Pre-Submission Review for Artificial Intelligence Papers

AI reviewers often start with contribution and evidence:

• what is new compared with prior work
• whether baselines are current and fairly tuned
• whether ablations isolate the contribution
• whether datasets are appropriate and leakage-free
• whether code, models, and training details are enough to reproduce results
• whether claims generalize beyond the tested setting
• whether limitations are honest

For an AI paper, vague novelty is dangerous. A polished manuscript with weak baselines still fails.

In our pre-submission review work, AI papers usually fail in the gap between performance and proof. The numbers look strong, but reviewers ask whether the comparison was fair, whether the method is new enough, and whether the result will survive outside the chosen benchmark.

The common failure patterns are:

• Benchmark theatre: the test setup flatters the method but does not answer the real task.
• Baseline weakness: comparison methods are outdated, undertuned, or selectively reported.
• Ablation gap: the paper claims a mechanism but does not isolate which component matters.
• Reproducibility fog: code exists in theory, but details needed to rerun the result are missing.
• Claim overreach: the abstract claims general AI value from narrow experiments.

Good field review should identify which pattern will dominate reviewer comments.

Nature Machine Intelligence policies include reporting standards and availability of data, materials, code, and protocols. JMLR author guidance tells authors to situate work in the broader literature and notes that submissions can be rejected without written review if clearly below the journal's standards or out of scope. Nature Methods has also published reproducibility standards for machine learning in the life sciences, emphasizing data, code, models, and documentation.

The practical standard is not "we can run it in our lab." It is "a reviewer can understand and evaluate enough of the pipeline to trust the result."

Send the manuscript, target venue, code repository if available, dataset links or access constraints, appendix, experiment logs, baseline details, hyperparameters, random seeds, and a short statement of what you believe the main contribution is.

That last statement matters. If the authors cannot state the contribution in one sentence, reviewers will not rescue it.

Before submission, authors often need to:

• add a stronger related-work contrast
• rerun baselines with fair tuning
• add ablations for the central architectural or training choice
• clarify dataset splits and leakage controls
• include compute, seed, hyperparameter, and environment details
• state limitations without weakening the contribution
• narrow claims about generalization

These fixes are more important than copyediting when acceptance depends on trust.

The best output is not a long list of every possible weakness. It should identify the reviewer's likely first objection and the fastest credible fix.

For AI papers, the response should be technically specific. "Add more experiments" is not enough. The review should say which baseline, which ablation, which leakage risk, or which claim is likely to matter.

AI manuscripts often overstate in predictable places. A model trained on one dataset becomes a claim about a class of problems. A benchmark improvement becomes a claim about practical deployment. A method with limited ablations becomes a claim about why the method works. A narrow dataset becomes language about human-level, general, or real-world performance.

Pre-submission review should cut that language before reviewers do. Narrower claims are not weaker when they match the evidence. They make the contribution easier to trust.

Before choosing a target, ask:

• Does the venue reward method novelty, application value, theory, or reproducibility?
• Are similar papers accepted there, or only cited there?
• Would a reviewer in this venue expect open code?
• Is the contribution incremental for a top ML venue but strong for an applied journal?
• Does the paper need a conference-first strategy or a journal-first strategy?

That venue-fit step keeps this page from cannibalizing the generic AI review product page. The page is about publishing AI research, not reviewing manuscripts with AI.

It also keeps conference strategy separate from journal readiness. A paper may be strong enough for a workshop or applied venue but not ready for a selective journal that expects fuller theory, stronger reproducibility, or broader external validation.