icml-reproducibility

ICML Reproducibility

Use this when the paper's acceptance risk is tied to whether experiments, code, or theory can be

trusted. ICML reviewers are asked to evaluate soundness, and ICML author instructions state that

reproducibility and code availability are considered in decisions.

Evidence checklist

• Data: source, license, preprocessing, splits, leakage checks, and access restrictions.
• Code: anonymous review package, environment, dependencies, exact commands, and expected runtime.
• Randomness: seeds, variance, confidence intervals, or explanation for deterministic runs.
• Compute: hardware, training budget, evaluation budget, and fairness relative to baselines.
• Baselines: tuning protocol, implementation source, and why missing baselines are not applicable.
• Theory: assumptions, theorem statements, proof dependencies, and edge cases.
• Impact: support claims in the impact statement with real evidence or narrow the statement.

ICML-specific public-record issue

Accepted papers may publish original supplementary material. Do not put unreleasable data, identity

leaks, or private credentials in the review package. If data cannot be public, document the access

path and ethics constraints in the paper.

Reproducibility scoring lens

ICML reviewers fold reproducibility into the soundness judgment rather than scoring it on a separate

axis, so the question is whether a skeptical reviewer could regenerate the headline number.

| Signal a reviewer checks | Strong evidence | Weak signal that invites doubt |

| --- | --- | --- |

| Code in review package | Anonymized, runnable, exact commands | "Code on acceptance" promise only |

| Variance | Seeds with intervals | Single run, no spread |

| Compute | Hardware and budget table | Unstated cost, unfair comparison |

| Theory | Assumptions and proof dependencies listed | Theorem with hidden conditions |

Worked vignette: theory-plus-benchmark paper

For an adaptive-optimizer paper, reproducibility means the convergence proof's assumptions are

written out, the benchmark scripts run from the anonymized supplement with a fixed seed, and the

compute table lets a reviewer judge whether the speedup is real or a tuning artifact. The recurring

failure is a clean theorem paired with benchmark code that silently relies on an unreleased internal

dataset; document the access path or move to a public dataset before the deadline.

Reviewer-pushback patterns and the ICML fix

| Pushback | ICML-specific fix |

| --- | --- |

| "Cannot reproduce without the code" | Ship the anonymized runnable package now, not a post-acceptance promise |

| "Hyperparameters undocumented" | Add the search protocol and final values to the appendix |

| "Speedup may be a seed artifact" | Report multiple seeds with confidence intervals |

| "Theorem assumptions hidden" | List assumptions, proof dependencies, and edge cases explicitly |

Because accepted ICML papers can publish the original supplementary material on OpenReview, the

reproducibility package is also a public commitment. Confirm before the deadline that every file is

releasable, every license is stated, and no private credential or identity path remains.

Output format

[Reproducibility status] strong / adequate / weak
[Weakest claim] <claim not yet supported>
[Required fix] <code/data/seed/compute/baseline/proof>
[Supplement/public-record risk] <none or issue>
[Reviewer-facing sentence] <concise reproducibility statement>