Journal of Machine Learning Research Submission Guide

For every headline result, fill in four cells before polishing prose.

In our editorial analysis, the recurring failure pattern is result without learning: the paper reports that a system wins but never isolates why, when, or at what cost. JMLR's author information explicitly asks papers to report what was learned. A causal claim about a component therefore needs a discriminating ablation, not merely a model-minus-component row with untuned baselines.

JMLR allows some submissions that extend workshop or conference work, but the official policy requires disclosure and a substantive delta. Build a comparison table that a skeptical reader could inspect:

Adding omitted proofs, background, or small experimental variations may be insufficient under the journal's stated examples. The cover letter should disclose the overlap without trying to market a cosmetic delta as a new contribution.

The current author information says submissions use the JMLR electronic system and the JMLR LaTeX style, with a PDF manuscript and a separate cover letter. It describes a 200-word abstract, five keywords, a short running title, overlap disclosures, coauthor consent, conflicts, and suggestions for action editors and reviewers. Recheck the live author page before upload because details can change.

Use this order:

1. confirm that the paper is ML theory or methods for a broad ML audience;
2. write the claim-evidence-boundary table and fix unsupported claims;
3. reconcile conference, workshop, preprint, and related-paper overlap;
4. make experiments reproducible enough to test the central claims;
5. apply the current JMLR format without modifying the style file;
6. assemble the cover letter disclosures and conflict-aware editor/reviewer suggestions;
7. verify the live submission system and inspect the final PDF.

Imagine a conference paper introduced an optimizer and showed gains on image classification. The journal version adds more datasets and longer appendices. That may make the study larger without making it a substantive JMLR extension.

A stronger version asks a question the original could not answer: under what curvature, noise, or scale conditions does the optimizer's advantage arise? It adds an analysis that predicts those conditions, a controlled experiment that can falsify the explanation, and failures outside the predicted regime. The new value is not the number of tables. It is a more general account of learning behavior.

JMLR novelty reduced to a leaderboard margin. A small improvement can be valuable, but the paper must show why it occurs, whether the comparison is fair, and what the result teaches beyond the selected benchmark. Add an explanation that produces a testable prediction, then evaluate that prediction under at least one condition where the method might fail.

JMLR breadth claimed from task count. More datasets do not automatically create a broad ML contribution. Identify the property shared across tasks and the evidence showing that the method addresses that property. If each gain has a different explanation, the central claim may need to be narrower.

JMLR reproducibility treated as a repository link. A repository is useful only when the evaluated version, environment, data path, configuration, and commands can be matched to the paper. Preserve a stable release and state the compute needed to reproduce central tables. Where data cannot be redistributed, separate what is restricted from what can still be inspected.

Cover-letter decision map

The cover letter should not repeat the abstract. It should help the editor resolve potential conflicts and route the paper responsibly.

This source-backed synthesis cannot predict acceptance or editorial priority. It converts public JMLR requirements into checks an author can falsify before submission. The live author information remains the controlling source if any package detail changes.