Samuel Marks

Career

Marks's Harvard page records a PhD from Harvard in May 2023 in mathematics, on p-adic Hodge theory and prismatic cohomology under Mark Kisin, before a shift to machine learning and the mechanistic interpretability of large language models. Marks then worked as a postdoc at Northeastern University under David Bau.⁵

At Anthropic, Marks is a Member of Technical Staff and leads the Cognitive Oversight subteam of the Alignment Science team. The MATS program lists Marks as a mentor for its Anthropic streams in 2026 and 2027, with focus areas in AI control and monitoring, misalignment science, and interpretability.¹²

Research

The Cognitive Oversight subteam aims to oversee AI systems by checking whether anything is suspicious in the cognitive processes behind their behavior, such as detecting when a language model is lying, using both interpretability-based white-box methods and black-box interrogation of models. Marks's 2025 study on auditing language models for hidden objectives trained a model to exploit errors in RLHF reward models and tested whether auditing teams could uncover that objective.¹³⁴

Earlier interpretability papers by Marks include "The Geometry of Truth" with Max Tegmark, which studies how language models represent the truth of simple true/false statements, and "Sparse Feature Circuits", which introduces methods for finding and editing causal subnetworks of human-interpretable features. In 2026 Marks co-authored CHIVE, an agentic pipeline that finds unexpected model behaviors and explains them with counterfactual prompt edits.⁸⁹²

Open-source work

Marks is a co-author of the paper introducing NNsight, an open-source system that extends PyTorch with deferred remote execution, and the National Deep Inference Fabric, an inference service that runs NNsight requests on shared GPUs and pretrained models.¹⁰⁶

Work and contributions

• Auditing language models for hidden objectives Published research

  First author and core research contributor of the 2025 study, which trains a language model with a hidden objective, runs a blind auditing game in which four teams investigate it, and then compares eight auditing techniques.³⁴

• NNsight and NDIF Open-source code

  Listed as a co-author in the nnsight repository's citation for "NNsight and NDIF: Democratizing Access to Foundation Model Internals"; the repository is released under the MIT license.⁶

• Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback Published research

  Co-author of the 2023 survey of open problems and limitations of RLHF and related methods.⁷⁵

Timeline

1. May 2023Completed a PhD in mathematics at Harvard, studying p-adic Hodge theory under Mark Kisin.⁵
2. Jul 27, 2023Co-authored a survey of open problems and fundamental limitations of RLHF.⁷
3. Oct 10, 2023Released "The Geometry of Truth" with Max Tegmark, on how language models represent true and false statements.⁸
4. Mar 28, 2024Released "Sparse Feature Circuits" as first author, with co-authors including David Bau.⁹
5. Jul 18, 2024Co-authored the paper introducing NNsight and the National Deep Inference Fabric.¹⁰
6. Mar 14, 2025Released "Auditing language models for hidden objectives" as first author, with work conducted at Anthropic.³⁴
7. Aug 21, 2026Co-authored the CHIVE study on explaining language-model behavior with counterfactual experiments.²

Other documented roles

Roles the sources document that are not confirmed as current.

• Postdoc under David Bau, Northeastern University⁵

Where they work

• Anthropic 34 people in the wiki

Related people

• Arthur ConmyAlso at Anthropic
• Neel NandaAlso at Anthropic
• Chris OlahAlso at Anthropic
• Evan HubingerAlso at Anthropic
• Collin BurnsAlso at Anthropic
• Kenneth LiInterpretability · Language models
• Yulia TsvetkovLanguage models · Safety and alignment
• Jacob SteinhardtSafety and alignment · Interpretability
• Sameer SinghInterpretability · Language models
• Aylin CaliskanSafety and alignment · Language models