New arXiv policy: 1-year ban for hallucinated references

https://twitter.com/tdietterich/status/2055000956144935055

by gjuggler • 2 months ago

659 230 2 months ago

imenani• 2 months ago

https://xcancel.com/tdietterich/status/2055000956144935055

Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated (Dieterrich, T. G.)

incognition• 2 months ago

coauthors about to get roasted

bigfishrunning4• 2 months ago

Good; academic literature is in crisis because of all of the slop. Forcing some consequences on easily-detectable hallucinations can only be a good thing

tengwar23• 2 months ago

It's not just AI, though. I did a doctorate in physics about 40 years back, and bad references were a problem back then.

lucb1e• 2 months ago

In what way? Surely something like the source not quite saying what was cited, or mixing up citations, rather than inventing them outright?

add-sub-mul-div• 2 months ago

Yes and ffs arrows kill people too but we don't bring that up every time we talk about what to do with guns.

dualvariable• 2 months ago

Doesn't matter if it is AI hallucinations or entirely human scientific fraud, the problem is the same, and the solution works fine for both cases.

If you can't validate that your bibliography is full of real articles, you shouldn't get published.

LLMs have just poured gasoline on the fire.

wrs3• 2 months ago

"Bad", like, you literally just made them up? I hope that would have been a problem.

asdff3• 2 months ago

Imagine how bad they are now then.

noobermin2• 2 months ago

Which is why the angry replies on Twitter from AI hype accounts is so funny. You should get penalised for fake references and profanity in your submissions, even if you wrote your slop longhand. I don't know why anyone would have an issue with this policy.

btown• 2 months ago

The penalty is a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue.

This is incredibly good for science. arXiv is free, but it's a privilege not a right!

I'm not seeing this clearly listed on https://info.arxiv.org/help/policies/index.html so it's possible this is planned but not live yet - or perhaps I'm not digging deeply enough?

As a certain doctor once said: the whole point of the doomsday machine is lost if you keep it a secret!

dataflow15• 2 months ago

This is incredibly good for science.

I disagree. It's just one darn hallucinated citation for heaven's sake, not fraud or something. It doesn't account for the substance or quality of their work at all. A one-year ban seems plenty sufficient for a minor first time mistake like this. People make mistakes and a good fraction of them can learn from those mistakes. There's no need to permanently cripple someone's ability to progress their life or contribute to humanity just because an AI hallucinated a reference one time in their life. That's punitive instead of rehabilitative.

ajkjk2• 2 months ago

It's not the kind of mistake that is possible unless you're engaging in fraud anyway.

wrs3• 2 months ago

A "mistake" would be a typo in a real citation. A hallucinated citation is evidence of just plain laziness and negligence, which taints the entire submission.

Loughla• 2 months ago

Don't use AI? Problem solved?

patcon• 2 months ago

A citation is where you derived knowledge... If you haven't checked it and you are submitting something that should represent a ton of labour (and which will consume labour to review), you don't understand what you're doing. It is not just crossing T's and dotting I'd.

Your being set behind is less important than the fact that your publishing is setting everyone else behind.

Such a banned person is being helped to "step out of the way", and someone more competent will assuredly step forward to consume the limited maintenance labour more thoughtfully

conartist6• 2 months ago

Yes, it is fraud

toast04• 2 months ago

It's just one darn hallucinated citation for heaven's sake, not fraud or something.

It is fraud.

It doesn't account for the substance or quality of their work at all.

References are part of the work. If you're making up the references, what else are you making up?

People make mistakes and a good fraction of them can learn from those mistakes. There's no need to permanently cripple someone's ability to progress their life or contribute to humanity just because an AI hallucinated a reference one time in their life.

A one year ban is not permanent. Having a negative consequence for making poor decisions seems like an inducement to learn from the mistake?

In an ideal world, one would be keeping notes on references used while doing the research that lead to writing the paper. Choosing not to do that is one poor decision.

