PeopleofSuperintelligence

Researcher

Ansh Radhakrishnan

Ansh Radhakrishnan is a researcher whose AI work examines language-model reasoning and scalable oversight. He earned a Statistics and Data Science degree in 2021 and coauthored work on reasoning faithfulness in 2023. His 2024 paper on model debate won an ICML Best Paper award. In November 2024, he coauthored a study of safeguards for untrusted language models; in May 2025, he coauthored research showing that reasoning traces often conceal models' use of hints. As of October 2026, he is a software engineer at Google.

Now at
Google
Works
7
Citations
126
h-index
6

Publication numbers from OpenAlex, as of 15 October 2026.

Timeline

  1. 2021

    A statistics degree

    He graduates with a Statistics and Data Science degree in 2021 after formal study focused on computer science and statistics.

  2. 2023

    Question decomposition

    He coauthors a 2023 paper that uses question decomposition to improve the faithfulness of language-model reasoning.

  3. November 2023

    Anthropic's debate research

    In November 2023, he publishes an update on Anthropic work in scalable oversight through AI safety via debate.

  4. 2024

    An ICML Best Paper

    In 2024, his coauthored debate paper wins an ICML Best Paper award. The study finds that debate helps non-expert models and humans answer questions.

  5. November 2024

    Adaptive deployment of untrusted LLMs

    He coauthors a study submitted in November 2024 on adaptive deployment of untrusted LLMs. The approach reduces backdoors by 80% at a given level of usefulness.

  6. May 2025

    Chain-of-thought faithfulness

    In May 2025, he coauthors a study of chain-of-thought faithfulness. The paper finds the reveal rate for prompt hints is often below 20%.

  7. As of October 2026

    Software engineer at Google

    As of October 2026, he is a software engineer at Google.

Sources and credits

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

  1. Inception anshradh.substack.com. Checked 15 October 2026.
  2. Question Decomposition Improves the Faithfulness of Model-Generated Reasoning arxiv.org. Checked 15 October 2026.
  3. Anthropic Fall 2023 Debate Progress Update alignmentforum.org. Checked 15 October 2026.
  4. Debating with More Persuasive LLMs Leads to More Truthful Answers proceedings.mlr.press. Checked 15 October 2026.
  5. ICML 2024 Awards icml.cc. Checked 15 October 2026.
  6. Adaptive Deployment of Untrusted LLMs Reduces Distributed Threats arxiv.org. Checked 15 October 2026.
  7. Reasoning Models Don't Always Say What They Think arxiv.org. Checked 15 October 2026.
  8. Ansh Radhakrishnan anshradhakrishnan.com. Checked 15 October 2026.