About me

I am a researcher on the interpretability team at Anthropic where I try to reverse engineer how language models compose microscopic building blocks into higher level computational circuits.

I completed my PhD at MIT where I worked with Dimitris Bertsimas, Max Tegmark, and Neel Nanda on interpretability. Before that, I was a software engineer at Google working on data pipelines for the storage analytics team, and researched fairness-optimized political redistricting with David Shmoys.

Selected Publications

See Google Scholar for a full and up-to-date list.

• Refusal in Language Models is Mediated by a Single Direction
  by Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, Neel Nanda
  Appeared at NeurIPS 2024 [arXiv]
• Not All Language Model Features are Linear
  by Joshua Engels, Eric J Michaud, Isaac Liao, Wes Gurnee, Max Tegmark
  Appearing at ICLR 2025 [arXiv]
• Confidence Regulation Neurons in Language Models
  by Alessandro Stolfo, Ben Wu, Wes Gurnee, Yonatan Belinkov, Xingyi Song, Mrinmaya Sachan, Neel Nanda
  Appeared at NeurIPS 2024 [arXiv]
• Universal Neurons in GPT2 Language Models
  by Wes Gurnee, Theo Horsley, Zifan Carl Guo, Tara Rezaei Kheirkhah, Qinyi Sun, Will Hathaway, Neel Nanda, Dimitris Bertsimas
  Published in TMLR [arXiv] [Twitter]
• Language Models Represent Space and Time
  by Wes Gurnee and Max Tegmark.
  Appeared at ICLR 2024 [arXiv] [Twitter]
• Finding Neurons in a Haystack: Case Studies with Sparse Probing
  by Wes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dimitrii Troitskii, and Dimitris Bertsimas
  Published in TMLR [Paper] [arXiv] [Twitter]
• Learning Sparse Nonlinear Dynamics via Mixed-Integer Optimization
  by Wes Gurnee and Dimitris Bertsimas.
  Published in Nonlinear Dynamics [Paper] [arXiv] [Twitter]
• Combatting gerrymandering with social choice: The design of multi-member districts
  by Nikhil Garg, Wes Gurnee, David Rothschild, David Shmoys
  Published in EC ‘22 [Paper] [arXiv] [Talk]
• Fairmandering: A column generation heuristic for fairness-optimized political districting
  by Wes Gurnee and David Shmoys
  Best paper award at SIAM ACDA ‘21 [Paper] [arXiv] [Talk]

Other Projects and Writing

• SAE reconstruction errors are (empirically) pathological (2024) - A preliminary research post on a potential issue with sparse autoencoder reconstructions.
• Inductive Biases of SGD Training (2022) - A review of inductive biases of stochastic gradient descent (SGD) when training deep neural network.
• Analytics for Health Security (2022) - An analytics enabled defense-in-depth strategy for health security.
• Optimal Political Districting: The Anchor Method (2022) - A formulation of optimal political districting using the anchor method.
• Fairmandering: Generating Fairness-optimized Political Districts (SIAM News; 2021)
• Scalable Approximation of k-medians for Political Districting (2020) - Using a linear programming relaxation to approximate the k-medians problem for political districting.