Having a positive outlook, if asking an AI to provide references that may have been missed, one should at least verify the references exist and are relevant. Choosing not to do that is also a poor decision, even if one did take notes on references used while researching.

mianos• 2 months ago

You are being ironic right?

goolz• 2 months ago

If you cannot be bothered to check your references when writing academic quality papers then you have no place writing them in the first place. The punishment is not chopping off a finger, it is a polite reminder to do the bare minimum.

themafia• 2 months ago

There's no need to permanently cripple someone's ability to progress their life or contribute to humanity

I don't think you need to publish on arXive to contribute meaningfully to humanity.

That's punitive instead of rehabilitative.

Unfortunately science is competitive. Yours is a race to the bottom where the people who can afford the most expensive models and who are least concerned with the truth can publish the most papers and benefit financially and professionally by doing so. This is not a zero sum arena, grant money and opportunities will possibly be rewarded to them, and not to another team who is producing more careful and genuine output.

vhantz• 2 months ago

What's the difference between a "hallucinated" citation and consciously inserting reference to a non-existent paper and hopping it goes unnoticed? How do we determine which one was done consciously and which was "a minor first time mistake"?

Your standards are lower than what they would accept at my high-school. Seriously.

And generally, if you are generating papers with LLMs, let other LLMs read them. Why would we waste human hours considering something that was generated? At this point publish your prompt because that's the actual work you're doing.

simianparrot• 2 months ago

In science, one hallucinated reference can corrupt the entire rest of the work. So you're completely wrong.

nkrisc2• 2 months ago

It’s easy to avoid this whole issue: write the paper yourself.

avanwyk• 2 months ago

We got pretty far without a single hallucinated citation (because they simply didn't exist). It's really simple: check your references, or don't bother of publishing. It's not a high hurdle.

paulpauper2• 2 months ago

My take: this seems excessive.

ArXiv doesn't even check the submission closely, so how can they know?

They say "errors, mistakes"

They use an automated system to check if the basic requirements were met, and sometimes papers are flagged for further superficial human review, but there is no way they can possibly do this at scale or check every reference. This would be like trying to do peer review, but for a preprint archive that gets easily 100x more volume than any journal.

Second, there is such a huuuuge gap between publishing on arvix and peer review. I can attest personally that it's not even close. I've gotten probably dozen rejections from peer review and no problems publishing in arxiv math. This is because peer review checks not just for if something is new or correct, but also if it's of "interest to math community," which is inherently subjective, but also makes peer review many magnitudes harder than publishing on arxiv.

Even when a well-known professor in number theory praised the paper when I got an endorsement and a second emailed me and and encouraged me to publish it, it still got rejected 3 times and still waiting.

Being required to publish in a peer reviewed journal will close off arxiv for many researchers for good. It also defeats the point of it being a pre-print.

helterskelter• 2 months ago

You could at least filter out hallucinated references which simply don't exist pretty trivially, I'd imagine.

eqvinox3• 2 months ago

You don't need to be actively enforcing a rule 100% on everyone. Speed cameras don't cover every stretch of road either.

It's enough for them to place this policy and enforce it when they become aware of violations. Someone reading the slopped paper (or, here, trying to follow a reference) will notice sooner or later.

Being required to publish in a peer reviewed journal will close off arxiv for many researchers for good. It also defeats the point of it being a pre-print.

You sound like it's impossible for researchers to write papers without slopped references, and inevitable to get hit by this policy.

rossjudson2• 2 months ago

This puts the burden to make sure it's right on the submitter, where it should be. Verification can come at any time after that; the submitter understands the consequences of hallucinated references. Verification can be crowd-sourced (and likely will be).

Nothing stops someone from putting a PDF on the internet. I'm fine with ArXiv holding a high standard.

nimih• 2 months ago

ArXiv doesn't even check the submission closely, so how can they know?

They can be informed by people who read the papers and check the citations. A zero-tolerance policy provides an incentive to report sloppy papers (namely, that you can be confident something will be done about it), and each time a paper is removed or an author is banned, it incrementally increases the value of the arXiv as a whole.

