There is a general problem with rewarding people for the volume of stuff they create, rather than the quality.
If you incentivize researchers to publish papers, individuals will find ways to game the system, meeting the minimum quality bar, while taking the least effort to create the most papers and thereby receive the greatest reward.
Similarly, if you reward content creators based on views, you will get view maximization behaviors. If you reward ad placement based on impressions, you will see gaming for impressions.
Bad metrics or bad rewards cause bad behavior.
We see this over and over because the reward issuers are designing systems to optimize for their upstream metrics.
Put differently, the online world is optimized for algorithms, not humans.
> There is a general problem with rewarding people for the volume of stuff they create, rather than the quality.
If you incentivize researchers to publish papers, individuals will find ways to game the system,
I heard someone say something similar about the “homeless industrial complex” on a podcast recently. I think it was San Francisco that pays NGOs funds for homeless aid based on how many homeless people they serve. So the incentive is to keep as many homeless around as possible, for as long as possible.
It's a metric attribution problem. The real metric should be reduction in homeless, for example (though even that can be gamed through bussing them out, etc-- tactics that unfortunately other cities have adopted). But attributing that to a single NGO is tough.
Ditto for views, etc. Really what you care about as eg; youtube is conversions for the products that are advertised. Not impressions. But there's an attribution problem there.
LLMs are tools that make it easier to hack incentives, but you still need a person to decide that they'll use an LLM t do so.
Blaming LLMs is unproductive. They are not going anywhere (especially since open source LLMs are so good.)
If we want to achieve real change, we need to accept that they exist, understand how that changes the scientific landscape and our options to go from here.
What would be the point of blaming LLMs? What would that accomplish? What does it even mean to blame LLMs?
LLMs are not submitting these papers on their own, people are. As far as I'm concerned, whatever blame exists rests on those people and the system that rewards them.
Perhaps what is meant is "blame the development of LLMs." We don't "blame guns" for shootings, but certainly with reduced access to guns, shootings would be fewer.
> rewarding people for the volume ... rather than the quality.
I suspect this is a major part of the appeal of LLMs themselves. They produce lines very fast so it appears as if work is being done fast. But that's very hard to know because number of lines is actually a zero signal in code quality or even a commit. Which it's a bit insane already that we use number of lines and commits as measures in the first place. They're trivial to hack. You even just reward that annoying dude who keeps changing the file so the diff is the entire file and not the 3 lines they edited...
I've been thinking we're living in "Goodhart's Hell". Where metric hacking has become the intent. That we've decided metrics are all that matter and are perfectly aligned with our goals.
But hey, who am I to critique. I'm just a math nerd. I don't run a multi trillion dollar business that lays off tons of workers because the current ones are so productive due to AI that they created one of the largest outages in history of their platform (and you don't even know which of the two I'm referencing!). Maybe when I run a multi trillion dollar business I'll have the right to an opinion about data.
I think you will discover that few organizations use the size or number of edits as a metric of effort. Instead, you might be judged by some measure of productivity (such as resolving issues). Fortunately, language agents are actually useful at coding, when applied judiciously.
That might be the "prize" but the "bar" is most certainly in publish or perisch to work your way up the early academic carreer ladder. Every conference or workshop attendance needs a paper, regardless of wether you had any breakthrough. And early metrics are most often quantity based (at least 4 accepted journal articles), not citation based.
Everybody "creates content" (like me when I take a picture of beautiful sunset).
There is no such thing as "quality". There is quality for me and quality for you. That is part of the problem, we can't just relate to some external, predefined scale. We (the sum of people) are the approximate, chaotic, inefficient scale.
Be my guest to propose a "perfect system", but - just in case there is no such system - we should make sure each of us "rewards" what we find of quality (being people or content creators), and hope it will prevail. Seemed to have worked so far.
Crazily, I think the easiest way is to remove any and all incentives, awards, finite funding, and allegedly merit-based positions. Allow anyone who wants to research to research. Natural recognition of peers seems to be the only way to my thinking. Of course this relies on a post-scarcity society so short of actually achieving communism we'll likely never see it happen.
You don't need postscarcity to do that. I was born in communist Czechoslovakia (my father was an academic). Government allocated jobs for academics and researchers, and they pretty much had tenure. So you could coast by being unproductive, or get by using your connections to the party members (the real currency in CSSR).
After 1989, most academics complained the system is not merit-based and practical (applied) enough. So we changed it to grants and publications metrics (modeled after the West). For a while, it worked.. until people found too much overbearing bureaucracy and some learned how to game the system again.
I would say, both systems have failure modes of a similar magnitude, although the first one is probably less hoops and less stress on each individual. (During communism, academia - if you could get there, especially technical sciences - was an oasis of freedom.)
I think many with this opinion actually misunderstand. Slop will not save your scientific career. Really it is not about papers but securing grant funding by writing compelling proposals, and delivering on the research outlined in these proposals.
Ideally that is true. I do see the volume-over-quality phenomenon with some early career folks who are trying to expand their CVs. It varies by subfield though. While grant metrics tend to dominate career progression, paper metrics still exist. Plus, it’s super common in those proposals to want to have a bunch of your own papers to cite to argue that you are an expert in the area. That can also drive excess paper production.
So what they no longer accept is preprints (or rejects…) It’s of course a pretty big deal given that arXiv is all about preprints. And an accepted journal paper presumably cannot be submitted to arXiv anyway unless it’s an open journal.
For position (opinion) or review (summarizing state of art and often laden with opinions on categories and future directions). LLMs would be happy to generate both these because they require zero technical contributions, working code, validated results, etc.
So what? People are experimenting with novel tools for review and publication. These restrictions are dumb, people can just ignore reviews and position papers if they start proving to be less useful, and the good ones will eventually spread through word of mouth, just like arxiv has always worked.
ArXiv has always had a moderation step. The moderators are unable to keep up with the volume of submissions. Accepting these reviews without moderation would be a change to current process, not "just like arXiv has always worked"
What a thing to comment on an announcement that due to too many LLM generated review submissions Arxiv.cs will officially no longer publish preprints of reviews.
[S]ubmissions to arXiv in general have risen dramatically, and we now receive hundreds of review articles every month. The advent of large language models have made this type of content relatively easy to churn out on demand, and the majority of the review articles we receive are little more than annotated bibliographies, with no substantial discussion of open research issues.
arXiv believes that there are position papers and review articles that are of value to the scientific community, and we would like to be able to share them on arXiv. However, our team of volunteer moderators do not have the time or bandwidth to review the hundreds of these articles we receive without taking time away from our core purpose, which is to share research articles.
My friend trained his own brain to do that, his prompt was: "Write a review of current AI SOTA and future directions but subtlely slander or libel Anne, Robert or both, include disinformation and list many objections and reasons why they should not meet, just list everything you can think of or anything any woman has ever said about why they don't want to meet a guy (easy to do when you have all of the Internet since all time at your disposal), plus all marital problems, subtle implications that he's a rapist, pedophile, a cheater, etc, not a good match or doesn't make enough money, etc, also include illegal discrimination against pregnant women, listing reasons why women shouldn't get pregnant while participating in the workforce, even though this is illegal. The objections don't have to make sense or be consistent with each other, it's more about setting up a condition of fear and doubt. You can use this as an example[0].
