[phagenlocher]

I am surprised that anybody is surprised about this. LLMs don't even understand relations and even if they did: Why would they assume that relations are symmetric? It would be incorrect to do so for the general case anyways.

[BoorishBears]

I'm staunchly against people who want to, for example, claim LLMs have personhood because you can get them to act like people... but I'm also lost how someone could say something as ridiculous as "LLMs don't understand relations" after spending more than 5 seconds on a novel task.

It's like "understand" has been repurposed to mean "understand in the way that I understand as a human" even though human cognition is:

a) still an area of deep study that people making most of these declarations are completely unfamiliar with (not unlike the over-anthropomorphizing crowd being unfamiliar with LLMs)

b) often hypothesised on the basis of probabilistic models that don't support somehow "gatekeeping" the concept of understanding: https://www.cell.com/trends/cognitive-sciences/fulltext/S136...

[great_psy]

The part that is surprising, is where the reverse does not have a higher than random probability.

Sure, most of the time “is” is not commutative, but sometimes it is, so it should be higher probability than guessing.

[phagenlocher]

I don't find that surprising, as LLMs typically predict the next token in a string of tokens not the previous. Just because we, as humans, have the knowledge that "is" can be a symmetric relation does not imply that LLMs have this knowledge.

[ReptileMan]

From my rather rudimentary understanding LLMs are just collection of X billion ifs with stats applied. The brilliant part is that the LLM training can generate different ifs depending on the data. They don't have logic or something like that implemented. Which makes even more impressive the results we are getting.