ICLR 2024 vs 2025

Same venue, two consecutive editions, same analysis method. What changed between 2024 and 2025 rejections: in volume, decision distribution, reviewer scores, and what reviewers critique.

1. Volume evolution

Objective figures from OpenReview's public API, independent of how we cluster weaknesses. The 2025 edition is markedly larger.

Visible submissions

7,404

→

11,672

+57.6%

With decision

5,780

→

8,727

+51%

Accepted

2,261

→

3,708

+64%

Acceptance rate

39.1%

→

42.5%

+3.4 pp

2. Decision distribution

How committee decisions split each edition. Lets you see whether accepts concentrate in one tier (poster, spotlight, oral) more than another.

ICLR 20245,780 decisions

60.9%

31.3%

ICLR 20258,727 decisions

57.5%

35.7%

Reject
Accept (oral)
Accept (Oral)
Accept (poster)
Accept (Poster)
Accept (spotlight)
Accept (Spotlight)

3. Reviewer score distribution

Distribution of reviewer scores. Normalised to relative share within each year, so both editions are comparable even though 2025 receives many more reviews.

4. Top 5 concerns per year

The five heaviest clusters per edition, without forcing matches between years. Useful to see what dominated each year as it emerged from the clustering.

ICLR 2024

Before

#01
Unclear definitions and algorithms
authorsequationtheoremalgorithm
22.1%
#02
Insufficient datasets and training setup
modeldatamodelsdataset
17.3%
#03
Unclear core contribution
papermaincontributionlacks
10.6%
#04
Baseline comparisons missing or weak
methodproposedproposed methodpaper
8%
#05
Positioning against existing work
methodsexistingcomparisonexisting methods
6.4%

ICLR 2025

After

#01
Loose definitions and evaluation
modelsmodelmethodsperformance
27.7%
#02
Algorithms and equations unclear
authorsalgorithmeqequation
17.2%
#03
Central contribution under-articulated
paperdoesweaknessespaper does
11.2%
#04
Incremental method
methodproposedproposed methodmethods
8.4%
#05
Poorly presented experimental results
resultsexperimentalexperimental resultspaper
6.4%

5. Shared-category projection

Projecting both years' clusters onto 11 shared categories yields the chart below. Read the deltas with caution: part of the change may be due to a large 2025 cluster absorbing themes that were split in 2024.

How to read this

Volume figures (Section 1) are the most reliable: they count records directly.
Decision distribution (Section 2) and rating distribution (Section 3) are objective but their comparison depends on each edition's editorial policy (e.g., spotlight or oral criteria can shift year to year).
The top concerns (Section 4) is the most useful section for researchers: it shows what dominated each year, as is.
The category projection (Section 5) is a visual guide, not quantitative truth. The cluster→category mapping is heuristic and can be reviewed in the site's source code.