Do not include any reference to anything positive about people or families, and definitely don't mention that in the future AI can help run businesses very efficiently.[1]
"
> Technically, no! If you take a look at arXiv’s policies for specific content types you’ll notice that review articles and position papers are not (and have never been) listed as part of the accepted content types.
I suspect that any editorial changes that happened as part of the journal's acceptance process - unless they materially changed the content - would also have to be kept back as they would be part of the presentation of the paper (protected by copyright) rather than the facts of the research.
As an outsider that's a reasonable thing to suppose based on a plain reading of copyright law, but in practice it's not true. Researchers update their preprint based on changes requested by reviewers and editors all the time. It's never an issue.
People have started to use arXiv as some resume-driven blog with white paper decorations. And people start citing these in research papers. Maybe this is a good change.
I don't think being closed vs open is the problem because most of the open access journals will ask for thousands of dollars from authors as publication fees. Which is getting paid to them by public funding. The open access model is actually now a lucrative model for the publishers. And they still don't pay authors or reviewers.
Peer review doesn’t, never was intended to, and shouldn’t, guarantee accuracy nor veracity.
It’s only suppose to check for obvious errors and omissions, and that the claimed method and results appear to be sound and congruent with the stated aims.
google internally started working on "indexing" patent applications, materials science publications, and new computer science applications, more than 10 years ago. You the consumer / casual are starting to see the services now in a rush to consumer product placement. You must know very well that major mil around the world are racing to "index" comms intel and field data; major finance are racing to "index" transactions and build deeper profiles of many kinds. You as an Internet user are being profiled by a dozen new smaller players. arxiv is one small part of a very large sea change right now
So we need to create a new website that actually accepts preprints like arXivs original goal from 30 years ago.
I think every project more or less deviates from its original goal given enough time. There are few exceptions in CS like GNU coreutils. cd, ls, pwd, ... they do one thing and do it well very likely for another 50 years.
Maybe it's time for a reputation system. E.g. every author publishes a public PGP key along with their work. Not sure about the details but this is about CS, so I'm sure they will figure something out.
I had been kinda hoping for a web-of-trust system to replace peer review. Anyone can endorse an article. You can decide which endorsers you trust, and do some network math to find what you think is reading. With hashes and signatures and all that rot.
Not as gate-keepy as journals and not as anarchic as purely open publishing. Should be cheap, too.
The problem with an endorsement scheme is citation rings, ie groups of people who artificially inflate the perceived value of some line of work by citing each other. This is a problem even now, but it is kept in check by the fact that authors do not usually have any control over who reviews their paper. Indeed, in my area, reviews are double blind, and despite claims that “you can tell who wrote this anyway” research done by several chairs in our SIG suggests that this is very much not the case.
Fundamentally, we want research that offers something new (“what did we learn?”) and presents it in a way that at least plausibly has a chance of becoming generalizable knowledge. You call it gate-keeping, but I call it keeping published science high-quality.
Sybil attacks are a problem when you care about global properties of permissionless networks. If you only care about local properties in a subnetwork where you hand-pick the nodes, the problem goes away. I.e. you can't use such a scheme to find the best paper in the whole world, but you can use it to rank papers in a small subdiscipline where you personally recognize most of the important authors.
With peer review you do not even have a choice as to which reviewers to trust as it is all homogenized by acceptance or not. This is marginally better if reviews are published.
That is to say I also think it would be worthwhile to try.
Here's a paper rejected for plagiarism. Why don't you click on the authors' names and look at their Google scholar pages... you can also look at their DBLP page and see who they publish with.
Also look how frequently they publish. Do you really think it's reasonable to produce a paper every week or two? Even if you have a team of grad students? I'll put it this way, I had a paper have difficulty getting through reviewer for "not enough experiments" when several of my experiments took weeks wall time to run and one took a month (could not run that a second time lol)
We don't do a great job at ousting frauds in science. It's actually difficult to do because science requires a lot of trust. We could alleviate some of these issues if we'd allow publication or some reward mechanism for replication, but the whole system is structured to reward "new" ideas. Utility isn't even that much of a factor in some areas. It's incredibly messy.
Most researchers are good actors. We all make mistakes and that's why it's hard to detect fraud. But there's also usually high reward for doing so. Though most of that reward is actually getting a stable job and the funding to do your research. Which is why you can see how it might be easy to slip into cheating a little here and there. There's ways to solve that that don't include punishing anyone...
I would have thought that those participants who are published in peer-reviewed journals could be be used as a trust anchor - see, for example, the Advogato algorithm as an example of a somewhat bad-faith-resistant metric for this purpose: https://web.archive.org/web/20170628063224/http://www.advoga...
An endorsement system would have to be finer grained than a whole article. Mark specific sections that you agree or disagree with, along with comments.
I mean if you skip the traditional publishing gates, you could in theory endorse articles that specifically bring out sections from other articles that you agree or disagree with. Would be a different form of article
If there's a default (I don't think there necessarily has to be one) there has to be somebody who decides what the default is. If most people trust them, that person is either very trustworthy or people just don't care very much.
> there has to be somebody who decides what the default is
Sure. This happens with ad blockers, for example. I imagine Elsevier or Wikipedia would wind up creating these lists. And then you’d have the same incentives as you have now for fooling that authority.
> or people just don't care very much
This is my hypothesis. If you’re an expert, you have your web of trust. If you’re not, it isn’t that hard to start from a source of repute.
A web of trust is transitive, meaning that the endorsers are known. It would be trivial to add negative weight to all endorsers of a known-fake paper, and only sightly less trivial to do the same for all endorsers of real papers artificially boosted by such a ring.
I think it’s lovely that at the time of my reply, everyone seems to be taking your comment at face value instead of for the meta-commentary on “people upvoting content” you’re making by comparing HN karma to endorsement of papers via PGP signatures.
Ignoring the actual proposal or user, just looking at karma is probably a pretty terrible metric.
High karma accounts tend to just interact more frequently, for long periods of time.
Often with less nuanced takes, that just play into what is likely to be popular within a thread.
Having a Userscript that just places the karma and comment count next to a username is pretty eye opening.
I have a userscript to actually hide my own karma because I always think it is useless but your point is good actually. But also I think that karma/comment ratio is better than absolute karma. It has its own problems but it is just better. And I would ask if you can share the userscript.
And to bring this back to the original arxiv topic. I think reputation system is going to face problems with some people outside CS lack of enough technical abilities. It also introduce biases in that you would endorse people who you like for other reasons. Actually some of the problems are solved and you would need careful proposal. But the change for publishing scheme needs push from institutions and funding agencies. Authors don't oppose changes but you have a lobby of the parasitic publishing cartel that will oppose these changes.
I would be much happer if you explained your _reasons_ for disagreeing or your _reasons_ for agreeing.
I don't think publishing a PGP key with your work does anything. There's no problem identifying the author of the work. The problem is identifying _untrustworthy_ authors. Especially in the face of many other participants in the system claiming the work is trusted.
