trace-paper-followups

Trace what later research did with a paper

The job: start from one seed work (a paper, model, method, or architecture) and produce an honest map of the research that came after it — separating real follow-ups (things that extend or improve the seed) from the large mass of papers that only cite it in passing.

This is a pipeline of cheap automated steps wrapped around one irreducibly model-driven step: reading and judging each candidate. Citation databases and keyword filters are noisy; the value you add is reading abstracts and deciding what each paper actually did relative to the seed. Don't try to automate that judgment away.

The pipeline

Work through these in order. Each script lives in scripts/ and is run with python3. Keep every stage's output in separate files so the lineage is traceable (raw citations → judged → prose), and pick a short <slug> for the seed (e.g. sam, mamba).

1. Identify the seed and read it first

Get the seed's arXiv id (or DOI / Semantic Scholar id). If you'll be making claims about what the seed is, read the seed paper before trusting how other papers describe it — secondhand summaries drift, and a comparison built on a wrong mental model of the seed is wrong everywhere. Download it with download_pdfs.py --ids <seed-id> and digest it with digest.py.

Then write a small <slug>_seed.json profile of the seed — its name, any aliases it's known by, arxiv id, a one-line what it is, and its baseline capabilities/modalities. This is what later questions are answered against: "did anyone add video to Moshi?" can only be answered honestly if the record says Moshi has no video to begin with. Shape:

{ "slug": "moshi", "name": "Moshi", "aliases": ["Moshi", "Mimi"],
  "arxiv": "2410.00037", "what": "Full-duplex speech-text model for real-time dialogue.",
  "capabilities": ["speech-in", "speech-out", "full-duplex", "real-time"],
  "modalities": ["audio", "text"],
  "code": "https://github.com/kyutai-labs/moshi" }

Include the seed's own reference implementation (code) if it has one — set it to null if there genuinely isn't one.