Being required to publish in a peer reviewed journal will close off arxiv for many researchers for good.

At the end of the day, demanding that people carefully proofread their LLM-generated papers before sharing them on the arXiv seems like a relative low bar to clear, and I sort of question whether it's reasonable to call individuals who find it too onerous "researchers" in the first place.

nomel3• 2 months ago

I bet, since this has been posted, someone here has already vibe coded a reference checker that they plan to put behind a subscription.

This is good for reference checking, but I doubt this will do much for the most likely shoddy science that accompanies hallucinated references.

nradov2• 2 months ago

The frontier LLMs are getting pretty good at checking this sort of thing. You could prompt them to not only verify the references are real but that they actually state what the article claims. Some human review will still be needed but I'll bet this approach could find a lot of academic fraud.

____mr____• 2 months ago

subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue. Isn't the whole point of arxiv to be somewhere you post your low-reviewed/unpublished/prepublished papers? doesnt this somewhat defeat the point and basically means a lifetime ban?

purerandomness• 2 months ago

If it isn't, it absolutely should be.

The penalty cannot be high enough.

jkman• 2 months ago

Eh guess so, I think it's appropriate

kordlessagain• 2 months ago

It's a gatekeepers dream, for sure.

dan-g• 2 months ago

Sorry maybe I’m missing something, but how is this gatekeeping?

bigfishrunning4• 2 months ago

I feel like requiring that your cited references are real is a pretty low bar

T-A• 2 months ago

> The penalty is a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue.

Catch-22 [1]:

You will need to provide an arXiv article id. This is necessary for your paper to be processed correctly, the submitted version must be the same as the arXiv version.

[1] https://jcap.sissa.it/jcap/help/helpLoader.jsp?pgType=author

chadgpt33• 2 months ago

Plausible deniability on banning the person from all science forever.

Are all venues like this, however?

_aavaa_• 2 months ago

That’s the journal’s doing, not arxiv’s.

random3• 2 months ago

It seems a good idea to ban cheating, but how hard is it, especially in new reasoning/agents contexts to validate references?

The deeper question is whether legitimate AI generated results are allowed or not? Test - In the extreme - think proof of Riemann Hypothesis autonomously generated (end to end) formally proven - is it allowed or not?

Ifkaluva• 2 months ago

This is not about banning cheating, it’s about banning inaccurate information.

Retric• 2 months ago

You don’t need to solve everything, catching a few thousand non existent citations with such a policy is on its own a net benefit.

pointlessone• 2 months ago

It is allowed as long as it’s verified.

The thread specifically points out that if authors can’t be arsed to simply proofread their text the rest can not be trusted either.

It’s a simple heuristic against low quality submissions, not an anti-ai measure.

pinkmuffinere2• 2 months ago

think proof of Riemann Hypothesis autonomously generated (end to end) formally proven - is it allowed or not?

Sorry to be rude, but this seems like a dumb question. I want science to progress. A primary purpose of these journals is to progress science. A full proof of the Riemann Hypothesis progresses science. I don't care how it was produced, if Hitler is coauthor, etc, I just care that it is correct. Whether the authors should be rewarded for whatever methods they used can be a separate question.

kingstnap• 2 months ago

Terence Tao had a nice talk from the Future of Mathematics conference posted yesterday [0] that shapes a lot of my own feelings on this matter.

The short of it is he argues how first to correctness shouldn't be the only goal / isn't a great optimisation incentive. Presentation and digestibility of correct results is a missing 1/3 when you've finished generation and verification. I completely agree with him. You don't just need an AI generated proof of the Reimann Hypothesis. You would really like it to be intentional and structured for others to understand.