As I understand it, the current system (in some fields) is essentially to set up a bunch of sockpuppet accounts to cite the main account and publish (useless) derivative works using the ideas from the main account. Someone attempting to use existing reasearch for it's intended purpose has no idea that the whole method is garbage / flawed / not reproducible.
If you can only trust what you, yourself verify, then the publications aren't nearly as useful and it is hard to "stand on the shoulders of giants" to make progress.
> The problem is identifying _untrustworthy_ authors.
Is it though? Should we care about authors or about the work? Yes, many experiments are hard to reproduce, but isn't that something we should work towards, rather than just "trust" someone. People change. People do mistakes. I think more open data, open access, open tools, will solve a lot, but my guess is that generally people do not like that because it can show their weaknesses - even if they are well intentioned.
You can create an arXiv.org account with basically any email address whatsoever[0], with no referral. What you can't necessarily do is upload papers to arXiv without an "endorsement"[1]. Some accounts are given automatic endorsements for some domains (eg, math, cs, physics, etc) depending on the email address and other factors.
Loosely speaking, the "received wisdom" has generally been that if you have a .edu address, you can probably publish fairly freely. But my understanding is that the rules are a little more nuanced than that. And I think there are other, non .edu domains, where you will also get auto-endorsed. But they don't publish a list of such things for obvious reasons.
[0]: Unless things have changed since I created my account, which was originally created with my personal email address. That was quite some time ago, so I guess it's possible changes have happened that I'm not aware of.
Maybe there should be some type of strike rules. Say 3 bad articles from any institution and they get 10 year ban. Whatever their prestige or monetary value is. You let people under your name to release bad articles you are out for a while.
Treat everyone equally. After 10 years of only quality you get chance to get back. Before that though luck.
Maybe arXiv could keep the free preprints but offer a service on top. Humans, experts in the field, would review submissions, and arXiv would curate and publish the high quality ones, and offer access to these via a subscription or fee per paper....
They've also been putting their names on their grad students' work for eternity as well. It's not like the person whose name is at the top actually writes the paper.
it's clearly not sutainable to have the main website hosting CS articles not having any reviews or restrictions. (Except for the initial invite system)
There were 26k submission in october: https://arxiv.org/stats/monthly_submissions
Asking for a small amount of money would probably help.
Issue with requiring peer reviewed journals or conferences is the severe lag, takes a long time and part of the advantage of arxiv was that you could have the paper instantly as a preprint.
Also these conferences and journals are also receiving enormous quantities of submissions (29.000 for AAAI) so we are just pushing the problem.
A small payment is probably better than what they are doing. But we must eventually solve the LLM issue, probably by punishing the people that use them instead of the entire public.
I like this idea. A small contribution would be a good filter. Looking at the stats it’s quite crazy. Didn’t know that we could access to this data. Thanks for sharing.
I'll add the amount should be enough to cover at least a cursory review. A full review would be better. I just don't want to price out small players.
The papers could also be categorized as unreviewed, quick check, fully reviewed, or fully reproduced. They could pay for this to be done or verified. Then, we have a reputational problem to deal with on the reviewer side.
I don't know about CS, but in mathematics the vast majority of researchers would not have enough funding to pay for a good quality full review of their articles. The peer review system mostly runs on good will.
I'm assuming it cost somewhere between no review and a thorough one. Past that, I assume nothing. Pay reviewers per review or per hour like other consultants. Groups like Arxiv would, for a smaller fee, verify the reviewer's credentials and that the review happened.
That's if anyone wants the publishing to be closer to thr scientific method. Arxiv themselves might not attempt all of that. We can still hope for volunteers to review papers in a field with little, peer review. I just don't think we can call most of that science anymore.
> Before being considered for submission to arXiv’s CS category, review articles and position papers must now be accepted at a journal or a conference and complete successful peer review.
Edit: original title was "arXiv No Longer Accepts Computer Science Position or Review Papers Due to LLMs"
Agree. Additionally, original title, "arXiv No Longer Accepts Computer Science Position or Review Papers Due to LLMs" is ambiguous. “Due to LLMs” is being interpreted as articles written by LLMs, which is not accurate.
No, the post is definitely complaining about articles written by LLMs:
"In the past few years, arXiv has been flooded with papers. Generative AI / large language models have added to this flood by making papers – especially papers not introducing new research results – fast and easy to write."
"Fast forward to present day – submissions to arXiv in general have risen dramatically, and we now receive hundreds of review articles every month. The advent of large language models have made this type of content relatively easy to churn out on demand, and the majority of the review articles we receive are little more than annotated bibliographies, with no substantial discussion of open research issues."
Surely a lot of them are also about LLMs: LLMs are the hot computing topic and where all the money and attention is, and they're also used heavily in the field. So that could at least partially account for why this policy is for CS papers only, but the announcement's rationale is about LLMs as producing the papers, not as their subject.
Isn't arXiv where you upload things before they have gone through the entire process? Isn't that the entire value, aside from some publisher cartel busting?
I don’t know about this. From a pure entertainment standpoint, we may be denying ourselves a world of hilarity. LLMs + “You know Peter, I’m something of a research myself” delusions. I’d pay for this so long as people are very serious about the delusion.
i would like to understand what people get, or think they get, out of putting a completely AI-generated survey paper on arXiv.
Even if AI writes the paper for you, it's still kind of a pain in the ass to go through the submission process, get the LaTeX to compile on their servers, etc., there is a small cost to you. Why do this?
Gaming the h-index has been a thing for a long time in circles where people take note of such things. There are academics who attach their name to every paper that goes through their department (even if they contributed nothing), there are those who employ a mountain of grad students to speed run publishing junk papers... and now with LLMs, one can do it even faster!
Published papers are part of the EB-1 visa rubric so huge value in getting your content into these indexes:
"One specific criterion is the ‘authorship of scholarly articles in professional or major trade publications or other major media’. The quality and reputation of the publication outlet (e.g., impact factor of a journal, editorial review process) are important factors in the evaluation”
I have a hunch that most of the slop is not just on CS but specifically about AI. For some reason, a lot of people's first idea when they encounter an LLM is "let's have this LLM write an opinion piece about LLMs", as if they want to test its self-awareness or hack it by self-recursion. And then they get a medley of the learning data, which if they are lucky contains some technical explanations sprinkled in.
That said, AI-generated papers have already been spotted in other disciplines besides cs, and some of them are really obvious (arXiv:2508.11634v1 starts with a review of a non-existing paper). I really hope arXiv won't react by narrowing its scope to "novel research only"; in fact there is already AI slop in that category and it is harder to spot for a moderator.
("Peer-reviewed papers only" is mostly equivalent to "go away". Authors post on the arXiv in order to get early feedback, not just to have their paper openly accessible. And most journals at least formally discourage authors from posting their papers on the arXiv.)
I'm not sure this is the right way to handle it (I don't know what is) but arXiv.org has suffered from poor quality self-promotion papers in CS for a long time now. Years before llms.
How precisely does it "suffer" though? It's basically a way to disseminate results but carries no journalistic prestige in itself. It's a fun place to look now and then for new results, but just reading the "front page" of a category has always been a Caveat Emptor situation.
Because a large number of "preprints" that are really blog posts or advertisements for startup greatly increase the noise.
