Most “trending papers” lists are annual cuts: how many times a paper got cited this year versus last. That misses the shape of the curve. A paper that quietly clocked 200 citations evenly across twelve months looks identical to one that went from 5 per month to 40 per month. The second is actually taking off; the first is already famous.

We built a monthly-resolution citation index from OpenAlex’s 357 million papers — the full citation graph for every paper with at least 10 career citations, aggregated to per-paper-per-month totals. 1.67 billion citation edges, 513 million (paper, month) rows, 3 GB on disk. The leaderboard ranks papers by the gap between their recent monthly citation rate and the nine months before that. Here are six of the top movers across six distinct research communities, along with the citation curve that flagged them.

AlphaFold 3 — biology gets a new baseline

What it found. Abramson et al. (Nature 2024, DOI 10.1038/s41586-024-07487-w) extended AlphaFold 2 — which predicts single-protein structures — into a diffusion model that handles whole complexes in one pass: proteins binding to small molecules, DNA, RNA, ions, and modified residues. The paper reports substantially better accuracy on protein-ligand docking than specialized tools like DiffDock and AutoDock Vina that had been the field’s standard.

So what. Most drug discovery is figuring out how a candidate molecule physically sits in a target protein — will it bind, how tightly, which orientation. For thirty years that meant running a stack of specialized docking software that made educated guesses. AlphaFold 3 does not guess; it predicts the joint 3D structure directly from sequences. For pharma and biotech R&D this rewrote the cost structure of the first six months of a discovery program.

What the field did. 43 citations in the month after publication (Nov 2024), rising to a steady 80–120 per month from April 2025 onward. That flat-high plateau is the fingerprint of a paper that stopped being a new finding and became lab infrastructure — every group running structure prediction now cites it in their methods.

Semaglutide — still accelerating four years after publication

What it found. Wilding et al.’s 2021 STEP 1 trial (NEJM, DOI 10.1056/nejmoa2032183) showed that weekly 2.4 mg subcutaneous semaglutide — originally a diabetes drug — produced sustained double-digit-percent weight loss in adults with overweight or obesity. A 2025 Phase 3 companion paper (NEJM, DOI 10.1056/nejmoa2413258) reported that the same dose improved liver histology in patients with metabolic dysfunction-associated steatohepatitis (MASH), a disease category with no approved therapy before this year.

So what. The 2021 paper launched what is now the largest pharmaceutical revenue category in history — Ozempic and Wegovy from Novo Nordisk, plus Eli Lilly’s tirzepatide competitor Zepbound — and has pushed GLP-1 drugs into trials for cardiovascular disease, Alzheimer’s, alcoholism, and pain. The 2025 liver paper extended the same molecule into a multi-billion-dollar indication that pharma had been chasing for two decades without a win.

What the field did. STEP 1 is four years old and still accelerating. It picked up 245 citations in the last 12 months versus 102 the year before, and the second derivative of its citation rate is positive — researchers keep finding new applications to measure against the baseline. The 2025 MASH paper already has 39 citations in its first eight months, the shape of a pivotal trial with regulatory implications.

GLOBOCAN 2022 — when the cancer baseline changes, everyone re-cites

What it found. Bray et al. (CA: A Cancer Journal for Clinicians 2024, DOI 10.3322/caac.21834) is the International Agency for Research on Cancer’s triennial update on worldwide cancer incidence and mortality. For 2022: 20 million new cases globally, 9.7 million cancer deaths, lung cancer most frequently diagnosed at 2.5 million. Lifetime risk: one in five men or women develops cancer; one in nine men and one in twelve women die from it.

So what. GLOBOCAN is the default citation every cancer paper uses to establish the scale of the problem its intervention is aimed at. When IARC releases new numbers, thousands of researchers pivot their introductions within a few months. The paper is not a new discovery; it is the canonical dataset that reframes every other paper’s claim of relevance.

