Skip to main content
v2026.11,772 entries · CC-BY 4.0
Dictionary termTrack CProposedv2026.2

Reproducible AI experiment

An AI experiment for which sufficient artefacts and metadata are released (data, code, seed, environment, hyperparameters, training procedure) that an independent investigator can re-run it and obtain numerically equivalent or statistically indistinguishable results.

ByCASRAI Editorial Board
· Last updated 22 Aug 2026
Share this

Ask CASRAI · included with Regulatory Radar

Ask about Reproducible AI experiment

Ask CASRAI answers research-administration questions and cites the passages behind every claim — and says so when the corpus does not cover something, instead of guessing. It comes with a Regulatory Radar subscription at $29 a month, alongside the daily digest of regulatory changes and the dashboard of what changed.

150 questions a day, on this site, over the API, or inside your own tools through the CASRAI MCP server.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays free to read, with no account and no card.

Examples

Worked examples

  • Is an instance

    A published model with full training script, locked-version dependencies, random seed, and Docker image producing the reported accuracy on 5 seeds.

  • Is an instance

    An RL experiment with deterministic environment seeds and reported variance across 30 runs.

Counter-examples

Looks similar, but isn't

  • Not an instance

    A paper reporting a single accuracy with no seed disclosure.

  • Not an instance

    A model trained with a closed API call to an undisclosed model version.

Editorial commentary

A reproducible AI experiment is the machine-learning-specific application of computational reproducibility: an independent party re-runs the released training and evaluation code, on the released data, with the released configuration, and obtains results that match the reported ones, either exactly or within the variability the original authors themselves reported across seeds. It differs from a general reproducibility audit mainly in what “matching” is allowed to mean, because ML experiments carry sources of legitimate non-determinism that most other computational research does not.

The obstacles are structural, not just researcher diligence. Stochastic initialisation and data-shuffling mean two runs of identical code on identical data can legitimately differ. Floating-point arithmetic is hardware- and library-version-dependent, so a GPU-trained model may not reproduce bit-for-bit on different hardware, and distributed training adds its own non-determinism from operation ordering across devices. A growing share of published work also depends on closed, versioned commercial model APIs whose weights can change or be withdrawn, making the original experiment permanently un-repeatable regardless of disclosure.

Because exact numerical equivalence is often not achievable even in good faith, ML venues have converged on a weaker but checkable bar: results falling within the reported variability across a stated number of random seeds, rather than an exact match. Pineau and colleagues’ widely cited 2021 paper in the Journal of Machine Learning Research on improving reproducibility in ML research, and the NeurIPS reproducibility checklist that grew out of it, formalised this into concrete disclosure requirements: report the number of seeds run and the variance across them, specify the hardware and software environment (a container image or equivalent), and release the full training and evaluation pipeline, not only the final model weights.

This is a narrower, engineering-oriented bar than replicability in the NASEM sense, which asks whether an independent team collecting genuinely new data reaches the same substantive conclusion. A reproducible AI experiment can be perfectly reproducible in the seed-variance sense while the underlying claim — that the architecture or method generalises beyond the benchmark it was tested on — remains untested. Reviewer checklists and reproducibility requirements at venues such as NeurIPS, ICML, and ICLR are aimed squarely at the narrower, checkable claim, precisely because the broader generalisation claim is much harder to adjudicate at review time.

Also known as

reproducible ML experiment

Machine-readable encodings

Use in your systems

JATS XML <role> element
xml
<role vocab="credit"
      vocab-identifier="https://casrai.org/dictionary/"
      vocab-term="Reproducible AI experiment"
      vocab-term-identifier="https://casrai.org/dictionary/term/reproducible-ai-experiment" />
Schema.org DefinedTerm (JSON-LD)
json
{
  "@context": "https://schema.org",
  "@type": "DefinedTerm",
  "@id": "https://casrai.org/dictionary/term/reproducible-ai-experiment",
  "name": "Reproducible AI experiment",
  "identifier": "https://casrai.org/dictionary/term/reproducible-ai-experiment",
  "description": "An AI experiment for which sufficient artefacts and metadata are released (data, code, seed, environment, hyperparameters, training procedure) that an independent investigator can re-run it and obtain numerically equivalent or statistically indistinguishable results.",
  "inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
  "url": "https://casrai.org/dictionary/term/reproducible-ai-experiment",
  "sameAs": [
    "reproducible ML experiment"
  ],
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "publisher": {
    "@id": "https://casrai.org/#organization"
  },
  "author": {
    "@id": "https://casrai.org/#editorial-team"
  },
  "datePublished": "2026-05-21T02:22:51",
  "dateModified": "2026-08-22T16:05:01",
  "inLanguage": "en-GB",
  "isAccessibleForFree": true
}

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →

Regulatory Radar

Stop finding out after the fact

$29/month, cancel anytime. Daily digest updates from our analysis, a dashboard holding the same items, and a cited assistant for everything they raise.

  • Federal Register, Federal Register+, Grants.gov, Regulations.gov, NSF News, UKRI, plus CASRAI’s own published content.
  • 72,264 indexed passages, and every answer cites the ones it drew on.