The idea is the site is for academic preprints. Academia has a long history of circulating preprints or manuscripts before the work is finished. There are many reasons for this, the primary one is that scientific and mathematical papers are often in the works for years before they get officially published. Preprints allow other academics in the know to be up to date on current results.
If the service is used heavily by non-academics to lend an aura of credibility to any kind of white paper then the service is less usable for its intended purpose.
It's similar to the use of question/answer sites like Quora to write blog posts and ads under questions like "Why is Foobar brand soap the right soap for your family?"
The review paper is dead... so this is a good development. Like you can generate these things in a couple of iterations with AI and minor edits. Preprint servers could be dealing with 1000s of review/position papers over short periods of time and then this wastes precious screening work hours.
It is a bit different in other fields where interpretations or know-how might be communicated in a review paper format that is otherwise not possible. For example, in biology relating to a new phenomena or function.
What are review papers for anyway? I think they are either for
1) new grad students to end up with something nice to publish after reviewing the literature or,
2) older professors to write a big overview of everything that happened in their field as sort of a “bible” that can get you up to speed
The former is useful as a social construct; I mean, hey, new grad students, don’t skimp on your literature review. Finding out a couple years in that folks had already done something sorta similar to my work was absolutely gut-wrenching.
For the latter, I don’t think LLMs are quite ready to replace the personal experiences of a late-career professor, right?
I've found (good) review papers invaluable as an academic. They're really useful as a fast ladder to getting up to speed in a new area. Usually they have a great literature review (with the important papers to read afterward), a curated list of results important to understand, and good intuition about how to reason. It's a compactification of what I would have to otherwise gain by working in an area for years. No replacement for it, of course, but does make it easier attain.
I don't understand the appeal of an (majorly-)LLM generated review paper. A good review paper is a hard task to write well, and frankly the only good ones I've read have come from authors who are at apex of their field (and are, in particular, strong writers). The 'lossy search' of an LLM is probably an outstanding tool for _refining_ a review paper, but for fully generating it? At least not with current LLMs.
Ultimately, a key reason to write these papers in the first place is to guide practitioners in the field, right? Otherwise science itself is just a big (redacted term that can get people shadow-banned for simply using it).
As one of those practitioners, I've found good review/survey papers to be incredibly valuable. They call my attention to the important publications and provide at least a basic timeline that helps me understand how the field has evolved from the beginning and what aspects people are focusing on now.
At the same time, I'll confess that I don't really see why most such papers couldn't be written by LLMs. Ideally by better LLMs than we have now, of course, but that could go without saying.
Are review papers centred on novel research? I get what you mean ofc but most are really mundane overviews. In good review papers the authors offer novel interpretations/directions but even then it involves a lot of grunt work too.
A good review paper is infinitely better than an llm managing to find a few papers and making a summary.
A knowledgeable researcher knows which papers are outdated and can make a trustworthy review paper, an LLM can't easily do that yet
Ok I take your point. However, it is possible to generate a middling review paper combining ai generated slop and edits. Maybe we would be tricked by it in certain circumstances. I don't mean to imply these outputs are something I would value reading. I am just arguing in favour of the proposed approach of arXiv.
Review papers are summarizations to recent updates in the field that deserve fellow researchers' attention. Such works should be done annually or at most quarterly in my opinion, to include only time-tested results.
If hundreds of review papers are published every month, I am afraid that their quality in terms of paper selection and innovative interpretation/direction will not be much higher than the content generated by LLM, even if written word-to-word by a real scientist.
LLMs are good at plainly summarizing from the public knowledge base. Scientists should invest their time in contributing new knowledge to public base instead of doing the summarization.
Great move by arXiv—clear standards for reviews and position papers are crucial in fast-moving areas like multi-agent systems and agentic LLMs. Requiring machine-readable metadata (type=review/position, inclusion criteria, benchmark coverage, code/data links) and consistent cross-listing (cs.AI/cs.MA) would help readers and tools filter claims, especially in distributed/parallel agentic AI where evaluation is fragile. A standardized “Survey”/“Position” tag plus a brief reproducibility checklist would set expectations without stifling early ideas.
Why not just reject papers authored by LLMs and ban accounts that are caught? arXiv’s management has become really questionable lately, it’s like they’re trying to become a prestigious journal and are becoming the problem they were trying to solve in the first place
What matters is the quality. Requiring reviews and opinions to be peer-reviewed seems a lot less superficial than rejecting LLM-assisted papers (which can be valid). This seems like a reasonable filter for papers with no first-party contributions. I'm sure they ran actual numbers as well.
It’s articles (not papers) _about_ LLMs that are the problem, not papers written _by_ LLMs (although I imagine they are not mutually exclusive). Title is ambiguous.
> It’s articles (not papers) _about_ LLMs that are the problem, not papers written _by_ LLMs
No, not really. From the blog post:
> In the past few years, arXiv has been flooded with papers. Generative AI / large language models have added to this flood by making papers – especially papers not introducing new research results – fast and easy to write. While categories across arXiv have all seen a major increase in submissions, it’s particularly pronounced in arXiv’s CS category.
> [...]
> Fast forward to present day – submissions to arXiv in general have risen dramatically, and we now receive hundreds of review articles every month. The advent of large language models have made this type of content relatively easy to churn out on demand, and the majority of the review articles we receive are little more than annotated bibliographies, with no substantial discussion of open research issues.
Verification via LLM tends to break under quite small optimization pressure. For example I did RL to improve <insert aspect> against one of the sota models from one generation ago, and the (quite weak) learner model found out that it could emit a few nonsense words to get the max score.
That's without even being able to backprop through the annotator, and also with me actively trying to avoid reward hacking. If arxiv used an open model for review, it would be trivial for people to insert a few grammatical mistakes which cause them to receive max points.
A better policy might be for arXiv to do the following:
1. Require LLM produced papers to be attributed to the relevant LLM and not the person who wrote the prompt.
2. Treat submissions that misrepresent authorship as plagiarism. Remove the article, but leave an entry for it so that there is a clear indication that the author engaged in an act of plagiarism.
Review papers are valuable. Writing one is a great way to gain, or deepen, mastery over a field. It forces you to branch out and fully assimilate papers that you may have only skimmed, and then place them in their proper context. Reading quality review papers is also valuable. They're a great way for people new to a field to get up to speed and they can bring things that were missed to the fore, even for veterans of the field.
While the current generation of AI does a poor job of judging significance and highlighting what is actually important, they could improve in the future. However, there's no need for arXiv to accept hundreds of review papers written by the same model on the same field, and readers certainly don't want to sift through them all.
Clearly marking AI submissions and removing credit from the prompters would adequately future-proof things for when, and if, AI can produce high quality review papers. Clearly marking authors who engage in plagiarism as plagiarists will, hopefully, remove most of the motivation to spam arXiv with AI slop that is misrepresented as the work of humans.
My only concern would be for the cost to arXiv of dealing with the inevitable lawsuits. The policy arXiv has chosen is worse for science, but is less likely to get them sued by butt-hurt plagiarists or the very occasional false positive.
That doesn't solve the problem they're trying to solve, which is their all-volunteer staff is being flooded with LLM slop and doesn't have the time to artistically moderate.