A really beautiful quote I learned of in the talk is this:

"We are not trying to meet some abstract production quota of definitions, theorems, and proofs. The measure of our success is whether what we do enables people to understand and think more clearly and effectively about math." - William Thurston

[0] https://www.youtube.com/watch?v=Uc2zt198U_U

lionkor• 2 months ago

If you use AI correctly, nobody should be able to tell that it was used at all.

beloch• 2 months ago

There already exists multiple tools for automatically verifying references. This measure will likely only filter out the laziest and most incompetent of AI slop submissions. It's a very modest raising of the bar, but comes at zero cost to honest researchers.

I expect arXiv will still have problems with slop submissions but, at least, their references should actually exist going forward.

llm_nerd• 2 months ago

It isn't "cheating" they're concerned with, it's sloppiness. This dictum isn't some sort of AI ban, but instead simply that if there is evidence that it was so low effort that the work includes such blatant problems, it's just adding noise.

david_shi• 2 months ago

In that case, you would just not do a reference. End to end autonomous science might have fewer concrete citations as the contributing knowledge is just the sum of the training data of the model.

squirrelon• 2 months ago

Had a colleague submit a paper with literal AI slop left in the text, got hit with a nasty revision request. Check your drafts before you submit, people. The reviewers will find it.

miki1232112• 2 months ago

Also check your LaTeX comments, Arxiv makes those publicly visible!!!

I'm a screen reader user and usually read papers as raw TeX. I've seen everything: slurs, demeaning comments towards reviewers and professors, admissions of fraud, instructions to coauthors to commit further fraud before paper submission to mask the earlier fraud... it's all there. There's far less of it than I would think, definitely <1% of papers, but it's there.

I think it would be useful to run an LLM anti-fraud pass on the TeX source of all new arxiv papers. It wouldn't catch everything, but it would catch some of the dumbest fraudsters.

On the positive side, you can also find stronger claims that didn't survive review, additional explanations that didn't make the cut due to the conference's page limit, as well as experimental results that the authors felt weren't really worth including. Those need to be approached with an abundance of caution, but are genuinely useful sometimes.

andrepd4• 2 months ago

https://xcancel.com/leaksph

cozzyd2• 2 months ago

That's why my forarxiv make target includes a run through latexpand

SchemaLoad• 2 months ago

Sad the suggestion here is to just disguise the slop to make it harder for reviewers to spot rather than not submitting slop to begin with.

MinimalAction• 2 months ago

There needs be to a careful vetting before such adverse actions. If somebody includes a name and pushed it without express permission, does everyone get the ban? I agree that implemented the right way, this is good.

vasco• 2 months ago

Plus afaik you can add any co-author you want without validation. So you can ban everyone on arxiv with one paper with one sentence.

lou13064• 2 months ago

As I mentioned in another thread:

To be a coauthor on a preprint that you have not submitted, you have to actively "claim" it (using a password given to the author who submitted). It's on you to double-check before claiming.

I surely hope that only "confirmed" coauthors will get the ban, it's only logical.

noobermin2• 2 months ago

Seeing the usual LLM hypers angry replying to this on twitter is such a tell. Just like the comments on the LLM poisoning articles, some people just can't accept that some people don't like LLMs and get upset when you put any amount of hindrance to their rapid acceptance.

lbrito2• 2 months ago

Crazy that this is graytexted. So basically HN consensus is that we need to be hyper and accelerate llm adoption everywhere.

Bonkers. At the same time peak hn

ethin2• 2 months ago

It's also hilarious that they complain about this because, from what I've seen, most LLM hypers will talk about something being irrelevant or taken over by AI with no understanding of what that something really is or involves.

lou13064• 2 months ago

some people don't like LLMs

It's not even that they "don't like LLMs". They just don't like academic fraud! If references were fabricated with a Markov chain it would be just as bad!

tdeck• 2 months ago

It's hard for me to even understand their perspective. Researching references for a published academic paper isn't some incidental busywork task, it's supposed to be a core part of doing research which is the core of the job. If you don't have sympathy for someone who, say, paid a person on Fiverr to cook up a paper rather than writing it themselves and then didn't even bother to check the references, why is using an LLM and not checking any better?

boccaff• 2 months ago

There is a lot of "throw it against the wall, and if it sticks, write it up" empirical work against benchmarks. It leads to post-hoc rationalization of the work and browser plugins using LLMs to find references for work that is already written. It is a bureaucratic view about "you need a citation for this", where people misunderstand the citation as a checkbox, instead of "you need to substantiate this claim, as I, the reviewer, do not accept this as a fact".

user34283• 2 months ago

While this arXiv policy seems reasonable enough, I don't care for the kind of drivel some post on HN because they don't like LLMs.

