PeopleofSuperintelligence

Researcher

Andrew Tang

Andrew Tang is a Columbia University computer science PhD student, advised by Vishal Misra and Dan Rubenstein, who studies how language models and agents use retrieved information. In 2022, they built MineDojo’s multimodal Minecraft Wiki and Reddit data pipeline, and the benchmark paper won a NeurIPS Outstanding Paper Award. They coauthored ClusterSC for AISTATS in May 2025 and worked on TokenProbe, a dashboard that streams token logits and probabilities, by June 2025. In September 2026, they coauthored an EMNLP paper on identity signals in retrieval. They are developing MIDLLMAT, a project that measures whether sensitive information can be rebuilt from less-sensitive documents.

Links
Website, Google Scholar, X, LinkedIn and GitHub

Timeline

  1. 2022

    A data pipeline for MineDojo

    In 2022, Tang coauthors the ten-author arXiv paper MineDojo, a Minecraft framework with thousands of open-ended tasks and an internet-scale knowledge base of videos, tutorials, wiki pages and forum discussions. They build its multimodal data pipeline for Minecraft Wiki and Reddit.

  2. 2022

    A NeurIPS outstanding paper award

    MineDojo wins the 2022 Outstanding Paper Award in the NeurIPS Datasets and Benchmarks track, and Tang is one of its ten authors.

  3. May 2025

    ClusterSC at AISTATS

    In May 2025, they coauthor ClusterSC with Saeyoung Rho, Noah Bergam, Rachel Cummings and Vishal Misra for the 28th International Conference on Artificial Intelligence and Statistics. The method clusters potential donors so each target uses only relevant units, provides theoretical guarantees and consistently outperforms classical synthetic control on synthetic and real-world datasets.

  4. June 2025

    Building TokenProbe

    By June 2025, they work on TokenProbe with Amy Wu, Rashfiqur Rahman, Charlie Kerfoot and Vishal Misra. The lightweight dashboard streams token logits, probabilities and entropy during inference to reveal when models learn concepts, experience mode collapse or forget.

  5. September 2026

    A paper on retrieval bias

    In September 2026, they coauthor the EMNLP 2026 camera-ready paper “Retrieval Sensitivity to Identity Signals in Queries” as first author, with Nicholas Deas, Kathleen McKeown and Vishal Misra. Across five dense retrievers and a sparse baseline, every retriever returns articles aligned with the query’s political lean and performs worse for African American Language questions than for White Mainstream English.

  6. September 2026

    MIDLLMAT and the mosaic effect

    In September 2026, they develop MIDLLMAT with Matthew Connelly, Siddhartha Dalal and Vishal Misra under IARPA BENGAL and NSF EAGER. The project measures whether sensitive facts can be rebuilt by aggregating individually less-sensitive documents.

  7. As of October 2026

    PhD student at Columbia

    As of October 2026, they are a Computer Science PhD student at Columbia University, advised by Vishal Misra and Dan Rubenstein. They study how LLMs and agents use retrieved information.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Columbia University. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. Nonprofit roles are 2.3 times as common in AI safetyIncludes Columbia University. We estimate where people focused on safety, alignment and evaluation work. Nonprofit roles appear for 25.1% of safety people with known affiliations, compared with 11.0% in other areas.
  3. Universities had a net outflow of 2,068 moves in AI careersIncludes Columbia University. Our estimates of career flows and median h-index across universities and companies in AI, including the role of joint appointments and earlier academic jobs.
  4. People shaping AI work at a median of 3 organizations over a career: 3 for university starters and institute and lab starters, versus 2 for company startersIncludes Columbia University. Across 6,853 people shaping AI with a dated work role, we estimate a median of 3 parent organizations over a career (mean 3.6), comparing role groups, starting sectors and starting decades from the 1950s to 2020–2026.

Sources and credits

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

  1. Andrew Tang andrewtcr.github.io. Checked 22 October 2026.
  2. MineDojo: Building Open-Ended Embodied Agents with Internet-Scale Knowledge arxiv.org. Checked 22 October 2026.
  3. Announcing the NeurIPS 2022 Awards – NeurIPS Blog blog.neurips.cc. Checked 22 October 2026.
  4. ClusterSC: Advancing Synthetic Control with Donor Selection proceedings.mlr.press. Checked 22 October 2026.
  5. ClusterSC: Advancing Synthetic Control with Donor Selection arxiv.org. Checked 22 October 2026.
  6. TokenProbe: Visualizing LLM Learning in Real-Time tokenprobe.cs.columbia.edu. Checked 22 October 2026.
  7. Retrieval Sensitivity to Identity Signals in Queries arxiv.org. Checked 22 October 2026.
  8. Vishal Misra - Students | Columbia University cs.columbia.edu. Checked 22 October 2026.