The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A" (2023)

by Anon84 • 30 days ago

25 46 30 days ago

eurekin• 30 days ago

I remember seeing this popping up in discussions the first time, but never noticed any resolution (other than to train both sides). Has SOTA advanced?

ACCount37• 30 days ago

No, it's just that no one really cares. It doesn't seem to cause identifiable faults in model reasoning in the real world.

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.

turzmo• 30 days ago

(2023)

gipp• 30 days ago

As I and several other people pointed out last time this was posted, "A is B," in natural language, does not imply "B is A." "Is" can denote any of many different shades of relationship weaker than logical identity.

zmgsabst2• 30 days ago

Even in strict logic, “is” can denote membership, as in, “squares are rectangles” does not entail “rectangles are squares”.

17186274403• 29 days ago

Because it is "(all)squares are (some)rectangles", so the reverse "(some)rectangles are (all)squares" is also true.

The difference is how you speak about a group term, not the meaning of the word 'to be'.

beardyw3• 30 days ago

Yes, "Who is Mary Lee Pfeiffer's son?" happens to have just one answer, whereas "Who is Mary Lee Pfeiffer's child?" would have several.

altmanaltman4• 30 days ago

Why would it have several answer when you are asking for a singular child?

Isn't the right way to phrase that question be "Who are Mary Lee Pfeiffer's children?" to get multiple answers?

whilenot-dev2• 30 days ago

It could have several correct answers, yes, where one should be logically deducible.

"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.

latexr• 30 days ago

"Who is Mary Lee Pfeiffer's child?" would have several.

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”.

sigmoid10• 30 days ago

I wouldn't resort to language examples, as the other comments show how this gets imprecise and lost in semantic details quickly. Instead think of a basic logic example: Consider an OR gate with inputs A and B and output X. If B=1 that means X=1. But if X=1 you can't infer that B=1, because there is an alternative (i.e. A=1,B=0). So from B=1→X=1, the inverse simply does not follow. This extends to all statements where a relation is not symmetric. Of course you can also go beyond and find cases where the relationship is not transitive or not even reflexive. There's a whole branch of language based IQ test puzzles (e.g. "all X are Y, some Y are Z" kind of stuff) that exploit this rabbit hole. Any LLM that does good on these will not jump to conclusions about reverse equalities quickly.

Dylan168072• 30 days ago

That's a fine reaction to the title alone but a very bad reaction to the abstract, where the failure they're criticizing is real and not a childish misunderstanding of the word "is".

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.

clcaev• 30 days ago

As a stylistic comment, on HN I am seeing more and more of "As I have always said..." or similar opening constructs. Concisely documenting a phenomena or affecting change requires detailed, sustained effort. Merely observing a pattern isn't sufficient to be notable.

A succinct point, ideally noting counter arguments, is most welcome. Further, if there is substantial prior discussion or relevant literature, a link is productive.

bayindirh• 30 days ago

As a better version of this, I like to link my prior comments, or if I'm referring to another comment in the same thread, I like to link that one.

H in HTML & third W in WWW is meant to denote connections.

zmgsabst2• 30 days ago

“A is B” doesn’t generally entail “B is A”.

“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).

WithinReason• 30 days ago

That completely misses the point. The point is that "Valentina Tereshkova was the first woman to travel to space" does imply "The first woman to travel to space was Valentina Tereshkova", which LLMs fail to recognise.

IshKebab2• 30 days ago

Right, but that could be because the fact that that implication exists is not actually as trivial as they are implying.

Does "Flargbler was blorglargh" imply "blorglargh was Flargbler"? Maybe. You need more context to know.

Dibby053• 30 days ago

"A is the B" does entail "the B is A", because "the" establishes an identity/bijection.

ralferoo3• 30 days ago

Not only is the inverse not generally true (as others have pointed out), their examples requires several mental leaps.

"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.

trumpdong3• 30 days ago

Pretraining could be reasonably expected to make it learn that mother/father and son/daughter are inverse relationships and Tom is usually a male name.

ralferoo3• 30 days ago

I'd argue that that's not an easy task in and of itself, but even if someone adds a special exception, there's still the issue that there are many other types of inverse relationship that we understand, but a machine that's just doing pattern matching can't be expected to understand. For instance "boss" and "employee". For instance "waiter" and "customer". For instance "manager" and "player" (in a football context) or "manager" and "artist" (in a music context) or "manager" and "customer" (in a bank context). And what's the inverse of "customer" now? And so on and so on...

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.

soco• 30 days ago

We have both Sean Young and Sean Bean. Black swans still exists and the pretraining cannot rely on assumptions - provided if you want answers, not hallucinations.

YetAnotherNick• 29 days ago

So by extension, the question "The first letter of Oman is _", and "O is the first letter of country _" the same for humans?

whilenot-dev2• 30 days ago

Here's a short example from A. Karpathy in a 2024 video: https://www.youtube.com/watch?v=zjkBMFhNj_g&t=750s

josefritzishere• 30 days ago

The premise here is false. AI does not learn. It is a word guessing machine. I know that some of this is the semantics of how we describe these analogs but pretending that an LLM can learn does not advance the topic.

bigstrat2003• 29 days ago

Yes, because they can't reason. This is well known, and should be completely unsurprising. LLMs don't "learn" anything except that some token is statistically likely to be followed by some other token.

nh23423fefe• 29 days ago

cool assertion from 2023

dang• 29 days ago

Discussed at the time:

LLMs trained on “A is B” fail to learn “B is A” - https://news.ycombinator.com/item?id=37621999 - Sept 2023 (158 comments)

jeremysalwen• 29 days ago

I am surprised nobody links this blog post demonstrating that the paper's conclusion is not true (even for gpt3.5): https://andrewmayne.com/2023/11/14/is-the-reversal-curse-rea...

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.

The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A" (2023)

Related Stories