PeopleofSuperintelligence

Researcher

Johannes Welbl

Johannes Welbl is a Research Scientist at Google DeepMind who works on language models, reading comprehension, model toxicity, compute-optimal scaling and text watermarking. During his PhD in the Machine Reading Group at University College London, he assembled the SciQ science-question dataset in July 2017, published a multi-hop reading comprehension method in 2018, and studied neural reading comprehension undersensitivity in November 2020. In 2021, he was first author of "Challenges in Detoxifying Language Models" and coauthored the Gopher language model study. His March 2022 work as a coauthor of the Chinchilla paper on compute-optimal training received a NeurIPS 2022 Outstanding Paper award. In October 2024, he coauthored a Nature paper on the SynthID-Text watermarking scheme for language model outputs.

Badges

Timeline

  1. 2014–2016

    Random forests and complex embeddings

    In 2014, he publishes "Casting Random Forests as Artificial Neural Networks (and Profiting from it)" at GCPR. With Théo Trouillon, Sebastian Riedel, Éric Gaussier and Guillaume Bouchard, he coauthors "Complex Embeddings for Simple Link Prediction" at ICML in June 2016, using complex embeddings for symmetric and antisymmetric relations.

  2. 2017–2018

    SciQ and multi-hop datasets

    In July 2017, with Nelson F. Liu and Matt Gardner, he assembles SciQ, a crowdsourced set of 13.7K multiple choice science exam questions. Along with Pontus Stenetorp and Sebastian Riedel, he publishes a method for building multi-hop reading comprehension datasets in TACL in 2018 and creates two datasets based on Wikipedia and PubMed. He does this work as a PhD student in the Machine Reading Group at University College London, supervised by Riedel, Stenetorp and John Shawe-Taylor.

  3. November 2020

    Undersensitivity in reading comprehension

    Welbl is first author of "Undersensitivity in Neural Reading Comprehension" in November 2020, examining cases where meaningful changes to input text fail to change a model's prediction.

  4. 2021

    Toxicity study and Gopher

    In September 2021, he is first author of "Challenges in Detoxifying Language Models", a study with Amelia Glaese, Jonathan Uesato, Pushmeet Kohli, Po-Sen Huang and others. The study finds that basic toxicity-reduction strategies improve automatic metrics at the cost of reduced coverage for texts about, and dialects of, marginalized groups. In December 2021, he is also an author of the Gopher paper, which analyzes Transformer language models of up to 280 billion parameters.

  5. March 2022

    The Chinchilla compute-optimal study

    In March 2022, he is one of 22 authors of "Training Compute-Optimal Large Language Models", which finds that model size and training tokens should scale equally. The paper trains Chinchilla, a 70B-parameter model from the research team at Google DeepMind that uses the same compute budget as Gopher with four times more data. Chinchilla reaches 67.5% average accuracy on the MMLU benchmark.

  6. 2022

    A NeurIPS Outstanding Paper award

    NeurIPS 2022 gives an Outstanding Paper award to the Chinchilla paper, titled "An empirical analysis of compute-optimal large language model training" in the program. He is one of its 22 authors, alongside Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch and others.

  7. October 2024

    SynthID-Text watermarking in Nature

    Nature publishes "Scalable watermarking for identifying large language model outputs" in October 2024, with Welbl among the authors from Google DeepMind in London. The paper describes SynthID-Text, a production-ready text watermarking scheme tested in a live experiment on nearly 20 million Gemini responses.

  8. As of October 2026

    Research Scientist at Google DeepMind

    At Google DeepMind, he works as a Research Scientist with research interests in machine learning, natural language processing and artificial intelligence.

Career

  1. NowGoogle DeepMindResearch Scientist

Education

  1. 2018University College LondonPhD

From the biography.

  1. 16% of AI career moves in the 2020s went from a university to a companyIncludes Johannes Welbl. We estimate that 16% of AI career moves in the 2020s went from a university to a company; 29% of moves starting at a university went to a company, and 76% ended the university role.
  2. Google had the most people shaping AI every year from 2014 to 2025Includes Google. Our estimates of which organizations employed the most people shaping AI each year from 1955 to 2025, tracing the shift from university labs to companies.
  3. Google has the highest combined h-index among AI organizationsIncludes Google. We estimate that Google holds the highest combined h-index at 8,016, while universities account for 73% of the deduplicated total across AI organizations.
  4. 20% of people at companies who shape AI also hold a university role, led by Google and Stanford UniversityIncludes Google. We estimate that 539 of the 6,766 people with a known current organization (8%) hold both a university role and a company role, including 20% of people at companies and 13% of people at universities.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

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

  1. Johannes Welbl jowel.gitlab.io. Checked 22 October 2026.
  2. Complex Embeddings for Simple Link Prediction proceedings.mlr.press. Checked 22 October 2026.
  3. [1707.06209] Crowdsourcing Multiple Choice Science Questions arxiv.org. Checked 22 October 2026.
  4. [1710.06481] Constructing Datasets for Multi-hop Reading Comprehension Across Documents arxiv.org. Checked 22 October 2026.
  5. Constructing Datasets for Multi-hop Reading Comprehension Across Documents - ACL Anthology aclanthology.org. Checked 22 October 2026.
  6. Undersensitivity in Neural Reading Comprehension - ACL Anthology aclanthology.org. Checked 22 October 2026.
  7. [2109.07445] Challenges in Detoxifying Language Models arxiv.org. Checked 22 October 2026.
  8. [2112.11446] Scaling Language Models: Methods, Analysis & Insights from Training Gopher arxiv.org. Checked 22 October 2026.
  9. [2203.15556] Training Compute-Optimal Large Language Models arxiv.org. Checked 22 October 2026.
  10. Chinchilla (language model) - Wikipedia en.wikipedia.org. Checked 22 October 2026.
  11. NeurIPS 2022 Outstanding Paper Awards neurips.cc. Checked 22 October 2026.
  12. Scalable watermarking for identifying large language model outputs | Nature nature.com. Checked 22 October 2026.
  13. Johannes Welbl - Google Scholar scholar.google.com. Checked 22 October 2026.
  14. UCL NLP nlp.cs.ucl.ac.uk. Checked 22 October 2026.