reproducibility
In plain wordsWhether other people, or the same people later, can get the same results by rerunning an AI experiment. Many published AI results have proved hard to reproduce, which makes real progress harder to judge.
The property that a reported result can be obtained again: narrowly, by rerunning the original code on the original data; more broadly (often called replicability), by independent researchers using new implementations or data. Fields and organizations define the two words in opposite ways, and the ACM reversed its own definitions in 2020 to align with other bodies. In machine learning, failures stem from unreleased code, data or checkpointscheckpoint, n.: A saved snapshot of a model partway through or at the end of training, like a save point in a video game: training can resume from it after a crash, and finished snapshots are what people download and run.; unreported hyperparametershyperparameter, n.: A setting chosen by people before training, such as how fast a model learns or how big it is, as opposed to the internal values the model learns by itself from data., preprocessing and tuning budgets; variance across random seedsrandom seed, n.: A number that fixes the 'random' choices a computer makes; reuse the same seed and settings, and an image generator produces the same picture again. that can exceed the claimed improvement; weak or untuned baselinesbaseline, n.: A simple reference method that a new model must beat to show it is worth anything, such as always guessing the most common answer or reusing yesterday's value.; data leakagedata leakage, n.: A mistake in building a model that lets it peek at clues it will not have in real use, such as hints about the very answer it must predict, so it shines in testing and then disappoints. and test-set contaminationdata contamination, n.: When the test questions used to grade an AI system were already in its training data, so a high score may reflect memorized answers rather than real ability.; and nondeterminismnondeterminism, n.: The fact that an AI model can give different answers to exactly the same question, partly because it deliberately chooses words with some randomness and partly because of tiny computing differences. in parallel floating-point arithmetic. Evaluations of language modelslarge language model, n.: A computer program trained on enormous amounts of text to predict the words that come next. At vast scale, that simple skill lets it chat, answer questions, translate, summarize and write code. add prompt sensitivity and proprietary models that change or are retired without notice. Countermeasures include releasing code, data and weights, reporting several seeds with confidence intervals, conference reproducibility checklists, independent reproduction challenges, model cardsmodel card, n.: A short document published with an AI model explaining what it is for, how it was trained and tested, how well it works for different groups of people, and its known limits. and datasheetsdatasheet for datasets, n.: A standard document that travels with a dataset, describing why it was made, what is in it, how it was collected and labeled, and what it should and should not be used for., and deterministic or batch-invariant computation.
We could not reproduce the reported gain: across five seeds the new method was within noise of the tuned baseline.
Origin
An ideal of experimental science since at least the 17th century; the geophysicist Jon Claerbout championed 'reproducible research' for computational work around 1990. In machine learning, concern peaked around 2018: Odd Erik Gundersen and Sigbjørn Kjensmo found that few AI papers shared code, Peter Henderson and colleagues showed deep RL results swinging with random seeds, and Joelle Pineau's reproducibility checklist was adopted for NeurIPS 2019.
Usage note
Sayash Kapoor and Arvind Narayanan (2023) documented leakage-driven reproducibility failures in machine-learning-based science across 17 fields. Reproducing a result does not show that it is correct or general: a bug or a flawed benchmark reproduces perfectly.
Related
- random seedrandom seed, n.: A number that fixes the 'random' choices a computer makes; reuse the same seed and settings, and an image generator produces the same picture again.
- nondeterminismnondeterminism, n.: The fact that an AI model can give different answers to exactly the same question, partly because it deliberately chooses words with some randomness and partly because of tiny computing differences.
- baselinebaseline, n.: A simple reference method that a new model must beat to show it is worth anything, such as always guessing the most common answer or reusing yesterday's value.
- data leakagedata leakage, n.: A mistake in building a model that lets it peek at clues it will not have in real use, such as hints about the very answer it must predict, so it shines in testing and then disappoints.
- ablationablation, n.: Removing or switching off a part of an AI system to see what changes — whether an ingredient of an experiment's design or a single unit inside a trained network.
- model cardmodel card, n.: A short document published with an AI model explaining what it is for, how it was trained and tested, how well it works for different groups of people, and its known limits.
- Papers with CodePapers with Code, prop. n.: A free website, run by Meta from 2019 until it closed in 2025, that linked AI research papers to their computer code and kept leaderboards of the best results.
- benchmarkbenchmark, n.: A fixed, shared test — a set of questions or tasks with an agreed way of scoring — used to compare AI systems with one another and to track progress over time.
Sources
- Henderson, P., Islam, R., Bachman, P., Pineau, J., Precup, D. & Meger, D. (2018). Deep Reinforcement Learning that Matters. Proc. AAAI 2018.
- Pineau, J. et al. (2021). Improving Reproducibility in Machine Learning Research (A Report from the NeurIPS 2019 Reproducibility Program). Journal of Machine Learning Research 22.
- Kapoor, S. & Narayanan, A. (2023). Leakage and the Reproducibility Crisis in Machine-Learning-Based Science. Patterns 4(9).