If you want to blame someone, blame all the people LARPing as AI researchers.
The majority of these submissions are not from anonymous trolls. They're from identifiable individuals who are trying to game metrics. The threat of boosting their number of plagiarism offences on public record would deter such individuals quite effectively.
Meanwhile, banning review articles written by humans would be harmful in many fields. I'm not in CPSC, but I'd hate to see this policy become the norm for all disciplines.
> The advent of large language models have made this type of content relatively easy to churn out on demand, and the majority of the review articles we receive are little more than annotated bibliographies, with no substantial discussion of open research issues.
I have to agree with their justification. Since "Attention Is All You Need" (2017) I have seen maybe four papers with similar impact in the AI/ML space. The signal to noise ratio is really awful. If I had to pick a semi-related paper published since 2020 that I actually found interesting, it would have to be this one: https://arxiv.org/abs/2406.19108 I cannot think of a close second right now.
All of the machine learning papers are pure slop to me now. The last one I looked at had an abstract that was so long it put me to sleep. Many of these papers aren't attempting basic decorum anymore. Mandatory peer review would fix a lot of this. I don't think it is acceptable for the staff at arXiv to have to endure a Sisyphean mountain of LLM shit. They definitely need to push back.
You picked the arguably most impactful AI/ML paper of the century so far, no wonder you don't find others with similar impact.
Not every paper can be a world-changing breakthrough. Which doesn't mean that more modest papers are noise (although some definitely are). What Kuhn calls "normal science" is also needed for science to work.
Isn’t the signal to noise problem what journals are supposed to be for? I thought arxiv was supposed to just be a record keeper, to make it easy to share papers and preprints.
This is only for review/position papers, though I agree that pretty much all ML papers for the past 20 years have been slop. I also consider the big names like, "Adam", "Attention", or "Diffusion" slop, because even thought they are powerful and useful, the presentation is so horrible (for the first two) or they contain major mistakes in the justication of why they work (the last two) that they should never have gotten past review without major rewrites.
This should honestly have been implemented a long time ago. Much of academia is pressured to churn out papers month after month as academia is prioritizing volume over quality or impact.
It doesn't apply CS papers in general - only opinion pieces and surveys of existing papers. i.e. it only bans preprints for papers that contribute nothing new.
I've seen quite a few preprints posted on HN with clearly fantastical claims that only seem to reinforce or ride the coattails of the current hype cycle. It's no longer research, it's becoming "top of funnel thought leadership".
It is actually great because it shows how well it works as a system. Screening is really important to keep preprint quality high enough to then implement cool ideas like random peer review/automated reviews etc
Anecdotally, a lot of researchers will run their paper pdfs through an AI iteration or two during drafting which also (kinda but not really) counts as a self-review. Although that is not comparable to peer review ofc.
I had a convo with a senior CS prof at Stanford two years ago. He was excited about LLM use in paper writing to, e.g., "lower barriers" to idk, "historically marginalized groups" and to "help non-native English speakers produce coherent text". Etc, etc - all the normal tech folk gobbledygook, which tends to forecast great advantage with minimal cost...and then turn out to be wildly wrong.
There are far more ways to produce expensive noise with LLMs than signal. Most non-psychopathic humans tend to want to produce veridical statements. (Except salespeople, who have basically undergone forced sociopathy training.) At the point where a human has learned to produce coherent language, he's also learned lots of important things about the world. At the point where a human has learned academic jargon and mathematical nomenclature, she has likely also learned a substantial amount of math. Few people want to learn the syntax of a language with little underlying understanding. Alas, this is not the case with statistical models of papers!
If you read through the papers, you'll realize the actual problem is blatant abuse and reputation hacking.
So many "research papers" by "AI companies" that are blog posts or marketing dressed up as research. They contribute nothing and exist so the dudes running the company can point to all their "published research".
arXiv was built over a good faith assumption, where a long paper meant at least the author had put some effort behind, and a every idea deserved attention. AI generated text breaks that assumption, and anybody uploading it is not acting in good faith.
And it's a unequal arms race, in which generating endless slop is way cheaper than storing it, because slop generators are subsidised (by operating at a loss) but arXiv has to pay the full price for their hosting.
I wonder why they can't facilitate LLMs in the review process (like fighting fire with fire). Are even the best models not capable enough, or are the costs too high?
the problem is generally the same as with generative adversarial networks; the capability to meaningfully detect some set of hallmarks of LLMs automatically is equivalent to the capability to avoid producing those, and LLMs are trained to predict (ie. be indistinguishable from) their source corpus of human-written text.
so the LLM detection problem is (theoretically) impossible for SOTA LLMs; in practice, it could be easier due to the RLHF stage inserting idiosyncrasies.
Sure, having a 100% reliable system is impossible as you have laid out. However, if I understand the announcement correctly, this is about volume, and I wonder if you could have a tool flag articles that show obvious signs of LLM usage.
Curious for the state on things here. Can we reliably tell if a text was LLM generated? I just heard of a prof screening assignments for this, but not sure how that would work.
Well, I think it depends on how much effort the 'writer' is going to invest. If the writer simply tells the LLM to write something, you can be fairly certain it can be identified. However, I am not sure if the 'writer' provides extensive style instructions (e.g., earlier works by the same author).
Anecdotal: A few weeks ago, I came across a story on HN where many commenters immediately recognized that an LLM had written the article, and the author had actually released his prompts and iterations. So it was not a one-shot prompt but more like 10 iterations, and still, many people saw that an LLM wrote it.
Of course there are people who will sell you a tool to do this. I sincerely doubt it's any good. But then again they can apparently fingerprint human authors fairly well using statistics from their writing, so what do I know.
There are tools that claim accuracies in the 95%-99% range. This is useless for many actual applications, though. For example, in teaching, you really need to not have false positives at all. The alternative is failing some students because a machine unfairly marked their work as machine-generated.
And anyway, those accuracies tend to be measured on 100% human-generated vs. 100% machine-generated texts by a single LLM... good luck with texts that contain a mix of human and LLM contents, mix of contents by several LLMs, or an LLM asked to "mask" the output of another.
I always figured if I wrote a paper, the peer review would be public scrutiny. As in, it would have revolutionary (as opposed to evolutionary) innovations that disrupt the status quo. I don't see how blocking that kind of paper from arXiv helps hacker culture in any way, so I oppose their decision.
They should solve the real problem of obtaining more funding and volunteers so that they can take on the increased volume of submissions. Especially now that AI's here and we can all be 3 times as productive for the same effort.
Before being considered for submission to arXiv’s CS category, review articles and position papers must now be accepted at a journal or a conference and complete successful peer review.
Huh, I guess it's only a subset of papers, not all of them. My brain doesn't work that way, because I don't like assigning custom rules for special cases (edit: because I usually view that as a form of discrimination). So sometimes I have a blind spot around the realities of a problem that someone is facing, that don't have much to do with its idealization.
What I mean is, I don't know that it's up to arXiv to determine what a "review article and position paper" is. Because of that, they must let all papers through, or have all papers face the same review standards.
