arXiv cs.LG · cs.CL · cs.AI · cs.CV · cs.NE · cs.RO · cs.IR · stat.ML — 2012 to 2026
Every term below is measured the same way: the share of that year's AI papers whose title or abstract mentions it. Share, not count — the corpus grew 31-fold across this window, and raw counts would make everything look like a boom. Some of these words won. Some died. A few won so completely they stopped being worth saying.
Press play, or drag the year. Words hold their position across years so you can watch individual terms swell and shrink in place — the usual random-placement word cloud makes that impossible to see. Only each year's 40 largest terms are drawn, so the canvas stays readable; the cutoff climbs from 0.02% in 2013 to 1.04% in 2026 as the field's vocabulary concentrates. Click any word to pin its trajectory.
The same data stacked by vocabulary family. The classical band shrinking to a hairline while the LLM band swallows the top half is the whole story in one shape. Toggle families in the legend above; the key underneath lists what's in each.
Word clouds are bad at trajectory — area scales nonlinearly and the eye can't compare across frames. These can. Each sparkline is that term's share from 2012 to 2026, scaled to its own maximum; the marker sits on the peak year. The percentage is the drop from peak to now. Reading left to right and top to bottom walks forward through the field's successive crests: classical methods first, then the CNN years, then the LLM era.
Peaked at 0.24% in 2018, three years after Hinton's push. In 2026 the figure rounds to 0.00% — a 98% collapse and the only term in the set that effectively hits zero. GANs are the runner-up: 2.86% in 2018 down to 0.27%, killed outright by diffusion.
From 1.89% in 2017 to 0.11% now — a 94% fall that means the opposite of failure. Neural machine translation did the same (95% down). Both became so standard that naming them stopped carrying information. A falling line is ambiguous: it can mean death or total victory.
Bare agent sits flat at 3.5–5.5% for twelve years — that's RL and multi-agent systems, nothing to do with LLMs. Then 6.3 → 9.6 → 14.6% in three years. Query the naive term and you inherit a decade of unrelated baseline; agentic was near-zero until 2023.
Normalizing flows peaked at 0.33% in 2021 and faded — the invertibility constraint was never worth it. Flow matching, its unconstrained descendant, is already at 1.08%: more than triple its ancestor's all-time high, reached in a third of the time.
XAI never crashed; it climbed to 1.95% and plateaued. Mechanistic interpretability sat in noise at 0.03% for a decade, then went 0.21 → 0.57 → 0.99%. Same research goal, disjoint vocabulary, communities that barely cite each other.
Bayesian methods peak in 2012 at 6.44% and fall 92%. SVMs fall 89%, topic models 93% — monotonic decline, exactly as a control should. Meanwhile neuro-symbolic (0.52%) and spiking networks (0.58%) never inflect at all. The trends are real, not an artifact of corpus growth.
Source: full arXiv metadata snapshot (librarian-bots/arxiv-metadata-snapshot,
dated 2026-07-27), filtered to the eight AI/ML categories listed above — 625,296 papers with
a v1 submission date between 2012 and 2026. Each term is a regex matched against title +
abstract; a paper counts once per term regardless of how many times the phrase appears.
Denominator is all papers in that year, so every figure is a share and terms may overlap.
2026 is partial — roughly seven months, 80,856 papers. Shares are still comparable; a full-year count would not be.
Acronyms that collide with English words (KAN, LoRA,
ViT) are matched case-sensitively. RAG is matched as
"retrieval-augmented" and GAN as "generative adversarial" to avoid
rag and gan. Bare agent is kept deliberately as a
demonstration of what naive matching costs you.
Known soft spots: prompting catches generic uses of "prompt"; attention catches some prose usage; diffusion picks up physical and graph diffusion in the early years. Terms measure vocabulary, not quality, citations, or influence — a word can fall because the idea failed or because it won so thoroughly it went without saying, and this chart cannot tell you which.