What the field did. 67 citations per month in November 2024, climbing to 304 in June 2025, settling to 180–240 per month through October. That arc is what happens when a reference paper becomes the new baseline — labs replaced their 2020-era estimates, and anyone whose paper cites global cancer figures now cites this one.

PRISMA 2020 — a reporting standard that became infrastructure

What it found. Page et al. (BMJ 2021, DOI 10.1136/bmj.n71) published an updated 27-item checklist for how to transparently report a systematic review or meta-analysis, replacing the 2009 PRISMA guideline that had become the field standard. The 2020 version incorporates newer methods for searching, screening, bias assessment, and synthesis.

So what. Major medical and biomedical journals — BMJ, JAMA, Lancet, Cochrane — require systematic reviews to follow PRISMA formatting as a condition of acceptance. The 2020 version became the required format across most journals in 2023–2024. Every systematic review published now cites this paper in its methods section.

What the field did. PRISMA 2020 is the single fastest-accumulating paper in our dataset. In the last 12 months it picked up 5,264 citations, an average of 438 per month, with recent peaks over 640 — at peak, more than 20 citations a day. For context, that rate exceeds the annual citation count of the median published paper in under a week.

SciPy 1.0 — a software paper that will not stop growing

What it found. Virtanen et al. (Nature Methods 2020, DOI 10.1038/s41592-019-0686-2) is the overview paper for the SciPy scientific computing library — the open-source toolkit providing Python implementations of standard algorithms from statistics, optimization, linear algebra, signal processing, and sparse matrices. The paper is not a single discovery; it is the citeable reference that makes SciPy “the thing everyone uses.”

So what. Any Python-based research paper that does statistics, curve fitting, FFTs, optimization, numerical integration, or sparse linear algebra cites SciPy. It shows up in physics, biology, chemistry, economics, finance, and most applied ML papers. The paired NumPy paper (Harris et al., Nature 2020) picked up 1,269 citations in the same window with a similar slope — the two play the same role at different layers of the scientific Python stack.

What the field did. A five-year-old software paper still gaining about 200 citations per month, with the rate increasing year-over-year. That is the signature of an expanding user base. As AI and data science push Python into more research domains — materials, climate, clinical informatics — SciPy’s TAM grows with them. Its acceleration (93.9) is higher than many papers with ten times its absolute citation count.

“Agentic AI” — a definition paper for an emerging vocabulary

What it found. A 2025 survey paper (IEEE Access, DOI 10.1109/access.2025.3532853) that attempts to formalize “Agentic AI” as a research category — autonomous AI systems that pursue complex goals with minimal human intervention, as distinct from traditional AI that runs on structured instructions. It maps foundational concepts, characteristics, and methodologies others are using.

So what. Survey papers landing in the middle of a hype wave play a specific role: they give a new field shared vocabulary. When VCs, corporate R&D teams, and academics all need to cite something to define “what we mean by AI agents,” a comprehensive survey published early in 2025 becomes the common reference before anyone has agreed on a better one.

What the field did. 27 citations in the last three months (vs. zero before 2025), with a slope that has doubled from April to October. This is the early-stage hype cycle citation pattern — a few per month at publication, then a rapid climb as other researchers tag their work into the new category. It is the weakest signal on this list in absolute terms but the most interesting behaviorally. Whether it stays exponential or collapses will be visible in the next two to three months of data.

Method note

Source: OpenAlex works snapshot ingested to PostgreSQL with deterministic work_ids. Citation graph: 1.67 billion edges extracted from per-paper referenced_works arrays, aggregated to (cited_work_id, year_month, citations) triples. Signal: slope = mean monthly citations in the last 3 months minus mean monthly citations in the prior 9 months of the last-12-month window, anchored at October 2025. Filters: publication year 2018–2025, at least 20 citations in the last 12 months, at least 3 in the last 3 months, citations received in at least 6 distinct months of the window (kills batch-import artifacts). Field clustering: MiniBatchKMeans in BAAI/bge-base-en-v1.5 embedding space, k=20, over the top 300 papers by slope.