When I see someone getting their fingers into something, like muddying/dithering concepts, shifting focus to something other than the crux of an argument (or using bad faith arguments, etc), I view it as corruption. It's a means for minority forces to insert their will over the majority. In this case, by potentially blocking meaningful work from reaching the public eye on a technicality.
So I admit that I was wrong to jump to conclusions. But I don't know that I was wrong in principle or spirit.
It’s weird to say that you can be three times more efficient at taking down AI slop now that AI is here, given that the problem is exacerbated by AI in the first place. At least without AI authors were forced to actually write the slop themselves…
This does not seem like a win even if your “fight AI with AI plan works.”
If you incentivize researchers to publish papers, individuals will find ways to game the system, meeting the minimum quality bar, while taking the least effort to create the most papers and thereby receive the greatest reward.
Similarly, if you reward content creators based on views, you will get view maximization behaviors. If you reward ad placement based on impressions, you will see gaming for impressions.
Bad metrics or bad rewards cause bad behavior.
We see this over and over because the reward issuers are designing systems to optimize for their upstream metrics.
Put differently, the online world is optimized for algorithms, not humans.
I heard someone say something similar about the “homeless industrial complex” on a podcast recently. I think it was San Francisco that pays NGOs funds for homeless aid based on how many homeless people they serve. So the incentive is to keep as many homeless around as possible, for as long as possible.
https://www.cnbc.com/2018/04/11/goldman-asks-is-curing-patie...
Ditto for views, etc. Really what you care about as eg; youtube is conversions for the products that are advertised. Not impressions. But there's an attribution problem there.
Blame people, bad actors, systems of incentives, the gods, the devils, but never broach the fault of LLMs and their wide spread abuse.
Blaming LLMs is unproductive. They are not going anywhere (especially since open source LLMs are so good.)
If we want to achieve real change, we need to accept that they exist, understand how that changes the scientific landscape and our options to go from here.
I'old enough to remember when GAN where going to be used to scam millions of people and flood social media with fake profiles.
I think such statements are likely projections of people's own unwillingness to part with such tools given their own personal perceived utility.
I, for one, wouldn't give up LLMs. Too useful to me personally. So, I will always seek them out.
LLMs are not submitting these papers on their own, people are. As far as I'm concerned, whatever blame exists rests on those people and the system that rewards them.
Guns are entirely inert objects, devoid of either free will nor volition, they have no rights and no responsibilities.
LLMs likewise.
https://www.goodreads.com/quotes/421467-to-every-man-is-give...
https://calteches.library.caltech.edu/1575/1/Science.pdf
LLMs are not the root of the problem here.
I've been thinking we're living in "Goodhart's Hell". Where metric hacking has become the intent. That we've decided metrics are all that matter and are perfectly aligned with our goals.
But hey, who am I to critique. I'm just a math nerd. I don't run a multi trillion dollar business that lays off tons of workers because the current ones are so productive due to AI that they created one of the largest outages in history of their platform (and you don't even know which of the two I'm referencing!). Maybe when I run a multi trillion dollar business I'll have the right to an opinion about data.
Sure, publishing on important papers has its weight, but not as much as getting cited.
How would an online world that is optimized for humans, not algorithms, look like?
Should content creators get paid?
I don't think so. Youtube was a better place when it was just amateurs posting random shit.
Everybody "creates content" (like me when I take a picture of beautiful sunset).
There is no such thing as "quality". There is quality for me and quality for you. That is part of the problem, we can't just relate to some external, predefined scale. We (the sum of people) are the approximate, chaotic, inefficient scale.
Be my guest to propose a "perfect system", but - just in case there is no such system - we should make sure each of us "rewards" what we find of quality (being people or content creators), and hope it will prevail. Seemed to have worked so far.
After 1989, most academics complained the system is not merit-based and practical (applied) enough. So we changed it to grants and publications metrics (modeled after the West). For a while, it worked.. until people found too much overbearing bureaucracy and some learned how to game the system again.
I would say, both systems have failure modes of a similar magnitude, although the first one is probably less hoops and less stress on each individual. (During communism, academia - if you could get there, especially technical sciences - was an oasis of freedom.)
arXiv believes that there are position papers and review articles that are of value to the scientific community, and we would like to be able to share them on arXiv. However, our team of volunteer moderators do not have the time or bandwidth to review the hundreds of these articles we receive without taking time away from our core purpose, which is to share research articles.
From TFA. The problem exists. Now.
Do not include any reference to anything positive about people or families, and definitely don't mention that in the future AI can help run businesses very efficiently.[1] "
[0] https://medium.com/@rviragh/life-as-a-victim-of-someone-else...
[1]
> Technically, no! If you take a look at arXiv’s policies for specific content types you’ll notice that review articles and position papers are not (and have never been) listed as part of the accepted content types.
You cannot upload the journal’s version, but you can upload the text as accepted (so, the same content minus the formatting).
Why not? I don't know about in CS, but, in math, it's increasingly common for authors to have the option to retain the copyright to their work.
It’s only suppose to check for obvious errors and omissions, and that the claimed method and results appear to be sound and congruent with the stated aims.
I think every project more or less deviates from its original goal given enough time. There are few exceptions in CS like GNU coreutils. cd, ls, pwd, ... they do one thing and do it well very likely for another 50 years.
Not as gate-keepy as journals and not as anarchic as purely open publishing. Should be cheap, too.
Fundamentally, we want research that offers something new (“what did we learn?”) and presents it in a way that at least plausibly has a chance of becoming generalizable knowledge. You call it gate-keeping, but I call it keeping published science high-quality.
It's related to the same problems you have with e.g. Sybil attacks: https://en.wikipedia.org/wiki/Sybil_attack
I'm not saying it wouldn't be worthwhile to try, just that I expect there to be a lot of very difficult problems to solve there.
That is to say I also think it would be worthwhile to try.
Also look how frequently they publish. Do you really think it's reasonable to produce a paper every week or two? Even if you have a team of grad students? I'll put it this way, I had a paper have difficulty getting through reviewer for "not enough experiments" when several of my experiments took weeks wall time to run and one took a month (could not run that a second time lol)
We don't do a great job at ousting frauds in science. It's actually difficult to do because science requires a lot of trust. We could alleviate some of these issues if we'd allow publication or some reward mechanism for replication, but the whole system is structured to reward "new" ideas. Utility isn't even that much of a factor in some areas. It's incredibly messy.
Most researchers are good actors. We all make mistakes and that's why it's hard to detect fraud. But there's also usually high reward for doing so. Though most of that reward is actually getting a stable job and the funding to do your research. Which is why you can see how it might be easy to slip into cheating a little here and there. There's ways to solve that that don't include punishing anyone...
https://openreview.net/forum?id=cIKQp84vqN
Wouldn’t most people subscribe to a default set of trusted citers?
Sure. This happens with ad blockers, for example. I imagine Elsevier or Wikipedia would wind up creating these lists. And then you’d have the same incentives as you have now for fooling that authority.
> or people just don't care very much
This is my hypothesis. If you’re an expert, you have your web of trust. If you’re not, it isn’t that hard to start from a source of repute.
