Data and topic assignment
The unit is an observed accepted or rejected paper from ICLR 2017–2026. Fixed regular expressions match available title/abstract text. One paper can match several groups; groups do not exhaust research. Counts were checked against the stored text for all 280 topic–year cells. This verifies implementation consistency, not semantic classification accuracy.
Annual coverage
Of 42,123 decision records, 41,902 have usable text. Missing text affects 217 records in 2025 and four across the remaining years. All plotted shares use the measurable-text denominator.
Fixed phrase definitions
Vocabulary changes, ambiguous abbreviations, and renaming can affect matches. First observed use is not a measure of a field's origin or maturity. No human semantic-accuracy estimate is claimed.
Measures and comparison population
Let n be a topic group's paper count and a its accepted count; N and A are the corresponding counts across all measurable records in that year. All shares and rate differences below are expressed in percentage points or percent as indicated.
Share (%)s = 100 n / N
Share change (pp)mt = st − st−1
Acceptance-rate difference (pp)g = 100 [a/n − (A−a)/(N−n)]
Comparing with records outside the focal group avoids the mechanical attenuation induced by including a growing group in its own reference mean. It does not adjust for differences in paper, author, or task composition.
Uncertainty and sensitivity checks
Annual acceptance-rate differences use Newcombe-style intervals constructed from Wilson intervals under an independent-paper working assumption. These are pointwise 95% intervals, not simultaneous bands over the atlas. Endpoint changes use Wald working intervals and a joint Holm correction across 45 eligible comparisons: 20 groups for 2023–2026 and 25 for 2024–2026.
What the additional checks establish
The LLM endpoint decline remains supported under both starting years in that working analysis. GNN and diffusion acceptance-rate declines do not survive the same joint correction (adjusted p = 0.541 and 0.273 for 2023–2026). Their count and share trajectories remain descriptive observations, not evidence of a general crowding mechanism.
Full-name and abbreviation variants retain several reported directions, but exact magnitudes and uncertainty vary. The count-up/share-down pattern is 15/20 at a 30-paper threshold, 17/24 at 20, and 13/17 at 50 for 2023–2026.
For diffusion in 2025, assigning all 217 missing-text records to the topic gives at most 859 papers, still below 898 in 2026. Assigning all to other groups gives a minimum 2025 share of 642/8,726 = 7.36%, still above 6.55% in 2026. These are extreme-allocation bounds, not confidence intervals.
Interpretation
This is an exploratory, post-hoc analysis. Phrase groups, years, and authors can overlap. The intervals do not account for all author/laboratory dependence, measurement error, unobserved composition, or choices made during exploration. The complete set of topic trajectories is provided to contextualize selected examples.
Growth-persistence results in the wider analysis are sensitive to topic definitions, LLM-group leverage, and inference with few years. This project does not establish a universal momentum–crowding–reversal mechanism. Reviewer familiarity, individual entry timing, counterfactual publication outcomes, and long-term knowledge production are not measured here.