International Conference on Machine Learning (ICML)

Conference positioning

International Conference on Machine Learning (ICML) is a top computer-science conference venue for core machine learning methods, theory, applications, and responsible deployment. It rewards a technically rigorous ML contribution with clean experiments, clear limitations, and current-cycle formatting discipline. Treat this skill as a fit / venue-selection / re-framing tool for conference submission strategy, not as a substitute for the current year's CFP, author kit, ethics policy, or submission portal.

Because CS conferences change deadlines, templates, page limits, review workflow, artifact rules, AI-use policy, and rebuttal formats every cycle, always verify the live official instructions before making a submission-ready recommendation. Start from the official source anchor recorded for this venue in ../../resources/conference-roster.md and ../../resources/official-source-map.md.

When to trigger

• The author names ICML / International Conference on Machine Learning as the target venue.
• A manuscript in core machine learning methods needs a conference-fit read before being formatted or submitted.
• The paper must be re-framed from journal style or arXiv style into a selective CS conference narrative.
• The author needs an evidence-gap, anonymity, artifact, rebuttal, or re-routing diagnosis for this venue.

Scope & topic fit

• Core fit: core machine learning methods, theory, applications, and responsible deployment.
• Best submissions make a precise contribution type visible: algorithm, theorem, system, dataset, benchmark, empirical finding, design artifact, tool, or socio-technical analysis.
• The paper should explain why the result matters to ICML's reviewers, not just why it is interesting to the authors' lab or product context.
• Position related work against the most recent conference-cycle papers in this venue and its closest siblings; stale comparisons are a common early-review weakness.
• If the contribution is interdisciplinary, state which part is CS research and which part is domain evidence.

Venue-specific calibration

• Reviewer lens: Read reviewers as cross-area AI specialists. The paper needs a clear ML or AI research contribution, strong baselines, honest limitations, and enough breadth to matter outside one lab benchmark.
• Contribution hook to foreground: the venue-specific contribution bar.
• Scope vocabulary to use naturally in the abstract and introduction: core machine learning methods, theory, applications, and responsible deployment.
• Distinctive fingerprint for reviewer calibration: core, machine, learning, methods, theory, applications, responsible, deployment, venue-specific, contribution, flagship, icml.
• Official anchor domain: icml.cc. Quote annual rules only after opening that source and the current-year CFP/author kit.

Close-neighbor routing guardrail

• Route to ICML when the contribution is a machine-learning method, theory, evaluation protocol,

or empirical result with rigorous ML baselines and ablations.

• Compare ICLR for representation/deep-learning framing, NeurIPS for broad ML/AI impact,

AISTATS/UAI/COLT for statistical or theoretical claims, and MLSys for systems bottlenecks.

What distinguishes this venue from its closest siblings

• What ICML is. The International Conference on Machine Learning (IMLS) — the broad ML flagship spanning theory and applications.
• vs ICLR / NeurIPS. ICLR is deep-learning-forward (OpenReview) and NeurIPS is the broad ML/neuro flagship; the three overlap heavily — route by cycle and community.
• Routing. Learning theory → COLT; vision → CVPR/ICCV/ECCV; NLP → ACL-family.

ICML-specific routing detail

• Prefer ICML when the paper is a broad machine-learning contribution: algorithms, theory-informed methods, optimization, probabilistic learning, evaluation, or generalization across tasks.
• Route representation-learning framing to ICLR, broader AI/society/systems mix to NeurIPS, statistical inference to AISTATS, and ML systems to MLSys.
• ICML evidence should include strong baselines, ablations, theoretical or empirical justification, compute/reporting transparency, and task diversity matching the claim.

Method & evidence bar

• Compare against current strong baselines and explain exactly what changes in the algorithm, objective, data, or inference procedure.
• Report ablations that isolate the claimed mechanism; do not rely on aggregate benchmark wins alone.
• Document data, compute, hyperparameters, model selection, and failure cases so the result can be reviewed as science rather than demo output.
• For ICML, the evidence must support the venue-specific signature: a technically rigorous ML contribution with clean experiments, clear limitations, and current-cycle formatting discipline.
• Include limitations, negative results, compute/resource reporting, data provenance, and ethics details when they affect the claim.

Structure & house style

• Frame the contribution as a reusable idea: method, theory, benchmark, dataset, system, or socio-technical finding.
• Separate main claims from exploratory results; reviewers at top AI venues punish overclaiming and hidden cherry-picking.
• Use the current official template exactly; do not guess page limits, font sizes, supplement rules, anonymity exceptions, or camera-ready requirements from old cycles.
• The introduction should answer: problem, why now, what is new, why this venue, and what evidence proves the claim.
• Put the strongest result in the main paper, not only in the appendix or supplement; reviewers should not have to reconstruct the contribution.

Official-cycle checklist

• Open the live official venue page: https://icml.cc/
• Re-check the current cycle's CFP, author kit, submission system, abstract/paper deadlines, page limits, supplementary-material rules, anonymity policy, dual-submission policy, ethics policy, AI-use policy, artifact/code/data expectations, rebuttal/author-response format, and camera-ready requirements.
• Confirm the review workflow and portal: OpenReview and the current-year official author guide.
• Check whether accepted papers require in-person presentation, separate registration, artifact badges, proceedings copyright, or post-acceptance release forms.
• If the live official instructions conflict with this skill, the official instructions win.

Pre-submission self-check

• ] One sentence states why this manuscript belongs at ICML, using the venue's scope rather than generic "top conference" language.
• ] The claim is calibrated to the evidence: no broader than the datasets, proofs, systems, user studies, deployments, or threat model support.
• ] Related work includes the nearest current-cycle AI/ML flagship papers and explains the technical delta.
• ] The paper satisfies the current official template, anonymity, ethics, artifact, and rebuttal requirements.
• ] The main paper is self-contained enough for reviewers to evaluate novelty and correctness without hunting through external links.

Common desk-reject triggers

• Leaderboard-only novelty with weak explanation of why the method works.
• Unclear data contamination, missing baselines, or evaluation that cannot be reproduced.
• Claims about safety, fairness, health, or society without matching evidence and limitations.
• Formatting, anonymity, dual-submission, external-link, or supplement violations under the current-year policy.
• A contribution framed for a neighboring field while giving ICML reviewers too little technical or empirical substance.

Re-routing decision

If the paper misses ICML's bar, compare against neural-information-processing-systems / international-conference-on-learning-representations / aaai-conference-on-artificial-intelligence / international-joint-conference-on-artificial-intelligence. Re-route based on contribution type, not prestige: theory to a theory venue, systems to a systems venue, application-heavy work to a domain venue, and early ideas to workshops or shorter tracks when the official CFP supports them.

Output format

text
[Fit] High / Medium / Low (one-line reason)
[Target] International Conference on Machine Learning (ICML)
[Contribution type] algorithm / theory / system / dataset / benchmark / empirical / design / security / other
[Main evidence gap] <single most important missing proof, experiment, study, artifact, or policy check>
[Official items to re-check] CFP / author kit / deadline / format / anonymity / ethics / AI-use / artifact / rebuttal / camera-ready
[Top rejection risk] <venue-specific risk>
[Re-route suggestion] <better-matched conference or journal if not a fit>