PeopleofSuperintelligence

Researcher

Wilson Wu

Wilson Wu is an AI researcher whose work studies language-model planning, neural-network internals and data attribution. He studied future-token planning in 2024 and coauthored a compact-proof analysis of group-operation networks that appeared at ICLR 2025. In 2025, he coauthored a Bayesian influence method for data attribution and an ARC post about mechanistic estimation. His May 2026 work examined expected outputs of wide random MLPs and random-product expectations; he is a researcher at ARC as of October 2026.

Timeline

  1. 2024

    Myopic training and future tokens

    In April 2024, he coauthors “Do language models plan ahead for future tokens?” with John X. Morris and Lionel Levine. His affiliation is the Department of Mathematics at the University of Colorado Boulder; Morris and Levine are affiliated with Cornell University's computer science and mathematics departments. The paper tests pre-caching and breadcrumbs by training language models with myopic gradients that stop propagation to past timesteps.

  2. 2024–2025

    Compact proofs for group networks

    In October 2024, he coauthors “Towards a unified and verified understanding of group-operation networks” with Louis Jaburi, Jacob Drori and Jason Gross. The paper analyzes one-hidden-layer networks trained on finite group operations and turns its explanation into compact proofs of model performance; for the symmetric group S5, it gives an accuracy bound of at least 95% for 45% of the trained models. It appears at ICLR in 2025.

  3. September 2025

    Bayesian influence functions

    In September 2025, he coauthors the University of Colorado Boulder paper “Bayesian Influence Functions for Hessian-Free Data Attribution” with Philipp Alexander Kreer, Maxwell Adam, Zach Furman and Jesse Hoogland. It proposes the local Bayesian influence function, which replaces Hessian inversion with loss landscape statistics estimated by stochastic-gradient MCMC sampling.

  4. November 2025

    Competing with sampling

    In November 2025, he coauthors ARC's “Competing with sampling,” a post that introduces the matching sampling principle. The post reports progress on mechanistic estimates for random MLPs and trained two-layer MLPs.

  5. May 2026

    Estimating outputs of wide MLPs

    In May 2026, he coauthors ARC’s paper “Estimating the expected output of wide random MLPs more efficiently than sampling” with Victor Lecomte, Michael Winer, George Robinson, Jacob Hilton and Paul Christiano. It estimates randomly initialized MLP outputs over Gaussian inputs without running samples through the network. For sufficiently wide networks, the estimator reaches a target mean squared error with substantially fewer FLOPs than Monte Carlo sampling.

  6. May 2026

    Mechanistic estimates for random products

    In May 2026, he coauthors ARC's “Mechanistic estimation for expectations of random products” with Jacob Hilton, George Robinson, Eric Neyman, Paul Christiano, Michael Winer, Victor Lecomte and Gabriel Wu. The work covers random halfspace intersections, #3-SAT and random permanents; Wu writes up methods based on symmetric random functions and deduction–projection estimators for permanents.

  7. As of October 2026

    Researcher at ARC

    In April 2024, he is a full-time researcher at ARC. As of October 2026, he is a researcher at ARC, which works on a systematic and theoretically grounded approach to mechanistic interpretability.

Career

  1. Since 2024Alignment Research CenterResearcher
  2. 2024University of Colorado Boulder

From the biography.

  1. Direct Form 990 pay in the latest returns of 39 nonprofits with AI people on this site is highest at $2.1M at the Allen Institute for AI (Ai2) (2024)Includes Alignment Research Center. IRS Form 990 compensation across 39 nonprofits with AI people on this site from 2012 to 2025: in the latest returns, Mozilla reports $6.1M including related-company pay (2024) and the Allen Institute for AI (Ai2) leads direct nonprofit pay at $2.1M (2024).
  2. Google alumni have founded or led 133 other organizations, more than any other AI labIncludes Alignment Research Center. We estimate that alumni of 17 major AI labs went on to found or lead 350 other organizations, including 226 they founded, and we trace 472 career moves directly between 23 labs since 2010.
  3. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other universityIncludes University of Colorado. The Massachusetts Institute of Technology trained or employed 44 of the 187 AI major-prize winners, more than any other university. We compare doctorate and 5+ year employment links, career timing and current workplaces for 187 major-prize and 1,111 best-paper winners.
  4. More people shaping AI studied at Massachusetts Institute of Technology than at any other institutionIncludes University of Colorado. We estimate where the people shaping AI studied and where their careers led. Massachusetts Institute of Technology leads with 467 former students, including 291 recorded doctorates.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

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

  1. Do language models plan ahead for future tokens? arxiv.org. Checked 21 October 2026.
  2. Do Language Models Plan Ahead for Future Tokens? arxiv.org. Checked 21 October 2026.
  3. Towards a unified and verified understanding of group-operation networks arxiv.org. Checked 21 October 2026.
  4. Towards a Unified and Verified Understanding of Group-Operation Networks proceedings.iclr.cc. Checked 21 October 2026.
  5. Bayesian Influence Functions for Hessian-Free Data Attribution arxiv.org. Checked 21 October 2026.
  6. Bayesian Influence Functions for Hessian-Free Data Attribution arxiv.org. Checked 21 October 2026.
  7. Competing with sampling alignment.org. Checked 21 October 2026.
  8. Estimating the expected output of wide random MLPs more efficiently than sampling arxiv.org. Checked 21 October 2026.
  9. Mechanistic estimation for wide random MLPs alignment.org. Checked 21 October 2026.
  10. Mechanistic estimation for expectations of random products alignment.org. Checked 21 October 2026.
  11. Team — Alignment Research Center alignment.org. Checked 21 October 2026.
  12. Wilson Wu - MATS Mentor Program matsprogram.org. Checked 21 October 2026.