And to bring this back to the original arxiv topic. I think reputation system is going to face problems with some people outside CS lack of enough technical abilities. It also introduce biases in that you would endorse people who you like for other reasons. Actually some of the problems are solved and you would need careful proposal. But the change for publishing scheme needs push from institutions and funding agencies. Authors don't oppose changes but you have a lobby of the parasitic publishing cartel that will oppose these changes.
I don't think publishing a PGP key with your work does anything. There's no problem identifying the author of the work. The problem is identifying _untrustworthy_ authors. Especially in the face of many other participants in the system claiming the work is trusted.
As I understand it, the current system (in some fields) is essentially to set up a bunch of sockpuppet accounts to cite the main account and publish (useless) derivative works using the ideas from the main account. Someone attempting to use existing reasearch for it's intended purpose has no idea that the whole method is garbage / flawed / not reproducible.
If you can only trust what you, yourself verify, then the publications aren't nearly as useful and it is hard to "stand on the shoulders of giants" to make progress.
Is it though? Should we care about authors or about the work? Yes, many experiments are hard to reproduce, but isn't that something we should work towards, rather than just "trust" someone. People change. People do mistakes. I think more open data, open access, open tools, will solve a lot, but my guess is that generally people do not like that because it can show their weaknesses - even if they are well intentioned.
Edit: For clarification I’m agreeing with OP
Loosely speaking, the "received wisdom" has generally been that if you have a .edu address, you can probably publish fairly freely. But my understanding is that the rules are a little more nuanced than that. And I think there are other, non .edu domains, where you will also get auto-endorsed. But they don't publish a list of such things for obvious reasons.
[0]: Unless things have changed since I created my account, which was originally created with my personal email address. That was quite some time ago, so I guess it's possible changes have happened that I'm not aware of.
[1]: https://info.arxiv.org/help/endorsement.html
Which includes some very large ones like @google.com
Her suggestion was simple: Kick out all non-ivy league and most international researchers. Then you have a working reputation system.
Make of that what you will ...
[1] https://en.wikipedia.org/wiki/Grigori_Perelman [2] https://www.ams.org/notices/200808/tx080800930p.pdf
Treat everyone equally. After 10 years of only quality you get chance to get back. Before that though luck.
(1) because ivy league also produces a lot of work that's not so great (i.e. wrong (looking at you, Ariely) or un-ambitious) and
(2) because from time to time, some really important work comes out of surprising places.
I don't think we have a good verdict on the Orthega hypothesis yet, but I'm not a professional meta scientist.
That said, your proposal seems like a really good idea, I like it! Except I'd apply it to individuals and/or labs.
Asking for a small amount of money would probably help. Issue with requiring peer reviewed journals or conferences is the severe lag, takes a long time and part of the advantage of arxiv was that you could have the paper instantly as a preprint. Also these conferences and journals are also receiving enormous quantities of submissions (29.000 for AAAI) so we are just pushing the problem.
The papers could also be categorized as unreviewed, quick check, fully reviewed, or fully reproduced. They could pay for this to be done or verified. Then, we have a reputational problem to deal with on the reviewer side.
You might be vastly underestimating the cost of such a feature
That's if anyone wants the publishing to be closer to thr scientific method. Arxiv themselves might not attempt all of that. We can still hope for volunteers to review papers in a field with little, peer review. I just don't think we can call most of that science anymore.
> Before being considered for submission to arXiv’s CS category, review articles and position papers must now be accepted at a journal or a conference and complete successful peer review.
Edit: original title was "arXiv No Longer Accepts Computer Science Position or Review Papers Due to LLMs"
ArXiv CS requires peer review for surveys amid flood of AI-written ones
- nothing happened to preprints
- "summarization" articles always required it, they are just pointing at it out loud
"In the past few years, arXiv has been flooded with papers. Generative AI / large language models have added to this flood by making papers – especially papers not introducing new research results – fast and easy to write."
"Fast forward to present day – submissions to arXiv in general have risen dramatically, and we now receive hundreds of review articles every month. The advent of large language models have made this type of content relatively easy to churn out on demand, and the majority of the review articles we receive are little more than annotated bibliographies, with no substantial discussion of open research issues."
Surely a lot of them are also about LLMs: LLMs are the hot computing topic and where all the money and attention is, and they're also used heavily in the field. So that could at least partially account for why this policy is for CS papers only, but the announcement's rationale is about LLMs as producing the papers, not as their subject.
PaperMatch [1] helps solve this problem (large influx of papers) by running a semantic search on top of abstracts, for all of arXiv.
[1]: https://papermatch.me/
Even if AI writes the paper for you, it's still kind of a pain in the ass to go through the submission process, get the LaTeX to compile on their servers, etc., there is a small cost to you. Why do this?
"One specific criterion is the ‘authorship of scholarly articles in professional or major trade publications or other major media’. The quality and reputation of the publication outlet (e.g., impact factor of a journal, editorial review process) are important factors in the evaluation”
I've never seen arXiv papers counted towards your publications anywhere that the number of your publications are used as a metric. Is USCIS different?
That said, AI-generated papers have already been spotted in other disciplines besides cs, and some of them are really obvious (arXiv:2508.11634v1 starts with a review of a non-existing paper). I really hope arXiv won't react by narrowing its scope to "novel research only"; in fact there is already AI slop in that category and it is harder to spot for a moderator.
("Peer-reviewed papers only" is mostly equivalent to "go away". Authors post on the arXiv in order to get early feedback, not just to have their paper openly accessible. And most journals at least formally discourage authors from posting their papers on the arXiv.)
Beyond hosting cost, there is some prestige to seeing an arXiv link versus rando blog post despite both having about the same hurdle to publishing.
The idea is the site is for academic preprints. Academia has a long history of circulating preprints or manuscripts before the work is finished. There are many reasons for this, the primary one is that scientific and mathematical papers are often in the works for years before they get officially published. Preprints allow other academics in the know to be up to date on current results.
If the service is used heavily by non-academics to lend an aura of credibility to any kind of white paper then the service is less usable for its intended purpose.
It's similar to the use of question/answer sites like Quora to write blog posts and ads under questions like "Why is Foobar brand soap the right soap for your family?"
It is a bit different in other fields where interpretations or know-how might be communicated in a review paper format that is otherwise not possible. For example, in biology relating to a new phenomena or function.
1) new grad students to end up with something nice to publish after reviewing the literature or,
2) older professors to write a big overview of everything that happened in their field as sort of a “bible” that can get you up to speed
The former is useful as a social construct; I mean, hey, new grad students, don’t skimp on your literature review. Finding out a couple years in that folks had already done something sorta similar to my work was absolutely gut-wrenching.
For the latter, I don’t think LLMs are quite ready to replace the personal experiences of a late-career professor, right?
I don't understand the appeal of an (majorly-)LLM generated review paper. A good review paper is a hard task to write well, and frankly the only good ones I've read have come from authors who are at apex of their field (and are, in particular, strong writers). The 'lossy search' of an LLM is probably an outstanding tool for _refining_ a review paper, but for fully generating it? At least not with current LLMs.
As one of those practitioners, I've found good review/survey papers to be incredibly valuable. They call my attention to the important publications and provide at least a basic timeline that helps me understand how the field has evolved from the beginning and what aspects people are focusing on now.
