arXiv Research Pulse

arXiv turns research attention into a measurable public time series. The useful question is not whether a topic is "good"; it is whether the public preprint stream changed enough to deserve a closer read.

Research questions

Question DeTime component Reader action
Which categories are structurally growing? trend prioritize a monthly field scan
Which categories are deadline-driven? season compare against conference calendars before writing
Which categories had unusual bursts? residual inspect titles and abstracts for the event behind the spike
Which topic deserves a column this week? trend score plus residual event rank open the query notebook and build a source card

Query protocol

The category notebook uses arXiv category queries such as:

Query Label Main caveat
cat:cs.AI artificial intelligence broad category with mixed subfields
cat:cs.LG machine learning high volume can hide smaller subtopic moves
cat:cs.CL language and LLMs language-model releases can cluster around deadlines
cat:cs.CV computer vision and video models conference submission cycles matter
cat:stat.ML statistical machine learning overlaps with cs.LG for some methods
cat:q-bio.QM quantitative biology methods smaller base rate makes residuals more volatile
cat:q-fin.ST statistical finance useful comparison category for trading examples
cat:econ.EM econometrics lower volume requires careful spike reading

The notebooks count monthly windows from the first day of each month to that month's end. Cross-listed papers can appear in more than one category view, so category charts should be read as field-specific public activity rather than a deduplicated paper inventory.

The agent-topic notebook uses phrase queries such as all:"AI agent", all:"coding agent", all:"tool use", and all:"agentic workflow". Query wording is part of the measurement. If a phrase changes, treat the resulting chart as a new series.

Deadline context

Submission volume often clusters around conference deadlines and camera-ready periods. A category spike should be read against the relevant calendar before it is described as a research shift. The safest article structure is:

  1. state the query and date window;
  2. show the trend, seasonal cadence, and residual event;
  3. inspect the top titles or abstracts behind the event;
  4. say whether the spike looks deadline-driven, launch-driven, or topic-driven.

Paper counts are publication activity. They are not peer-review quality, benchmark performance, or community consensus.

Non-core categories such as q-bio.QM, q-fin.ST, and econ.EM are included as comparison baselines. They make it easier to see whether an AI spike is field specific, cross-disciplinary, or part of a broader submission-cycle effect.

Notebooks

Use 01 arXiv category pulse for the category view and 02 arXiv agent research pulse for query-based agent topics.