An engraved anatomical plate of the retina: rows of nerve cells drawn in stacked layers.

I study the layers inside neural networks and build the tools to see them.

Alan Zeng

I’m a student researcher in machine learning and a software developer. Most of my work sits where the two meet.

In research I work on interpretability: what each layer of a model represents, and when those representations change during training.

In software I build the instruments that kind of research needs. Small, fast, tested tools that other people can actually install.

This year I’m writing up a probing study, reading everything on sparse autoencoders, and looking for a research internship for summer 2027.

Layers are where a model builds its picture of the world, one step at a time. Why that matters to me.

Selected work:

Twelve projects and papers, 2023–2026. Each name opens its entry in the record below.

A 1957 computer room: operators at desks and tape units around an IBM 704.
The IBM 704 room at NASA Langley, 1957. NASA, public domain.

Research

Papers
  • When Features Settle: Tracking Representation Drift Across Training2026

    Alan Zeng. Preprint.

    PDFCopy BibTeX

  • Layer by Layer: Where Linear Probes Stop Working2025

    Alan Zeng and co-authors. Workshop paper.

    PDFCodeCopy BibTeX

  • Sparse Lenses for Small Transformers2024

    Alan Zeng and co-authors. Student research symposium.

    PosterCopy BibTeX

In progress
  • Counting Heads: how many attention heads does a fact need?2026

    Experiments running, write-up this winter.

  • Steamer: tiered scheduling for shared student GPUs2026

    Systems note, draft.

Software
  • Stratum2025

    Layer-by-layer probing for PyTorch models.

  • Tidewater2026

    Replays an agent’s tool calls one step at a time.

  • Footnote2026

    Checks whether a model’s citations say what it claims they say.

  • Steamer2025

    A job queue that shares a few GPUs among many students, in tiers.

  • Lantern2025

    Attention maps you can scrub through in the browser.

  • Halfstep2024

    An autodiff engine in 480 lines, written to teach backprop.

  • Kiln2024

    A small harness for running evals the same way twice.

  • Shelf2023

    A reading list that imports and exports BibTeX.

Talks
  • Reading a model one layer at a time2026

    Lab reading group.

  • Probing without fooling yourself2025

    Department machine learning seminar.

  • Building research tools people install2025

    Campus developer meetup.

Writing
  • What a probe can’t tell you2026
  • Notes from forty interpretability papers2025
  • Why I write the tool before the paper2025
Teaching
  • Teaching assistant, Introduction to Machine Learning2026
  • Teaching assistant, Data Structures2025
  • PyTorch from zero, a workshop for first-year students2024
Education
  • B.S. in Computer Science, minor in Mathematics2027

    Deep learning, optimization, probability, compilers.

Awards
  • Undergraduate research grant2025
  • Campus hackathon, first place2024

On layers

  1. Detail of an engraved anatomical plate: rows of retinal cells stacked in layers.
    The retina, drawn cell by cellSantiago Ramón y Cajal, 1894. Wellcome Collection.Biology found the layers first. Each row of cells in the eye hands the next a slightly more abstract picture of the world.
  2. An operator at the console of an IBM 704 while an engineer works inside an open cabinet.
    An IBM 704 at NASA, 1957NASA. Public domain.In 1958 Frank Rosenblatt ran the first perceptron as a simulation on this model of machine. One layer of weights, and the start of my field.
  3. Santiago Ramón y Cajal pointing at a bench of glassware while students in lab coats watch.
    Santiago Ramón y Cajal teachingPublic domain, via Wikimedia Commons.Cajal drew neurons by hand so that other people could see them. Good instruments for looking are still half of research.

Write to me about research, internships, or a tool you need.

I read every message.

GitHubGoogle ScholarLinkedIn