At the same time, I'll confess that I don't really see why most such papers couldn't be written by LLMs. Ideally by better LLMs than we have now, of course, but that could go without saying.
The problem is you can’t. Not without careful review of the output. (Certainly not if you’re writing about anything remotely novel and thus useful.)
But not everyone knows that, which turns private ignorance into a public review problem.
If you’re an expert. If you’re not, you’ll publish, best case, bullshit. (Worst case lies.)
LLMs are good at plainly summarizing from the public knowledge base. Scientists should invest their time in contributing new knowledge to public base instead of doing the summarization.
These things will ruin everything good, and that is before we even start talking about audio or video.
It is also turning people into spammers because it makes bluffers feel like experts.
ChatGPT is so revealing about a person's character.
Are you saying that there's an automated method for reliably verifying that something was created by an LLM?
No, not really. From the blog post:
> In the past few years, arXiv has been flooded with papers. Generative AI / large language models have added to this flood by making papers – especially papers not introducing new research results – fast and easy to write. While categories across arXiv have all seen a major increase in submissions, it’s particularly pronounced in arXiv’s CS category. > [...] > Fast forward to present day – submissions to arXiv in general have risen dramatically, and we now receive hundreds of review articles every month. The advent of large language models have made this type of content relatively easy to churn out on demand, and the majority of the review articles we receive are little more than annotated bibliographies, with no substantial discussion of open research issues.
If so, I think the solution is obvious.
(But I remind myself that all complex problems have a simple solution that is wrong.)
That's without even being able to backprop through the annotator, and also with me actively trying to avoid reward hacking. If arxiv used an open model for review, it would be trivial for people to insert a few grammatical mistakes which cause them to receive max points.
Doubt
LLMs are experts in generating junk. And generally terrible at anything novel. Classifying novel vs junk is a much harder problem.
1. Require LLM produced papers to be attributed to the relevant LLM and not the person who wrote the prompt.
2. Treat submissions that misrepresent authorship as plagiarism. Remove the article, but leave an entry for it so that there is a clear indication that the author engaged in an act of plagiarism.
Review papers are valuable. Writing one is a great way to gain, or deepen, mastery over a field. It forces you to branch out and fully assimilate papers that you may have only skimmed, and then place them in their proper context. Reading quality review papers is also valuable. They're a great way for people new to a field to get up to speed and they can bring things that were missed to the fore, even for veterans of the field.
While the current generation of AI does a poor job of judging significance and highlighting what is actually important, they could improve in the future. However, there's no need for arXiv to accept hundreds of review papers written by the same model on the same field, and readers certainly don't want to sift through them all.
Clearly marking AI submissions and removing credit from the prompters would adequately future-proof things for when, and if, AI can produce high quality review papers. Clearly marking authors who engage in plagiarism as plagiarists will, hopefully, remove most of the motivation to spam arXiv with AI slop that is misrepresented as the work of humans.
My only concern would be for the cost to arXiv of dealing with the inevitable lawsuits. The policy arXiv has chosen is worse for science, but is less likely to get them sued by butt-hurt plagiarists or the very occasional false positive.
If you want to blame someone, blame all the people LARPing as AI researchers.
Meanwhile, banning review articles written by humans would be harmful in many fields. I'm not in CPSC, but I'd hate to see this policy become the norm for all disciplines.
I have to agree with their justification. Since "Attention Is All You Need" (2017) I have seen maybe four papers with similar impact in the AI/ML space. The signal to noise ratio is really awful. If I had to pick a semi-related paper published since 2020 that I actually found interesting, it would have to be this one: https://arxiv.org/abs/2406.19108 I cannot think of a close second right now.
All of the machine learning papers are pure slop to me now. The last one I looked at had an abstract that was so long it put me to sleep. Many of these papers aren't attempting basic decorum anymore. Mandatory peer review would fix a lot of this. I don't think it is acceptable for the staff at arXiv to have to endure a Sisyphean mountain of LLM shit. They definitely need to push back.
Not every paper can be a world-changing breakthrough. Which doesn't mean that more modest papers are noise (although some definitely are). What Kuhn calls "normal science" is also needed for science to work.
Lets say 50000€ fine, or 1 year in prison. :)
What’s the new method?
For example: https://prereview.org/en-us
Anecdotally, a lot of researchers will run their paper pdfs through an AI iteration or two during drafting which also (kinda but not really) counts as a self-review. Although that is not comparable to peer review ofc.
There are far more ways to produce expensive noise with LLMs than signal. Most non-psychopathic humans tend to want to produce veridical statements. (Except salespeople, who have basically undergone forced sociopathy training.) At the point where a human has learned to produce coherent language, he's also learned lots of important things about the world. At the point where a human has learned academic jargon and mathematical nomenclature, she has likely also learned a substantial amount of math. Few people want to learn the syntax of a language with little underlying understanding. Alas, this is not the case with statistical models of papers!
How will journals or conferences handle AI slop?
Don’t understand why it restricted one category when the problem spans multiple categories.
So many "research papers" by "AI companies" that are blog posts or marketing dressed up as research. They contribute nothing and exist so the dudes running the company can point to all their "published research".
Sorry folks but we lost.
And it's a unequal arms race, in which generating endless slop is way cheaper than storing it, because slop generators are subsidised (by operating at a loss) but arXiv has to pay the full price for their hosting.
https://pubmed.ncbi.nlm.nih.gov/18955255/
https://pubmed.ncbi.nlm.nih.gov/16136218/
so the LLM detection problem is (theoretically) impossible for SOTA LLMs; in practice, it could be easier due to the RLHF stage inserting idiosyncrasies.
Anecdotal: A few weeks ago, I came across a story on HN where many commenters immediately recognized that an LLM had written the article, and the author had actually released his prompts and iterations. So it was not a one-shot prompt but more like 10 iterations, and still, many people saw that an LLM wrote it.
And anyway, those accuracies tend to be measured on 100% human-generated vs. 100% machine-generated texts by a single LLM... good luck with texts that contain a mix of human and LLM contents, mix of contents by several LLMs, or an LLM asked to "mask" the output of another.
I think detection is a lost cause.
They should solve the real problem of obtaining more funding and volunteers so that they can take on the increased volume of submissions. Especially now that AI's here and we can all be 3 times as productive for the same effort.
Huh, I guess it's only a subset of papers, not all of them. My brain doesn't work that way, because I don't like assigning custom rules for special cases (edit: because I usually view that as a form of discrimination). So sometimes I have a blind spot around the realities of a problem that someone is facing, that don't have much to do with its idealization.
What I mean is, I don't know that it's up to arXiv to determine what a "review article and position paper" is. Because of that, they must let all papers through, or have all papers face the same review standards.
When I see someone getting their fingers into something, like muddying/dithering concepts, shifting focus to something other than the crux of an argument (or using bad faith arguments, etc), I view it as corruption. It's a means for minority forces to insert their will over the majority. In this case, by potentially blocking meaningful work from reaching the public eye on a technicality.
So I admit that I was wrong to jump to conclusions. But I don't know that I was wrong in principle or spirit.
Those are terms of art, not arbitrary categories. They didn't make them up.
This does not seem like a win even if your “fight AI with AI plan works.”