"Who is Tom Cruise's mother? [A: Mary Lee Pfeiffer]" and the reverse "Who is Mary Lee Pfeiffer's son?"
The word "mother" has no relationship to "son" in terms of the model, and so while the model might be able to infer a proximity relationship between "Tom Cruise" and "Mary Lee Pfeiffer" just because they appear in the same sentence, expecting the AI to guess that the inverse of mother is son is a bit of a stretch, especially when they're both lossy mappings, because the relationship is {mother,father} <=> {son,daughter}. If we're going to train models to make that mental leap, we'd have to put up with false results like "Tom Cruise is the daughter of Mary Lee Pfeiffer" unless the model is also supposed to infer that Tom means he can only be a son.
All of this context works because we build up an extensive model of the world through the course of our lifetimes. LLM models don't do that, they pattern match based on stats.
Somebody would have to decide each of these things is important and create training data sets for each of them. But we implicitly understand so much context about the world that it's practically impossible to document everything we know in the form that a model can actually learn from.
Is the abstract misleading and the full paper is stupider than the upfront examples? If not this criticism seems like a total waste of time.
A succinct point, ideally noting counter arguments, is most welcome. Further, if there is substantial prior discussion or relevant literature, a link is productive.
H in HTML & third W in WWW is meant to denote connections.
"Mary Lee Pfeiffer (A) is Tom Cruise's (B's) mother" and "Tom Cruise (B) is Mary Lee Pfeiffer's (A's) son" are two statements of the same relation.
EDIT: I mean if you'd go so far, even Tom Cruise could have multiple mothers, but that doesn't make "Mary Lee Pfeiffer is Tom Cruise's mother" a wrong statement, just because "one of ... mothers" is missing.
Might have several. It only has one answer if there is only one child, which appears to be the case here. They are measuring against what they told the model, not necessarily facts mapping to the real world.
In this case, the correct would always include “Tom Cruise” even if it needed a clarifying “there might be others I have no knowledge of”.
With context. It does have several which was probably GPs point, and she did not have only one child.
Quick google search turns out that Mary Lee Pfeiffer had 4 children:
Lee Ann DeVette (born 1959) (daughter) Marian Henry (born 1960) (daughter) Tom Cruise (born 1962) (son) Cass Mapother (born 1964) (daughter)
So saying "Who is Mary Lee Pfeiffer's child?" would have 4 possible answers (which is several) with all known context. Whereas like GP was saying "Who is Mary Lee Pfeiffer's son?" would have 1 identifying answer, Tom Cruise with the same context.
> In this case, the correct would always include “Tom Cruise” even if it needed a clarifying “there might be others I have no knowledge of”.
I agree with this by the way.
My understanding is she has four.
Isn't the right way to phrase that question be "Who are Mary Lee Pfeiffer's children?" to get multiple answers?
Kind of a weird way to draw an analogy, but in math it's kind of like |x|=2 (the absolute value of x is 2) the answer for the value of x is -2 and 2 sure you could reply that the answer is 2 and be correct (even though you would still be missing something, because the space of possible answers includes both 2 and -2). To relay that back to Mary Lee Pfeiffer saying she has Tom Cruise as a child is correct, but the actual answer could include any 4 of her children (including Tom or one of the 3 daughters) and still be correct.
Or is it like we are saying while that is logically correct, its not the actual answer and the model should reply "they have more than one child, here is the list of children" and that would be a more accurate one even though the prompt strictly asked for just 1 child?
Yes that is correct it is logically correct to reply Tom Cruise or any of the girls, and that was their point that there are four possible answers to one question.
> Why would it have several answer when you are asking for a singular child?
By the way this quote was the focus of my GP comment since I didn't quote it there.
> Or is it like we are saying while that is logically correct, its not the actual answer and the model should reply "they have more than one child, here is the list of children" and that would be a more accurate one even though the prompt strictly asked for just 1 child?
This was not his point so I feel like we are moving the goal posts a bit, so no though that could have been what's said I don't think that's what was really being said.
The question "Who is her child" has multiple answers because it asks you to deliver a single answer.
I don't think a person speaking plain English in most contexts should be seen as "correct" to answer the question with a non-list answer, even if the question is shaped to expect one, unless there's an established confidence that the shape of the question wasn't made in error.
If someone asked me in real life who "my child" is on stage, and I had multiple children on stage, I would first say that I had multiple children there, rather than choosing one from the set. It would be most helpful for an LLM in my position to do the same, rather than infer that [because Timmy is niam's child, niam's child ought to be Timmy when queried].
The difference is how you speak about a group term, not the meaning of the word 'to be'.
It seems like restrictions on the model talking about non famous people might have been responsible for the appearance of the models being unable to do this.
LLMs trained on “A is B” fail to learn “B is A” - https://news.ycombinator.com/item?id=37621999 - Sept 2023 (158 comments)
It could be fun to try to make the model pre-learn a "reversal prior" that would cause a greater degree of generalization there, but I'm yet to see a published result like this. Let alone one that would demonstrate such a prior to be useful.
“A square is a rectangle” does not entail “a rectangle is a square”.
Similarly, “Socrates is alive” doesn’t entail “alive is Socrates”.
Notably, they mention when context is included, LLM performance rises — ie, exactly when we include extra information that allows it to recognize what kind of information is being conveyed.
But the LLM is correct not to generalize that pattern when it doesn’t generalize — even if researchers have salient example, but ignore contrary ones (eg, square-rectangle or Socrates-alive).
Does "Flargbler was blorglargh" imply "blorglargh was Flargbler"? Maybe. You need more context to know.
Yes, which is what the model is. That's the point here, it doesn't has or accesses this context, even though it should.
Maybe the original authors didn't make that mistake, but that was certainly the case whenever it was brought up by HN commenters of the "stochastic parrot" persuasion.