Research Like Stock Trading: 40K ICLR Papers Show Chasing Hot Topics Is Rational — But Alpha Decays Fast

The author draws an analogy between academic publishing and stock-market momentum investing: a research direction gains traction after a few influential papers, reviewers become familiar with it, acceptance rates rise, senior students publish there and juniors follow, and big labs release benchmarks and open-source models — adding "volume" to the trend.

However, momentum has crashes. In finance, the 2009 market rebound wiped out years of momentum profits in months. In academia, prompt engineering enjoyed a +19 percentage-point acceptance premium in 2023 (21 papers), but by 2025 the premium turned to –11 points (papers in that direction were accepted less often than others). The more crowded a direction, the faster its alpha decays because multiple groups pursue the same idea and race to arXiv.

Data and Metrics

The study uses all ICLR 2017–2026 papers with known accept/reject decisions — 42,123 papers . Each paper is tagged by fixed phrases in title/abstract into 28 non-exclusive directions (e.g., "graph neural network" → GNN). Three metrics are computed per direction per year:

Share : papers in the direction divided by all papers that year. LLM reached 34% in 2026 (one in three papers).

Share change (momentum): this year's share minus last year's. Positive = heating up, negative = cooling down.

Relative acceptance-rate difference : acceptance rate of papers in the direction minus acceptance rate of papers not in the direction, controlling for yearly overall acceptance-rate fluctuations. A value of +19 means 19 percentage points higher acceptance.

LLM: Trend Self-Reinforcement Confirmed

LLM papers grew from 78 (2% of conference) in 2023 to 863 in 2024, 2,357 in 2025, and 4,716 (34%) in 2026 — a 60× increase in three years. The mechanism described earlier (influential papers → reviewer familiarity → senior student success → junior follow-on → industry benchmarks) is visible in the data.

Acceptance Premium Disappears at Peak Heat

In 2023, LLM papers had a +19 pp acceptance premium. It fell to +7 in 2024, 0 in 2025, and +0.13 (confidence interval crossing zero) in 2026 — statistically indistinguishable from no advantage. After multiple-comparison correction across 45 tests (28 directions × several years), the decline remains significant. In momentum terms: the trend continues but alpha has been fully arbitraged away.

Agent and Prompt Engineering: Rapid Alpha Decay

Agent papers grew from 128 (2025) to 533 (2026) with a 38% acceptance rate, but the relative acceptance difference was –1 (CI crossing zero) — no premium. Prompt engineering showed a clear arc: +16 pp premium in 2023 (21 papers), +2 in 2024, –11 in 2025 , –6 in 2026. Papers multiplied sixfold while the premium turned into a discount.

Directions Left Behind: Growth Goes into Rejections

GNN papers rose from 210 (2023) to 347 (2026) (+65%), but acceptances only grew from 83 to 93 (+10) while rejections doubled from 127 to 254. Acceptance rate fell from 40% to 27%. This illustrates alpha decay: the same idea is pursued by many groups simultaneously; the losers become rejections. The decline is not statistically significant after multiple-comparison correction, so the author treats it as descriptive.

Counter-Intuitive: Niche Directions Are Not Necessarily Small

ICLR itself expanded 3.6× (3,792 to 13,704 decided papers) from 2023 to 2026. Among 20 directions with ≥30 papers in 2023, 18 grew in absolute count, but only LLM, diffusion models, and interpretability outpaced the 3.6× conference growth. 15 directions grew in absolute numbers yet shrank in share — they are "running slower than the market." Causal inference, continual learning, and fairness all more than doubled in paper count but became more marginal in the conference.

Three Flavors of "This Direction Is Cooling"

Share ebbing but headcount still rising — diffusion models: 642→898 papers, share 7.5%→6.6%. People still enter, just slower than LLM.

Actual headcount decline — GAN: 128 (2019) → 58 (2026), halved in one PhD cycle. Only GAN and NAS show this pattern.

Headcount rising but all growth lands in rejections — GNN (as above).

These three situations look the same from outside ("the field is cooling") but imply completely different prospects for a PhD student.

Backtest: Does Chasing Heat Actually Help?

Using only data available at each year-end, the author labels the top quartile of directions by share growth as "hot" and compares their average relative acceptance-rate difference over the next three years against the rest. From 2020–2023 starting points, the hot group outperformed by 1–3 pp; 2019 was negative but had only three eligible directions. Chasing heat is not wrong. However, both groups' average premiums were negative — entering an already-established direction yields slightly lower acceptance than the conference average. Hot directions merely let you lose less; they are not a free pass.

Full 28-Direction Heatmap

All 28 directions' annual share changes (green = heating, red = cooling) are shown, sorted by 2026 paper count. The cells represent share change, not acceptance rate or scientific value.

What Can and Cannot Be Concluded

Data scope: ICLR accept/reject records, fixed-phrase matching, non-exclusive tags, not all submissions. 2025 misses 217 papers but key numbers are robust. All findings are historical associations, not causal.

Cannot conclude:

"Late entrants get exploited" — no individual entry-time data.

"Crowding causes rejections" — no causal identification.

"GAN has no value" — no long-term contribution or industry data.

The single actionable takeaway: When choosing a research direction, a PhD student should ask not just "Is it hot?" but "Is this direction's growth landing in acceptances or in rejections?"