> This is Goodhart's law at scale.

Also, Brandolini's law. And Adam Smith's law of supply and demand. When the ability to produce overwhelms the ability to review or refute, it cheapens the product.

> Number of released papers/number of citations is a target

There was this guy, well connected in the science world, that managed to publish a poor study quite high (PNAS level). It was not fraud, just bad science. There were dozens of papers and letters refuting his claims, highlighting mistakes, and so... Guess what? Attending to metrics (citations, don't matter if they are citing you to say you were wrong and should retract the paper!), the original paper was even more stellar on the eyes of grants and the journal itself.

It was rage bait before Facebook even existed.

There’s an accurate way to confirm fraud: look for inconsistencies and replicate experiments.

If the fraudsters “fail to replicate” legitimate experiments, ask them for details/proof, and replicate the experiment yourself while providing more details/proof. Either they’re running a different experiment, their details have inconsistencies, or they have unreasonable omissions.

Of course this is slightly messy too. Fraudsters are probably always incorrect, of course they could have stolen the data. But being incorrect doesn't mean your intentionally committing fraud.

That would be great if journals bothered publishing replication studies. But since they don't, researchers can't get adequate funding to perform them, and since they can't perform them, they don't exist.

We can't look for failed replication experiments if none exist.

That only confirms a very small subset of fraud. There are many ways to do scientific fraud that will yield internally consistent papers that pass replication as practiced today.

An example is papers which claims of the form, "We proved X by doing Y" where Y is a methodology that isn't derived from and can't prove X. This sort of paper will replicate every time because if you re-derive a correct methodology the original authors say you didn't really replicate their study and your work should be ignored, but if you use their broken methodology you'll just give an intellectually fraudulent paper the stamp of replication approval.

This kind of problem is actually much more widespread than work that looks scientific but in which the data is faked.

that approach is accurate, but not scalable.

the effort to publish a fraudulent study is less (sometimes much less) than the effort to replicate a study.

Is it that easy?

Machine Learning papers, for example, used to have a terrible reputation for being inconsistent and impossible to replicate.

That didn't make them (all) fraudulent, because that requires intent to deceive.

What do you think it is about machine learning that makes it hard to replicate? I'm an outsider to academic research, but it seems like computer based science would be uniquely easy - publish the code, publish the data, and let other people run it. Unless it's a matter of scale, or access to specific hardware.

A lot of things are easy if you ignore the incentive structure. E.g. a lot of papers will no longer be published if the data must be published. You’d lose all published research from ML labs. Many people like you would say “that’s perfectly okay; we don’t need them” but others prefer to be able to see papers like Language Models Are Few-Shot Learners https://arxiv.org/abs/2005.14165

So the answer is that we still want to see a lot of the papers we currently see because knowing the technique helps a lot. So it’s fine to lose replicability here for us. I’d rather have that paper than replicability through dataset openness.

But the lab must publish at least the general category of data, and if that doesn't replicate, then the model only works on a more specific category than they claim (e.g. only their dataset).

Even with the exact same dataset and architecture, ML results aren't perfectly replicable due to random weight initialisation, training data order, and non-deterministic GPU operations. I've trained identical networks on identical data and gotten different final weights and performance metrics.

This doesn't mean the model only works on that specific dataset - it means ML training is inherently stochastic. The question isn't 'can you get identical results' but 'can you get comparable performance on similar data distributions.

Then researchers should re-train their models a couple times, and if they can't get consistent results, figure out why. This doesn't even mean they must throw out the work: a paper "here's why our replications failed" followed by "here's how to eliminate the failure" or "here's why our study is wrong" is useful for future experiments and deserves publication.

As per my previous comment - we are discussing stochastic systems.

By definition, they involve variance that cannot be explained or eliminated through simple repetition. Demanding a 'deterministic' explanation for stochastic noise is a category error; it's like asking a meteorologist to explain why a specific raindrop fell an inch to the left during a storm replication.

Lack of will. That was one of the main results from the survey from Whitaker in 2020. Making your code reusable and easy to understand is significant work that had no direct benefits for a researcher's career. Particularly because research code grows wildly as researchers keep trying thungs.