I'm here because I enjoy building things. And today this mostly happens with AI. I could do without the often thoughtless comments and conspiracy theories about "LLM hypers" posted by people who don't like LLMs.

jszymborski• 2 months ago

Should be more harsh in my opinion.

rgmerk2• 2 months ago

Good.

If it’s not worth your time to check the output of your LLM carefully, it’s not worth my time to read it.

djoldman• 2 months ago

Unfortunately, it's probably not worth your time to read 99% of arxiv papers, LLM generated or otherwise.

Ever pick a random one and really dive in?

rgmerk2• 2 months ago

Agreed. There was already too much human generated slop in academia.

And I’m not talking about good faith research that didn’t pan out, I mean research that is completely useless for any other purpose other than convincing a casual observer that the authors are doing research.

gammalost• 2 months ago

Well, yeah, 99% of arXiv papers were not written for me or you. They were written for someone who works in a niche within a niche. That's (in my view) the beauty of research.

nullc• 2 months ago

It's been pretty eye opening watching Craig Wright (of bitcoin fakery fame) flooding out LLM generated 'academic' papers and even having some of them accepted.

He's toast if SSRN were to adopt a similar policy.

mks_shuffle• 2 months ago

While this is certainly a welcome step, I hope there is more work done to fix the underlying problem of easily creating correct BibTeX entries for the cited papers. Citations for any given paper can come from a wide range of journals with various publishers, conferences, and preprints. The same paper can be available from multiple sources with varying details, e.g. arXiv and the conference website. Tools like Zotero have certainly made it significantly easier to extract citations from webpages of publication, but I still find issues with the extracted BibTeX details. While author names and titles are often extracted correctly, I still have to manually ensure that details like publication venue, year, volume number, page number, URL, etc. are extracted correctly and also shown correctly in LaTeX format. Different publications can use different citation styles. This can unfortunately lead to taking shortcuts with AI-generated citation data due to the lack of an easy and unified approach to extract consistent citation data. I am not sure whether hallucinated citations are being generated in the main manuscript or in a separate BibTeX file, so I may be a bit off in my understanding.

flexagoon2• 2 months ago

Note that Zotero also has a free online tool to generate citations in any format or BibTeX files from a URL/DOI/ISBN/...

https://zbib.org/

gucci-on-fleek• 2 months ago

Fun fact: if an article has a DOI, you can just use curl to get a BibTeX entry. An example using one of my articles:

  $ curl -L "https://doi.org/10.47397/tb/43-1/tb133chernoff-widows" -H 'Accept: application/x-bibtex'
  @article{Chernoff_2022, title={Automatically removing widows and orphans with <tt>lua-widow-control</tt>}, volume={43}, ISSN={0896-3207}, url={http://dx.doi.org/10.47397/tb/43-1/tb133chernoff-widows}, DOI={10.47397/tb/43-1/tb133chernoff-widows}, number={1}, journal={TUGboat}, publisher={TeX Users Group}, author={Chernoff, Max}, year={2022}, pages={28–39} }

This is the exact same method that Zotero uses internally, so this won't ever give you better results, but I still find it kinda neat.

jimmygrapes• 2 months ago

As of yet no comments here seem to address the "reputable" condition. Reputable review is based on what criteria?

