Why research papers get rejected: what 2,378 peer reviews reveal
We read 2,378 public peer reviews to find out what reviewers criticise most. The most common reason a paper struggles is not a weak idea. It is a paper that does not explain itself. Here is what the reviews show, and a checklist you can use on your own manuscript before you submit.
Most advice on why research papers are rejected comes from editors' opinions or personal experience. We wanted evidence. OpenReview is one of the few places where peer reviews are published in the open, next to the decision each paper received. That makes it possible to count what reviewers actually complain about, and to compare papers that were accepted with papers that were rejected.
How we did the study
We collected the public reviews of the first 600 submissions listed for ICLR 2025, a large international research conference that publishes every review on OpenReview. That gave us 2,378 reviews: 1,185 on papers that were rejected and 681 on papers that were accepted. The rest were on papers the authors withdrew.
For each review we read the "Weaknesses" section and tagged the kinds of criticism it raised, such as unclear writing, missing comparisons or unjustified assumptions. We then read a random sample of weakness comments from low-scored rejected papers in full, to see what the problems looked like in practice.
Two cautions before the results. First, the tagging was done by keyword, so treat the percentages as a ranking of problems rather than exact rates. Second, ICLR is a machine-learning conference. We have set aside the criticisms that only apply to machine learning, such as computing cost, and kept the ones that apply to research writing in any field.
The results: what reviewers criticise most
| Criticism | All reviews | Rejected papers | Accepted papers |
|---|---|---|---|
| Unclear writing or presentation | 44% | 46% | 38% |
| Narrow scope or overgeneralisation | 33% | 33% | 31% |
| Weak comparison with prior work | 32% | 33% | 27% |
| Figures and tables that do not explain themselves | 32% | 31% | 31% |
| Thin or outdated literature review | 24% | 24% | 19% |
| Assumptions and design choices not justified | 22% | 26% | 18% |
| Unclear or incremental contribution | 21% | 22% | 16% |
| Undefined notation and terms | 14% | 17% | 11% |
| Limitations not discussed | 11% | 11% | 12% |
| Not enough detail to repeat the study | 9% | 9% | 9% |
| Statistical rigour (variance, significance) | 7% | 8% | 6% |
Two things stand out. Accepted papers are criticised too: almost every paper gets a list of weaknesses. What separates rejected papers is that certain criticisms appear clearly more often. The biggest gaps are in unjustified assumptions (26% of reviews on rejected papers against 18% on accepted ones), weak comparisons with prior work (33% against 27%), unclear contribution (22% against 16%) and undefined terms (17% against 11%).
The 7 most common reasons papers are criticised, and how to fix each
1. The writing does not explain itself
Unclear writing was the single most common criticism, raised in almost half of all reviews. Reading the comments, the problem is rarely grammar. It is structure. A model, framework or equation appears without any motivation. A term is used three pages before it is defined. The reader is never told why a step matters.
Fix it: define every construct, term and abbreviation the first time it appears. Before introducing a framework or equation, say in one sentence what problem it solves. Keep one point per paragraph and link paragraphs with signposting ("Building on this...", "In contrast...").
2. The findings are generalised too far
One review in three questioned the scope of the work: results from one setting, one dataset or one sample presented as if they applied everywhere.
Fix it: state the context, sample and boundary conditions plainly. Say where your findings are likely to transfer and where they may not, and why. A clear boundary reads as rigour, not weakness.
3. The paper does not compare itself with the strongest prior work
Reviewers repeatedly asked how a study compares with specific, recent, strong work in the same area, and noticed when it only compared itself with older or weaker approaches.
Fix it: name the closest and most recent studies, and say precisely what is the same, what is different and why the difference matters. A short comparison table is often the clearest way to do this. Never claim your approach is better without a direct comparison.
4. Figures and tables need the text to make sense
Almost a third of reviews raised a problem with a figure or table: unlabelled axes or colours, captions that mention things the figure does not show, or numbers in a table that differ from the numbers in the text.
Fix it: refer to every figure and table in the text before it appears, and tell the reader what to look for. Write captions that make sense on their own. Check that every number is identical in the abstract, the text and the tables.
5. The literature review is thin or out of date
Reviewers noticed when related work stopped several years ago, when a key line of research was missing, or when a cited study was described inaccurately.
Fix it: cover both the foundational studies and the work of the last three to five years. Synthesise rather than list: group studies by theme, and say where they agree, where they conflict and why. Read what you cite.
6. Assumptions and choices are not justified
This criticism showed one of the largest gaps between rejected and accepted papers. Reviewers asked why a sample, a method, a measure or an assumption was chosen, and the paper did not say.
Fix it: for every important decision, explain why, not only what. Each hypothesis or proposition should follow visibly from the theory and literature before it.
7. The contribution is unclear or looks incremental
One review in five questioned what was genuinely new. A common pattern was a paper that combined two known ideas and presented the combination as novel without showing what it revealed.
Fix it: state your contribution in plain words in the introduction ("This study makes three contributions..."), and say what is new compared with the closest prior work. Avoid words such as "first", "novel" or "proves" unless you can show they are true.
A pre-submission checklist
Before you submit, go through your manuscript with these questions. Each one comes from a criticism that appeared again and again in the reviews.
- Is every term, construct and abbreviation defined at first use?
- Is every framework, model or equation motivated before it appears?
- Have I stated the context and limits of my findings?
- Have I compared my study with the strongest and most recent prior work, by name?
- Is every figure and table referenced in the text, with a caption that stands on its own?
- Are all numbers identical across the abstract, text and tables?
- Does my literature review include work from the last three to five years?
- Have I justified each major design choice and assumption?
- Is my contribution stated plainly, and is every claim in proportion to the evidence?
- Have I written an honest limitations section?
- Could another researcher repeat my study from the method section alone?
- For quantitative work, have I reported sample sizes, variance, effect sizes and the exact tests?
How Soba Researcher uses these findings
We ran this study to improve Soba Researcher, an AI Research Assistant. The criticisms above became thirteen writing rules that every paper it drafts now follows, from defining terms at first use to keeping claims in line with the evidence. Its panel of 120 AI Review Agents checks every draft for the same problems, and the paper is rewritten to answer them before you see the final version. In the journal article format, every reference is a real published paper found in OpenAlex.
An AI draft is a starting point, not a finished submission. No tool can guarantee acceptance, and you should check every claim, number and reference yourself. But a draft that has already faced the most common criticisms gives you a stronger place to start. You can try it free at sobaresearcher.com, and read more about how it works in our FAQ.
Key takeaways
- In 2,378 public peer reviews, unclear writing was the most common criticism, raised in 44% of reviews.
- Rejected papers were criticised more often for unjustified assumptions, weak comparisons with prior work, unclear contributions and undefined terms.
- Most of these problems can be fixed in the text, without new data: define terms, justify choices, compare by name, explain every figure, and keep claims modest.
- Use the checklist above on every manuscript before you submit it.
Method: public reviews of the first 600 ICLR 2025 submissions listed on OpenReview (2,378 reviews, collected 13 November 2026), tagged by keyword with a sample read in full. Percentages are approximate and describe one machine-learning conference; the lessons drawn here are those that apply to research writing in general. This article summarises our own analysis and does not reproduce any review. It is not affiliated with OpenReview or ICLR.