Transactions on Machine Learning Research Submission Guide

In our editorial analysis, the recurring failure pattern is correct result, incomplete comparison. The code may run and the reported number may be accurate, yet the paper's conclusion exceeds the experiment because baselines received different tuning, data, compute, or stopping rules. Correctness includes the inference connecting results to claims, not only the absence of implementation errors.

The official author guide requires anonymization and active OpenReview profiles. Before upload, inspect more than the author line:

• remove names, affiliations, acknowledgments, grant identifiers, and revealing repository owners;
• phrase self-citations in the third person when the citation is scientifically necessary;
• check PDF metadata, supplementary filenames, document properties, and embedded links;
• provide an anonymous reproducibility route when the current policy permits it;
• keep the eventual public artifact plan ready so anonymity does not become permanent opacity.

Anonymity and transparency are sequential constraints, not opposites. The review package must not identify the authors, while the accepted record should still support inspection and reuse under the venue's current policies.

Suppose a paper shows that a popular method's reported advantage disappears after equalizing hyperparameter search. The result can fit TMLR if the reproduction is faithful, the search budget is explicit, uncertainty is reported, and the paper explains which evaluation practice caused the earlier conclusion.

It becomes weak if the authors use a different code path, quietly change preprocessing, or frame one failed reproduction as evidence that the entire method family is invalid. The constructive contribution is a falsifiable evaluation lesson: under matched search and preprocessing, the advantage is absent on these tasks, and future comparisons should control the identified variable.

TMLR correctness reduced to code execution. A program can run while the scientific comparison is invalid. Check whether data, tuning, stopping rules, compute, and uncertainty are matched. Then ask whether the conclusion would still hold if the strongest baseline received the same care as the proposed method.

TMLR anonymity broken by the evidence route. Removing names from the PDF is not enough when a repository URL, model card, dataset record, acknowledgments file, PDF metadata, or highly specific self-reference reveals the authors. Perform an identity scan on every uploaded artifact. Preserve the scientific context needed for review while replacing identity-bearing routes with policy-compliant anonymous ones.

TMLR ethics treated as boilerplate. A generic statement does not resolve who can be affected, what failure can occur, or what safeguards exist. Tie the ethics discussion to the actual data, task, deployment assumptions, misuse possibilities, and excluded uses. When human-subjects oversight applies, make the submission information agree with the manuscript and institutional record.

OpenReview readiness table

A falsifier-first abstract test

Underline every comparative or generalization claim in the abstract. For each, write one plausible result that would make the sentence false. Confirm that the paper actually performs a relevant test. If no result could falsify the sentence, it may be promotional rather than scientific. If the falsifier is outside the study, state the boundary rather than implying it was tested.

This source-backed synthesis cannot predict acceptance or editorial priority. It translates current public requirements into a correctness and package audit; TMLR editors and reviewers retain the actual judgment. Before upload, preserve the dated policy pages and compiled package used for the final audit so later changes can be reconciled without guessing.