It is field dependent but I'm not entirely against what the parent said. I work in ML and I am positive that all this is going on[0]. There's lots of true believers though and that's what makes things extra hard. Sometimes the fraudsters take over by making the system become incompetent and everyone is in good company. In this was fraud isn't committed with intent, weirdly enough.
Just look at all the ML reasoning papers. Wither you believe LLMs reason or not, an important factor you have to disentangle when trying to prove this is what data the models were trained on. To distinguish memorization from reasoning. You won't find this analysis because it's almost impossible given that the data is a trade secret, even by Meta.
This year at ACL a paper (mission impossible language models) won best poste paper despite their results running contrary to their claim, and very obviously so too.
Or there is the HumanEval paper which proposed that they created a data set which was not spoiled because they "hand wrote" over a hundred "Leetcode style problems". 60 authors and they didn't bother to check... But why would you check when the questions are things like "calculate the mean". What fucking programmer thinks there isn't python code on GitHub pre 2021 that: calculates the mean, takes the floor, checks if a string is a palindrome, calculates greatest common devisor, or any similar question. How did this become an influential dataset‽
[0] the big reason I'm upset is because I love the field. I'm not in it for money. I'm in it because I grew up on Asimov books and because I want our community to work towards AGI. But now every person that can do print("hello world") feels that they can lecture me, a published researcher about what these machines do while they talk about the Turing test (lol, what is this, the 60's?) and how they're black boxes (opaque, but certainly not black). I'm fine with armchair experts, but not when they come in swinging with a baseball bat
Just look at all the ML reasoning papers. Wither you believe LLMs reason or not, an important factor you have to disentangle when trying to prove this is what data the models were trained on. To distinguish memorization from reasoning. You won't find this analysis because it's almost impossible given that the data is a trade secret, even by Meta.
This year at ACL a paper (mission impossible language models) won best poste paper despite their results running contrary to their claim, and very obviously so too.
Or there is the HumanEval paper which proposed that they created a data set which was not spoiled because they "hand wrote" over a hundred "Leetcode style problems". 60 authors and they didn't bother to check... But why would you check when the questions are things like "calculate the mean". What fucking programmer thinks there isn't python code on GitHub pre 2021 that: calculates the mean, takes the floor, checks if a string is a palindrome, calculates greatest common devisor, or any similar question. How did this become an influential dataset‽
[0] the big reason I'm upset is because I love the field. I'm not in it for money. I'm in it because I grew up on Asimov books and because I want our community to work towards AGI. But now every person that can do print("hello world") feels that they can lecture me, a published researcher about what these machines do while they talk about the Turing test (lol, what is this, the 60's?) and how they're black boxes (opaque, but certainly not black). I'm fine with armchair experts, but not when they come in swinging with a baseball bat