I wonder if this approach can be used to determine whether a text was generated by a specific LLM.
It should be obvious that LLMs would be able to beat this with ease. Not sure why this paper deliberately skipped comparing to current LLMs
Example of LLMs doing well in similar tasks: https://arxiv.org/abs/2602.16800
I might be misinterpreting but the LUAR model (which is a transformer) seems to do decently well
https://www.nature.com/articles/s41599-025-06340-3/figures/2
Yes, the paper itself tells a different story than the bullet points in this article.
The article seems quite editorialized, shifting between describing "large-scale AI models" and "neural network-based approaches".
The underlying paper itself is more precise, comparing against LUAR, a 2021 method based on bert-style embeddings (i.e. a model with 82M parameters, which is 0.2% the size of e.g. the recent OS Gemma models). I don't fault the authors of the paper at all for this, their method is interesting and more interpretable! But you can check the publication history, their paper was uploaded originally in 2024: https://arxiv.org/abs/2403.08462
A good example of why some folks are bearish on journals.
"AI bad" seems to sell in some circles, and while there are many level-headed criticisms to be made of current AI fads, I don't think this qualifies.
Are you prepared to demonstrate a superior result with models newer than those available when the research was done? Can you suggest a candidate experiment design to test your hypothesis?
Ha! To think that we're finally back to asking ourselves why we are using generative models for categorization and extraction. I wonder how much money has collectively been wasted by companies wittling away at square pegs.
> why we are using generative models for categorization and extraction
Because LLM models have already amortized the man-years cost of collecting, curating and training on text corpuses?
Yeah, LLMs are a solution to the cold start problem plus they are easy to integrate and if you know what you're doing in terms of evals, post processing and so on you can get excellent performance out of them, plus they can do semantic classification and reasoning that you won't get out of some bespoke traditional DS/ML model.
If there's one problem that LLMs have solved, it's language. While an LLM may hallucinate, it does so in grammatically correct English sentences. Additionally, even the local version of gemma-4-26B can seamlessly switch between languages in the midst of a conversation while maintaining context. That's perhaps the most exciting part for me: We have a bonafide universal translator (that's Star Trek territory) and people seem more focused on its factual accuracy.
Tbh. The accuracy of translation is, while much better than prior methods, not that great yet. For tamil atleast.
Kind of a tangent I guess, but the coolest thing about Star Trek’s universal translator to me was that it could translate new languages mid-conversation with an extremely small amount of data. Makes me wonder how close we might be able to get to that eventually
https://en.wikipedia.org/wiki/Darmok
TL;DR, probably never.
Using LLMs for everything is going to be seen as a big fad in a few years. First we try them for everything, then we find what use cases actually make sense, then we scale back. Woe betide our 401(k)s when it happens, though.
The paper did not compare against LLMs though.
This is a concise statement of what I've tried to articulate by analogizing it to railroad infra buildout.
What applications do you think make the most sense so far?
Search, code review, some form of translation...
> Woe betide our 401(k)s when it happens, though.
The stock market crashes once in a while. Shit happens. The long-term outlook is unlikely to change nearly as much, unless you think there will be systemic macroeconomic changes.
Long-term relative to lifespan of the 401K holder. Outcome changes a lot for those who are ready to retire.