Working on the next paper is seem as the better choice.

Moreover if your code is easy for others to run then you're likely to be hit with people wanting support, or even open yourself to the risk of someone finding errors in your code (the survey's result, not my own beliefs).

There are other issues, of course. Just running the code doesn't mean something is replicable. Science is replicated when studies are repeated independently by many teams.

There are many other failure modes SOTA-hacking, benchmarking, and lack of rigorous analysis of results, for example. And that's ignoring data leakage or other more silly mistakes (that still happen in published work! In work published in very good venues even)

Authors don't do much of anything to disabuse readers that they didn't simply get really look with their pseudorandom number generators during initialization, shuffling, etc. As long as it beats SOTA who cares if it is actually a meaningful improvement? Of course doing multiple runs with a decent bootstrap to get some estimation of the average behavior os often really expensive and really slow, and deadlines are always so tight. There is also the matter that the field converged on a experimentation methodology that isn't actually correct. Once you start reusing test sets your experiments stop being approximations of a random sampling process and you quickly find yourself outside of the grantees provided by statistical theory (this is a similar sort of mistake as the one scientists in other fields do when interpreting p-values). There be dragons out there and statistical demons might come to eat your heart or your network could converge to an implementation of nethack.

Scale also plays into that, of course, and use of private data as the other comment mentioned.

Ultimately Machine Learning research is just too competitive and moves too fast. There are tens of thousands (hundreds maybe?) of people all working on closely related problems, all rushing to publish their results before someone else published something that overlaps too much with their own work. Nobody is going to be as careful as they should, because they can't afford to. It's more profitable to carefully find the minimal publishable amount of work and do that, splitting a result into several small papers you can pump every few months. The first thing that tends to get sacrificed during that process is reliability.

Yeah, but this happens all the time.

>>95% of the time, the fraudsters get off scot-free. Look at Dan Ariely: Caught red-handed faking data in Excel using the stupidest approach imaginable, and outed as a sex pest in the Epstein files. Duke is still giving him their full backing.

It’s easy to find fraud, but what’s the point if our institutions have rotten all the way through and don’t care, even when there’s a smoking gun?

> Number of released papers/number of citations is a target

Only in stupid university leaderships is that truly what gets you hired or promoted. It's simply not true. Junior researchers in fact are believing it stronger than the facts actually support. Yes, you have to have a solid amount of publications, but doing a ridiculous amount of low-impact salami-sliced stuff or getting your name on a ton of papers where you did no real work is not going to win you a job. People who evaluate applications also live in this world and know that these metrics are being gamed. It's a cat and mouse game but the cats are also paying attention. You can only play this against really dumb government bureaucracies that mechanically give points for publications and have hard numerical criteria etc. Good institutions don't do that.

Good evaluators actually read the papers themselves. Of course you can't read the papers of every single applicant if there are many. But once the applicant gets into the a somewhat filtered down list, reading the paper(s) or having an interview about it, or having them give a talk is much more informative than the number of the papers. Still not perfect, because some people can't communicate well, but communicating is part of the job, so maybe that's super bad but somewhat bad.

Evaluators will use also other evidence such as recommendation letters (informally being aware of the reputation of the recommender), previous fellowships or grants obtained, etc.

None of these are foolproof in themselves. But someone who has super few publications relative to their career stage will need some other piece of evidence in favor.

In machine learning and AI, peer reviews are known to be quite random. If you have a good Arxiv-only paper that makes sense and you can give a good talk on it and answer questions, that will get you further than having a rubberstamp on some paper that's "meh, so what".

There are some players in this game (which includes funding agencies, journals, university administration, hiring committees, conference organizers, students, etc) that are more ossified and slow-moving than others.

And it's also true that double blind peer review and the rubberstamp of a top-tier conference was mostly beneficial to small, not well connected research groups, as it puts the paper on an equal footing with the big labs. The more this system erodes, to more we fall back to reputation and branding of big labs and famous researchers. Again, because there is no infinite time and infinite wisdom available to pick from applicants and there never will be. There are only tradeoffs.