AI Researcher

David Bau Portfolio

Assistant Professor of Computer Science at Northeastern Khoury College

Assistant Professor at Northeastern University (PhD, MIT) whose lab studies how deep networks internally represent and can be edited to change what they know; co-authored ROME (first author Kevin Meng), a method for locating and directly editing factual associations stored inside a language model, and is principal investigator and director of the $9M NSF-funded National Deep Inference Fabric (NDIF).

InterpretabilityModel EditingAI SafetyAcademic Research

Location

Boston, MA

Companies

Northeastern UniversityHarvard UniversityMIT CSAILGoogleMicrosoft

Credentials

  • Ph.D. in Electrical Engineering and Computer Science, MIT (2021)
  • M.S. in Computer Science, Cornell University
  • A.B. in Mathematics, Harvard College

Skills

InterpretabilityModel EditingAI SafetyAcademic Research

Notable

  • Sloan Research Fellowship (2025)
  • Principal Investigator and Director of the National Deep Inference Fabric, a $9M NSF-funded research infrastructure project
  • Co-author of "Locating and Editing Factual Associations in GPT" (ROME, NeurIPS 2022)
  • Co-author, with Lloyd N. Trefethen, of the textbook "Numerical Linear Algebra" (SIAM)

What makes it work

A breakdown of the choices that make this portfolio stand out.

1

A named method that other papers now build on

ROME (Locating and Editing Factual Associations in GPT, NeurIPS 2022) gave the field both a technique and a reusable name, and the lab's direct follow-ups (MEMIT for mass editing, then a Mamba version at COLM 2024) extend it; the paper has 3,900+ Google Scholar citations.

2

Public research infrastructure, not just papers

The National Deep Inference Fabric is shared, NSF-funded infrastructure for running interpretability experiments on large models — a contribution to the field's capacity to do research, not only a result within it.

3

A lab page that reads as a living research program

Listing multiple 2026 ICLR papers across distinct subtopics (symbolic in-context reasoning, list-filtering attention heads, theory-of-mind belief tracking) shows an active, multi-threaded lab rather than a single-topic research niche.

What AI Researchers can take from this

Specific, actionable tips to apply to your own portfolio — no generic advice.

  • Give a technique you develop a clear name — a named method is easier for the field to cite, extend, and build follow-up work on top of.

  • Consider building shared infrastructure (tools, compute access, datasets) for your subfield, not only publishing results that use infrastructure others built.

  • Keep several related research threads visibly active at once rather than one narrow topic — it signals a running lab, not a single finished project.

Projects

Nearly every project on David Bau's lab page gets its own subdomain (rome, memit, filter, algebra, belief), so the lab reads like a small library of named instruments rather than a list of citations.

ResearchadvancedMulti-year research program

ROME: Locating and Editing Factual Associations in GPT

PyTorchCausal TracingModel EditingGPT

Uses causal tracing to find where a language model stores a factual association, then edits those weights directly to change the fact.

A named method that other research builds on, which is the clearest sign of research impact.

  • First author Kevin Meng; Bau co-author

  • Causal tracing highlighted in a Nature news feature

  • Has its own project site

Open SourceadvancedOngoing infrastructure project

National Deep Inference Fabric (NDIF) and nnsight

PythonPyTorchHPCRemote Inference

NSF-funded infrastructure that gives researchers free remote access to large open models' internals, used through the nnsight library to read and alter activations and gradients.

Research infrastructure that expands what the whole field can study.

  • He is PI and Director

  • Runs on NCSA's Delta cluster

  • Published at ICLR 2025 as 'NNsight and NDIF'

ResearchadvancedMulti-year research program

LLMs Process Lists With General Filter Heads

TransformersAttention HeadsMechanistic Interpretability

Finds specialized attention heads that encode a filtering predicate abstractly enough to transfer the same filter to a new list, format, or language.

Shows a lab still producing new mechanistic findings, not only building on past work.

  • Lead author Arnab Sen Sharma

  • Explained with a programmer's filter-function analogy

  • Has its own project site

Ready to build your portfolio?

Follow the AI Researcher roadmap — skills, projects, and timeline to get hired.

AI Researcher Roadmap