Academic Momentum

Research project · ICLR 2017–2026

Academic Momentum

Topic Growth and Acceptance Outcomes at ICLR

Interactive atlasAggregate dataMeasurement & robustnessWebsite source

Exploratory analysis · Data snapshot: 7 September 2026

Overview

How does the growth of research topics relate to their subsequent prevalence and acceptance outcomes? We examine 42,123 observed accept/reject records from ICLR 2017–2026, including 41,902 records with usable title/abstract text. Using 28 fixed, overlapping phrase-defined topic groups, we distinguish changes in paper counts, changes in conference share, and differences in acceptance rates.

The analysis reveals a separation between topic expansion and relative acceptance outcomes. The LLM group expanded substantially while its early acceptance-rate advantage narrowed. Declining conference share often coexisted with increasing paper counts. A retrospective comparison of higher-growth topics across historical windows produced different results across starting years. These are descriptive and exploratory findings: they do not identify causal effects of topic choice, reviewer familiarity, or crowding.

Population and measurement. The sample comprises observed accepted and rejected papers, not all submissions. Topic groups are lexical measurements of available text, not mutually exclusive fields, author cohorts, or measures of scientific value.

Selected findings

Post-hoc descriptive examples; all 28 fixed topics are available below.

1. Topic expansion and relative acceptance outcomes can diverge

Between 2023 and 2026, the LLM group increased from 78 to 4,716 papers, or 2.06% to 34.41% of measurable records. Its acceptance-rate difference from same-year non-LLM papers decreased from +19.17 to +0.13 percentage points (pp). The 2026 pointwise 95% interval is [−1.59, +1.85] pp; a near-zero estimate is not evidence of exact equivalence.

(a) Share of measurable records

(b) Acceptance-rate difference vs. non-LLM papers

Figure 1. LLM-group expansion and relative acceptance outcomes, 2023–2026. Share uses all measurable accept/reject records in each year as its denominator. Error bars are pointwise 95% intervals under an independent-paper working assumption. The comparator excludes the focal group; the shrinking difference is therefore not merely the mechanical result of comparing a growing group with a mean that contains it.

The 2023–2026 change in the acceptance-rate difference is −19.04 pp (working 95% CI [−30.15, −7.93]); using 2024 as the start yields −6.88 pp [−10.85, −2.90]. Joint Holm-adjusted p-values across 45 eligible endpoint comparisons are 0.034 and 0.032, respectively. These checks do not resolve unmeasured composition, dependence, or post-hoc selection.

2. Absolute volume, conference share, and accepted counts describe different changes

(a) Graph neural networks

Accepted and rejected counts, 2023 vs. 2026

(b) Diffusion models

Paper counts and shares, 2025 vs. 2026

(c) Generative adversarial networks

Paper counts, 2019–2026

Figure 2. Contrasting count trajectories. GNN counts increased from 210 to 347, with accepted counts increasing from 83 to 93 and rejected counts from 127 to 254. Diffusion counts increased from 642 to 898 while share decreased from 7.54% to 6.55%. GAN counts decreased from 128 to 58 between the displayed endpoints. Groups are fixed phrase matches; these comparisons do not track individual entrants. The 2019 GAN endpoint was selected retrospectively.

This count–share distinction extends beyond individual examples. Among the 20 topic groups with at least 30 papers in 2023, 15 increased in count but decreased in share by 2026. The corresponding counts are 17/24 with a 20-paper threshold and 13/17 with a 50-paper threshold. These are counts of overlapping topic groups, not independent estimates of the proportion of research fields in decline.

Interactive topic atlas

All 28 fixed topic groups. Select a year to inspect counts, denominators, and uncertainty.

Topic share

Topic-group count divided by all measurable records in the same year.

Loading aggregate data…

Show annual data and confidence intervals

Figure 3. Annual topic trajectories. Acceptance-rate comparisons use same-year records outside the focal group. Acceptance curves are suppressed for groups with fewer than 30 papers; counts and shares remain visible. Suppressed or undefined values are not replaced by zero. Downloaded CSVs retain the underlying aggregate values, including low-count years.

Annual changes in topic share

A common color scale shows the magnitude and direction of share changes across all groups. A decline in share need not imply a decline in paper count.

Decrease No change IncreaseScale: −3 to +3 pp; colors saturate outside this range, values do not.

Figure 4. Year-to-year changes in topic share, in percentage points. No change is defined for 2017 because the preceding year is unavailable. Select a cell to inspect the complete trajectory. On narrow screens, scroll the table horizontally.

Historical growth and subsequent acceptance outcomes

At each starting year, we identify the highest-growth quartile of eligible topic groups using only share growth through that year, then compare their mean acceptance-rate differences over the following three years with those of other eligible groups. The estimated contrast is positive for 2020–2023 starting years and negative for 2019. This is a topic-level descriptive comparison, not an individual entry-strategy evaluation.

Figure 5. Subsequent mean acceptance-rate difference: higher-growth quartile minus other eligible groups (pp). A positive value indicates a less negative or more positive relative acceptance outcome, not necessarily an acceptance advantage over the conference. Topic groups and follow-up years overlap; no independent-sample significance claim or confidence interval is assigned to these window contrasts.
Eligibility, timing, and weighting

At starting year t, a group must have at least 20 papers in both t−1 and t. Groups with a percentile rank above 0.75 in the current share change are classified as higher-growth. We average each group's acceptance-rate differences over t+1, t+2, and t+3, then take an equal-weighted mean across groups within each class. The contrast subtracts the other-group mean from the higher-growth mean.

The 2019–2023 windows contain 3, 8, 10, 17, and 19 eligible groups, respectively. The 2018 window has one eligible group and no comparison group; it is retained in the downloadable data. All feasible windows are reported. Higher-growth is a relative rank and need not imply positive absolute growth. The frozen topic vocabulary and design were selected retrospectively; this is not a prospectively preregistered test.

Measurement, robustness, and limitations

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.

Resources and versioning

Topic–year aggregate dataCSV · 28 groups × 10 yearsComplete website snapshotJSON · trajectories, coverage, and window summariesSource checksumsJSON · provenance of the frozen aggregate sourcesWebsite source and historyHugging Face · presentation code, not the full analysis pipeline

This page documents an ongoing research project, not a published-paper record. No publication status or unavailable manuscript link is implied. The public release contains aggregate statistics and presentation code; it does not include paper-level records or the full analysis pipeline. Source checksums identify the frozen inputs but do not by themselves make the analysis fully reproducible.

Suggested project attribution: Nuo Chen, Academic Momentum: Topic Growth and Acceptance Outcomes at ICLR. Data snapshot: 7 September 2026. Website version: 2026-09-08.2.