aistats-reproducibility

AISTATS Reproducibility

Use this before submission and again before camera-ready. Reopen the current CFP and OpenReview forms to confirm whether a reproducibility checklist is required.

Evidence map

• Map each theorem, algorithmic claim, simulation claim, and empirical claim to a verifiable location in the paper, appendix, supplement, or artifact package.
• For theory, state assumptions, proof dependencies, convergence conditions, constants, and failure modes clearly enough for statistical readers.
• For experiments, report datasets, splits, preprocessing, evaluation metrics, baselines, hyperparameter ranges, final selected settings, seeds, repeated runs, compute, and runtime.
• For small performance differences, add uncertainty estimates: standard errors, confidence intervals, paired tests, bootstrap intervals, or repeated trials as appropriate.
• Explain missing code/data honestly and describe how a reader could reproduce the analysis in principle.
• Keep the checklist consistent with the manuscript; contradictions between checklist and paper are review-risk multipliers.

Checklist-to-claim audit table

Checklist item	Pure-theory answer	Theory-plus-experiments answer
Code availability	NA only if there is literally no computation	Anonymous archive, or an honest stated reason
Assumptions stated	Every theorem lists its conditions inline	Plus a note on which experiments satisfy them
Error bars	NA for deterministic results	Required for every stochastic figure and table
Compute resources	NA	Hardware, runtime, and total number of runs

Marking NA on an item the paper actually triggers is a recognizable AISTATS red flag, because reviewers cross-check checklist answers against the PDF and read contradictions as carelessness about the rest of the paper.

Vignette: a rates-plus-simulation paper

Consider a submission proving posterior contraction rates for a Bayesian nonparametric model, validated by MCMC simulation. Its reproducibility spine: prior hyperparameters and their selection rule, chain length, burn-in, convergence diagnostics, replication seeds, and a statement of which contraction-theorem conditions the simulated model satisfies — plus one honest sentence about the condition it does not.

Degrees of reproducibility

• Turnkey: one command regenerates each figure from logged seeds.
• Scripted: scripts exist but require documented manual steps or external data access.
• Descriptive: prose detailed enough that a competent reader could rebuild the pipeline.

For AISTATS, simulations should be turnkey because statistician reviewers actually rerun them; large real-data pipelines may stay scripted with deviations documented. Stating the achieved level honestly beats overpromising turnkey behavior that fails on a clean machine.

Output format

[Claim inventory] <claim -> evidence location>
[Checklist status] complete / inconsistent / missing
[Statistical reproducibility gaps] <assumptions/seeds/uncertainty/hyperparameters/compute>
[Paper fixes] <must appear in main PDF>
[Supplement fixes] <appendix or artifact additions>