Fastest-growing research topics on arXiv (November 2026)
Key numbers
- 115,442 papers submitted to arXiv from 17 August 2026 to 15 November 2026, compared with 76,489 over the same weeks a year earlier.
- on-policy distillation: 393 papers, share of all papers ×20.0.
- agent harness: 271 papers, share of all papers ×13.8.
- coding agent: 735 papers, share of all papers ×10.4.
- Fastest-growing in computer science and AI: on-policy distillation (×20.4).
- Fastest-growing in astrophysics: LZ (×2.8).
- Fastest-growing in biology (bioRxiv): Xenium (×2.9).
Counted by SummarizePaper from the titles and abstracts of every new arXiv, bioRxiv and medRxiv paper. Latest 13 weeks: 17 August 2026 to 15 November 2026; compared with the same weeks a year earlier (18 August 2025 to 16 November 2025). Updated every month; last update: 20 November 2026.
What are the fastest-growing research topics on arXiv?
The terms whose share of all new arXiv papers grew most, comparing the latest 13 weeks with the same weeks a year earlier. "Share" is how many times more often papers use the term, once arXiv's own growth is taken out.
| # | Term | What it is | Papers, latest 13 weeks | Papers, a year earlier | Share |
|---|---|---|---|---|---|
| 1 | on-policy distillation | Knowledge transfer using student-generated trajectories | 393 | 0 | ×20.0 |
| 2 | agent harness | Infrastructure for running and managing agents | 271 | 0 | ×13.8 |
| 3 | coding agent | Software agent that performs programming tasks | 735 | 46 | ×10.4 |
| 4 | recursive self-improvement | Iterative improvement of a system by itself | 194 | 0 | ×9.9 |
| 5 | Qwen3 | Family of large language models | 829 | 59 | ×9.2 |
| 6 | RoboTwin | Benchmark for bimanual robotic manipulation | 190 | 4 | ×8.4 |
| 7 | Codex | AI system for computer programming | 198 | 6 | ×8.2 |
Which topics are growing fastest in computer science and AI?
Computer Science (cs): 60,937 papers in the latest 13 weeks, 41,266 over the same weeks a year earlier.
| # | Term | What it is | Papers, latest 13 weeks | Papers, a year earlier | Share |
|---|---|---|---|---|---|
| 1 | on-policy distillation | Distillation using samples from the student policy | 392 | 0 | ×20.4 |
| 2 | agent harness | Software infrastructure for running and controlling agents | 268 | 0 | ×14.0 |
| 3 | coding agent | Software agent that performs programming tasks | 720 | 46 | ×10.4 |
| 4 | recursive self-improvement | Iterative improvement of a system by itself | 190 | 0 | ×9.9 |
| 5 | Qwen3 | Family of large language models from Alibaba | 825 | 59 | ×9.3 |
| 6 | RoboTwin | Benchmark for robotic bimanual manipulation | 190 | 4 | ×8.6 |
| 7 | action chunk | Sequence of actions predicted as a unit | 186 | 4 | ×8.4 |
Which topics are growing fastest in astrophysics?
Astrophysics (astro-ph): 6,864 papers in the latest 13 weeks, 6,093 over the same weeks a year earlier.
| # | Term | What it is | Papers, latest 13 weeks | Papers, a year earlier | Share |
|---|---|---|---|---|---|
| 1 | LZ | Liquid xenon dark matter detection experiment | 48 | 4 | ×2.8 |
| 2 | ALMA2040 | Development program for upgrading the ALMA observatory | 39 | 0 | ×2.7 |
Which topics are growing fastest in particle physics?
High Energy Physics - Phenomenology (hep-ph): 2,858 papers in the latest 13 weeks, 2,056 over the same weeks a year earlier.
| # | Term | What it is | Papers, latest 13 weeks | Papers, a year earlier | Share |
|---|---|---|---|---|---|
| 1 | LZ | Liquid xenon experiment searching for dark matter | 136 | 2 | ×7.0 |
| 2 | nuclear recoil | Motion of a nucleus following momentum transfer | 71 | 4 | ×3.4 |
| 3 | mass splitting | Mass difference between related particle states | 88 | 13 | ×3.2 |
| 4 | IceCube | Neutrino observatory embedded in Antarctic ice | 61 | 8 | ×2.6 |
Which topics are growing fastest in biology (bioRxiv)?
bioRxiv: 15,442 papers in the latest 13 weeks, 13,540 over the same weeks a year earlier.
| # | Term | What it is | Papers, latest 13 weeks | Papers, a year earlier | Share |
|---|---|---|---|---|---|
| 1 | Xenium | Platform for in situ spatial transcriptomics | 47 | 2 | ×2.9 |
Which topics are growing fastest in medicine (medRxiv)?
medRxiv: 4,675 papers in the latest 13 weeks, 4,113 over the same weeks a year earlier.
| # | Term | What it is | Papers, latest 13 weeks | Papers, a year earlier | Share |
|---|---|---|---|---|---|
| 1 | linear mixed-effects model | Linear model combining fixed and random effects | 65 | 10 | ×3.2 |
| 2 | mixed-effects model | Statistical model combining fixed and random effects | 107 | 25 | ×3.1 |
Which research topics are losing ground?
Terms that new arXiv papers use at most half as often as in the same weeks a year earlier.
| # | Term | What it is | Papers, latest 13 weeks | Papers, a year earlier | Share |
|---|---|---|---|---|---|
| 1 | natural language processing | Computational processing and analysis of human language | 150 | 440 | ×0.2 |
| 2 | GPT-4o | Multimodal generative model developed by OpenAI | 152 | 411 | ×0.2 |
How to follow a topic
- Each term links to its weekly curve in the rising topics explorer, free for all of arXiv.
- The rising topics page names the topics that jump each week in every arXiv category, bioRxiv and medRxiv; see how to read it.
- The daily paper alert sends the new papers of your categories every weekday, each with a two-sentence summary.
Methodology and limits
- Papers: every new arXiv paper, dated by the submission of its first version (arXiv OAI-PMH), and the new bioRxiv and medRxiv preprints. A paper counts in each archive of its categories.
- Terms: words and phrases of 1 to 3 words of the title and abstract, lowercased, with plurals cut to a stem; the readable form shown is restored afterwards. Terms a single paper used in a week are not stored, so a term absent from a stored week counts as one paper. Spelling variants ("language-model" for "language model") are left out when the other spelling is at least as common.
- Share: papers using the term divided by all papers of the scope, in the latest 13 weeks, divided by the same ratio over the same weeks a year earlier. Listed terms at least doubled their share and pass a likelihood-ratio test (far from chance).
- Filtering: the counts choose the terms. A language model only removes candidates that are not research topics (generic wording, template phrases) and writes the readable form and the short description.
- Words are not topics: a new name for an old idea rises too, and a term can grow because of one large collaboration.