dense language models, Voxtral Realtime for streaming speech recognition, and Voxtral TTS for multilingual speech generation.

  1. Google had the most people shaping AI every year from 2014 to 2025Our estimates of which organizations employed the most people shaping AI each year from 1955 to 2025, tracing the shift from university labs to companies.
  2. The United States employs an estimated 48% of people shaping AIWe estimate where the people shaping AI work today, how universities, companies and research institutes split across countries, and how country shares have changed since the 1960s.
  3. 43% of living people shaping AI work outside their country of birthWe estimate that 43% of living people shaping AI work outside their country of birth (520 of 1,198 people with known birth and employment countries).
  4. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionWe estimate where the people shaping AI studied and where their careers led. Massachusetts Institute of Technology leads with 467 former students, including 291 recorded doctorates.

Sources and credits

Photo: We have no published photo of Josselin Somerville Roberts yet. Send us a free-licensed photo.

  1. A Skeleton-based Approach For Rock Crack Detection Towards A Climbing Robot Application arxiv.org. Checked 21 October 2026.
  2. Projected Task-Specific Layers for Multi-Task Reinforcement Learning arxiv.org. Checked 21 October 2026.
  3. HELM Instruct: A Multidimensional Instruction Following Evaluation Framework with Absolute Ratings crfm.stanford.edu. Checked 21 October 2026.
  4. The First Steps to Holistic Evaluation of Vision-Language Models crfm.stanford.edu. Checked 21 October 2026.
  5. Image2Struct: Benchmarking Structure Extraction for Vision-Language Models proceedings.neurips.cc. Checked 21 October 2026.
  6. Magistral arxiv.org. Checked 21 October 2026.
  7. Voxtral arxiv.org. Checked 21 October 2026.
  8. Ministral 3 arxiv.org. Checked 21 October 2026.
  9. Voxtral Realtime arxiv.org. Checked 21 October 2026.
  10. Voxtral TTS arxiv.org. Checked 21 October 2026.