Arthur Conmy is a Member of Technical Staff at Anthropic; Conmy's site describes the role as working on aligning upcoming models as they are trained. From 2023 to 2026 Conmy was a Senior Research Engineer at Google DeepMind, working on post-training for Gemini and on interpretability tools including Gemma Scope, sparse autoencoders, probes, and model diffing. Earlier work on circuits includes first authorship of ACDC, a method for automated circuit discovery, and co-authorship of the indirect object identification circuit study in GPT-2 small.12345
At a glance
- Current roles
- Member of Technical Staff +1 more
- Company
- Anthropic
- Sector
- Company
- Work listed
- 3
- Sources cited
- 13
- Profile reviewed
Career
Conmy's site lists a software engineering internship at Meta in 2021 and research at Redwood Research in 2022–2023, followed by work as a Senior Research Engineer at Google DeepMind from 2023 to 2026. At Google DeepMind, Conmy worked on post-training for Gemini and on interpretability tools closer to production use, including probes, reward-model bias discovery, sparse autoencoders, model diffing and steering. Conmy has since joined Anthropic as a Member of Technical Staff, focused on alignment while models are being trained.12
Conmy has been a mentor in the MATS research program since MATS 6.0 and describes it as the main way Conmy mentors new researchers. The MATS mentor page lists Conmy's interests as automating interpretability, finding circuits and making model-internals techniques useful for AI safety.12
Research
Conmy's early work, in 2022–2023, was on circuits in small transformer models. Conmy co-authored the study of the indirect object identification circuit in GPT-2 small and was first author of ACDC, which automates the search for the model components behind a behavior. Related papers studied copy suppression in attention heads and successor heads, which increment tokens with a natural order such as days or months.4367
At Google DeepMind, Conmy co-authored the Gated Sparse Autoencoder paper and the Gemma Scope suite of open sparse autoencoders for Gemma 2. Later papers with MATS scholars examined unfaithful chain-of-thought reasoning, which reasoning steps matter in a trace, how thinking models differ from their base models, and automatic discovery of reward model biases. In January 2026 Conmy was the last-listed author of a paper on building activation probes for Gemini that hold up under production distribution shifts.85910111213
Work and contributions
Towards Automated Circuit Discovery for Mechanistic Interpretability (ACDC) Published research
First author of the 2023 paper, which systematizes the mechanistic interpretability workflow and automates the step of identifying the circuit of model components responsible for a behavior.31
Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2 Published research
Co-author of the 2024 Google DeepMind report introducing an open suite of JumpReLU sparse autoencoders trained on Gemma 2 models; Conmy's site lists Gemma Scope 1 and 2 among the Google DeepMind work.51
Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small Published research
Co-author of the 2022 paper explaining how GPT-2 small performs indirect object identification in terms of groups of attention heads.41
Timeline
- Co-authored the indirect object identification circuit paper on GPT-2 small.4
- Joined Google DeepMind as a Senior Research Engineer.1
- Posted ACDC, on automated circuit discovery for mechanistic interpretability, as first author.3
- Co-authored "Improving Dictionary Learning with Gated Sparse Autoencoders".8
- Co-authored the Gemma Scope report on open sparse autoencoders for Gemma 2.5
- Co-authored "Chain-of-Thought Reasoning In The Wild Is Not Always Faithful" as last author.9
- Co-authored "Thought Anchors", a study of which reasoning steps matter in language-model traces.10
- Co-authored "Building Production-Ready Probes For Gemini" as last author.13
Other documented roles
Roles the sources document that are not confirmed as current.
- Senior Research Engineer (2023–2026), Google DeepMind1
- Worked at Redwood Research (2022–2023), Redwood Research12
Where they work
- Anthropic 34 people in the wiki
Related people
- Samuel MarksAlso at Anthropic
- Amanda AskellAlso at Anthropic
- Collin BurnsAlso at Anthropic
- Ben MannAlso at Anthropic
- Jared KaplanAlso at Anthropic
- Kenneth LiInterpretability · Language models
- Eric Michael SmithSafety and alignment · Language models
- Jacob SteinhardtSafety and alignment · Interpretability
- Duen Horng "Polo" ChauInterpretability · Safety and alignment
- Sameer SinghInterpretability · Language models
Suggest a correction or removal
Profiles cover public professional work only, and each statement cites a source. If something is out of date, missing or wrong, tell us and link a source. People can also ask for their profile to be corrected or removed.
References
13 sources · profile reviewed Oct 21, 2026 · data updated Oct 21, 2026
- Arthur Conmy — homepage (role, past work)
- Arthur Conmy — MATS mentor page (Anthropic, Member of Technical Staff)
- Towards Automated Circuit Discovery for Mechanistic Interpretability (arXiv 2304.14997)
- Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small (arXiv 2211.00593)
- Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2 (arXiv 2408.05147)
- Copy Suppression: Comprehensively Understanding a Motif in Language Model Attention Heads (BlackboxNLP 2024)
- Successor Heads: Recurring, Interpretable Attention Heads In The Wild (ICLR 2024)
- Improving Dictionary Learning with Gated Sparse Autoencoders (arXiv 2404.16014)
- Chain-of-Thought Reasoning In The Wild Is Not Always Faithful (arXiv 2503.08679)
- Thought Anchors: Which LLM Reasoning Steps Matter? (arXiv 2506.19143)
- Base Models Know How to Reason, Thinking Models Learn When (arXiv 2510.07364)
- Automatically Finding Reward Model Biases (arXiv 2602.15222)
- Building Production-Ready Probes For Gemini (arXiv 2601.11516)