ElenaDaibunny• 2 months ago

how will they detect hallucinated refs at scale? Manual spot checks? Automated DOI verification? The policy seems right but enforcement is the hard part.

fc417fc8028• 2 months ago

However difficult it might be right now it's only going to get easier. Anyway I don't think proactive enforcement is the point. Rather now they have an official method by which to address incidents that are brought to their attention.

ddosmax556• 2 months ago

Enforcement is secondary and is allowed to take weeks / months / never at all if nobody reads the paper. It's about being able to ban if an issue arrises; not about keeping the database strictly clean.

druub• 2 months ago

What are reasonable alternatives to arXiv? It has become increasinbgly slow. Techrxiv?

az2262• 2 months ago

Next, for AI papers, a reproducibility requirement. So much code and details are fudged and paper's cannot be reproduced. Ran the training with some other config, or other data, etc. to make their mechanism or intervention seem better.

az2262• 2 months ago

Hurray!

cyclecycle2• 2 months ago

This has become such a problem in scholarly publishing that we have a business that provides citation checking https://groundedai.company/ that we've been buidling for a couple of years now

gizajob4• 2 months ago

What’s the hallucination rate of your AI?

cyclecycle2• 2 months ago

So far we basically just provide a very rule-based approach and try not use LLMs as much as possible. So we extract and parse the citations using various ML and rule-based approaches, and carry out a bunch of predetermined queries and do various fuzzy matching approaches on the metadata components, and have a bunch of rules around risk levels of things we should have found/matched based on what type of source it is, which venue we should have found it in, etc.

So there are absolutely a bunch of tasks that could be evaled/benchmarked, but "hallucination rate" isn't particularly applicable/interesting as a metric of how good the tool is

that said, we do use various LLMs (mostly local, fine-tuned, small, for things like NER/parsing/metadata comparison, etc.). and they can and do hallucinate, but we have very hard constraints on the validation, so any extraction results that don't match 1:1 back to the input text are discarded for example. so again, rather than hallucination risk we prefer hard constraints

thatjoeoverthr• 2 months ago

No mercy to brain slugs.

Foivos• 2 months ago

I just wish to anyone who is against this policy to be forced to review a paper that turns out to be unedited AI slop. Reviewers are experts volunteers who do it for free. It is incredibly frustrating to have spent 4 hours reading a paper where you try your best to make sense of what the authors are trying to prove just to realize that it is hallucinations.

The authors should value the time of the reviewers higher than their own time. So, if you include AI nonsense in your paper, it is insulting.

scirob• 2 months ago

Great, it's so easy to automate checking ref super bad to not check

soraminazuki• 2 months ago

It's not unexpected, but still sad to see so many comments opposing even the smallest step against low-effort fraud in academic publications. Is this what hacker culture has been reduced to in the age of the slop era? Open hostility against science and engineering?

skydhash2• 2 months ago

Something something about not understanding the problem when your salary depends on it…

clearstack2• 2 months ago

the hallucination problem is especially bad in financial AI. one made-up revenue number and someone makes a bad investment. the actual 10-K is the only source of truth for company data

kevincox• 2 months ago

That sounds like way less of a problem than science. If you take bad sources you hurt yourself financially. For science (especially anything medicine related) you can hurt many others physically.

clearstack2• 2 months ago

fair at the individual level. but now these tools are used in quant models and at retail scale — one systematically hallucinated metric can affect a lot of real portfolios

qarl• 2 months ago

I don't understand.

I have standing instructions with my agents: do not commit anything without review by at least two sub agents.

The chance of two agents hallucinating at the same time in the same way is almost nil. Why do people still do it?

vwkd• 2 months ago

Previously: „Generate a bibliography.“

Now: „Generate a bibliography. Make no mistakes.“

bigfishrunning4• 2 months ago

"You are an expert research assistant that never makes mistakes. Please write a..."

lbrito2• 2 months ago

As another comment said (not sure if sarcastically or not - impossible to tell): "just spin up another agent(s) to check the work of the first!"

VerifiedReports• 2 months ago

Fabricated, not "hallucinated."

kramit1288• 2 months ago

thats good but difficult to identify if its pure AI generated or not. identifying itself can be hallicunation.

New arXiv policy: 1-year ban for hallucinated references

Related Stories