🚫 Bye-Bye Junk Reviews? arXiv might react on LLM spam
-
It’s 2025, and arXiv, the beloved preprint server of the CS community, is facing a new kind of flood: a torrent of AI-generated review papers. The deluge has become so intense that arXiv moderators are seriously contemplating a controversial policy shift: only accepting review articles that have gone through peer review.
The AI Citation Scheme Invasion
Here’s what’s happening: With the rise of large language models (LLMs), it’s never been easier to churn out a polished-looking review paper. But along with this convenience has come a plague of citation farming — papers that exist primarily to game citation counts, either by over-citing the authors’ own work or linking to random irrelevant papers in what smells suspiciously like “citation-for-hire” schemes.
The result? arXiv moderators are overwhelmed. As they put it:
“We are stuck between a large amount of junk and a small amount of high-quality reviews.”
And because it’s so difficult to distinguish between genuinely helpful review papers and LLM-spam masquerading as scholarship, the editors are thinking of swinging the hammer: ban all review papers unless they’ve been peer-reviewed.
Collateral Damage
That sounds reasonable, right? Until you realize that high-quality, timely reviews, often collaborative efforts across institutions, are now getting caught in the crossfire.
One case that set off this firestorm was a review co-authored by researchers from 20+ top universities. It was “on hold” and potentially rejected, simply because it fell into the new gray area. The authors even contacted arXiv moderators for clarification, only to be told: there’s just too much junk to sort through.
The researchers’ frustration spilled over into Xiaohongshu (China’s Instagram-meets-Reddit), where the author “寒月灼华” summarized the mess:
“There are just too many LLM-generated, citation-spamming review papers. It's hard for moderators to judge what’s good or bad, and good ones are getting mistakenly blocked.”
The Bigger Questions
- Is arXiv still an open-access archive, or has it become a quality filter?
- Should LLMs be banned from generating reviews, or do we just need better detection tools?
- And more philosophically… is the concept of a review paper still relevant in the LLM era?
Here’s what the community had to say:
- "LLM should be used to detect other LLMs."
- "Review papers are just like directories now. Useful, but only if curated carefully."
- "The real solution? Know someone on the moderation team."
- "Deep Research reports are better than most reviews I read."
Ouch!
🧠 Are Review/Survey Papers Still Valuable?
There’s a divide in opinion. Some say review papers are obsolete in the age of ChatGPT and Deep Research. Others argue they’re more important than ever as entry points for cross-disciplinary learning.
“For someone doing LLM work, it's essential to quickly grasp what's happening in NLP, CV, RL, and Sys. A good review serves as a map.”
So, What Now?
The situation remains fluid. If arXiv enforces this new policy, we might see:
- A shift toward submitting review papers to journals like TMLR, Survey & Tutorials, or even Medium (kidding... kinda).
- More reviews being peer-reviewed by design, possibly lowering the bar for journal acceptance.
- New tools for LLM-detection being integrated directly into submission platforms.
- And a lot more whispered “do you know an editor?” backchanneling.
In short, we’re in the middle of an arms race: junk vs. quality, automation vs. moderation, speed vs. scrutiny.
Until then, if you’re planning to post a review on arXiv, maybe wait until you’ve got that Reviewer #2 approval in hand. Otherwise, your magnum opus might just get mistaken for an LLM’s 15-minute side project.
🧵 Have thoughts? Should arXiv be stricter with reviews? Or are they throwing out the baby with the bathwater? Let us know on Twitter, Mastodon, or your nearest hallway track.