Ian R. McKenzie
Lead author of the Inverse Scaling paper and of FAR.AI's STACK study of attacks on LLM safeguards
Ian R. McKenzie is first author of "Inverse Scaling: When Bigger Isn't Better," which reports tasks where language models get worse as they grow, drawn from the public Inverse Scaling Prize contest, and which released the winning datasets. McKenzie's papers from 2023 to 2025 list a FAR.AI affiliation, including "Scaling Trends in Language Model Robustness" and "STACK," a study that builds and red-teams an open-source safeguard pipeline for large language models.123456
Work
Inverse Scaling: When Bigger Isn't Better (external site: arxiv.org)Published research
First author of the paper, published in Transactions on Machine Learning Research, which analyzes 11 datasets from the Inverse Scaling Prize and releases the winning datasets.12
STACK: Adversarial Attacks on LLM Safeguard Pipelines (external site: arxiv.org)Published research
First author of the 2025 paper, which develops and red-teams an open-source defense pipeline and introduces a staged attack procedure against classifier-based safeguards.56
Scaling Trends in Language Model Robustness (external site: arxiv.org)Published research
Co-author, marked as an equal contributor with Nikolaus Howe, of the 2024 paper studying how attack and defense compute affect the robustness of language models as they scale.34
Other documented roles
Roles the sources document but that are not confirmed as current.
This profile covers public professional work only, from the sources below. It is not an endorsement or a ranking. To correct or remove it, use the correction form.