Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Language models still struggle with the concept of negation

quantamagazine.org

Post reply on HN
Print view View original

131–140 of 172 posts

Re: Language models still struggle with the concept of negation

#131

Post by bluepoint »

akavi wrote: post #14

Does this increase anyone else's sense that LLMs are actually modeling something similar to the way humans think? It certainly does for me. I find that negations/conjugations of negations are surprisingly difficult for to parse. Beyond two negations at most, I can feel myself having to switch into "code" mode, where I'm explicitly casing out the booleans, to understand them. Obviously the examples given in the articl…

The negations have multiple roles. An important one is that they add “color saturation” to the words. “It is bad”, “It is good”, “It is not bad”, “I can’t say it’s good”, “I can’t say it’s not bad”, are all different statements. So negations are difficult, maybe even for humans…

Re: Language models still struggle with the concept of negation

#132

Post by petepete »

School-Cotton wrote: post #103
petepete wrote: post #99

Earlier quoted context omitted.

An incorrect stock phrase that means the exact opposite of what's been said. I don't think it's unreasonable to say people who get things entirely wrong are struggling.

The whole point of stock phrases, idioms, expressions, etc. is that their meaning is “non-compositional”; i.e., they mean something different from the literal combination of words that make them up. Including even extreme cases where they can mean the opposite, like this one. That doesn’t mean they’re “wrong”, it’s just a quirk of human language.

It does. "I couldn't care less" is a common term and "I could care less" is a misunderstanding.

I'd hope more people would realise they're saying they _do care_ about something they're claiming to not care about. It sounds stupid to a native English speaker.

Re: Language models still struggle with the concept of negation

#133

Post by stubybubs »

famouswaffles wrote: post #43
tedunangst wrote: post #36

Earlier quoted context omitted.

Variants of common riddles remain the final frontier. I'm trying to cross the river in a canoe with a carrot, cabbage, and cucumber...

Variants of common riddles actually can be solved with GPT-4, but you have to rewrite it so it doesn't look like the riddle from memory(sometimes, it's as easy as changing names to something completely different). Turns out Language models trust their memory quite a bit. Slightly related - they won't actually use the results of tools if it differs a lot from what it expects the output to be - https://vgel.me/posts/to…

This really just makes it seem like it's not reason at all. The trick (or rather un-trick) here is that you can look at the bulbs as many times as you want. Even if I explicitly tell GPT4 that, it doesn't get it.

It's not reason, it's mapping.

Re: Language models still struggle with the concept of negation

#134

Post by nathan_compton »

famouswaffles wrote: post #114
nathan_compton wrote: post #105

Earlier quoted context omitted.

"Language models trust their memory quite a bit." All they have is memory, either in the weights or the input prompt. To the extent that these models appear to reason, it is precisely in the ability to successfully substitute information from the prompt into reasoning patterns in the training data. It shouldn't be any surprise that this fails when patterns in the prompt strongly condition the model to reproduce parti…

LLMs trained on code reason better. Perform better on reasoning benchmarks even if the benchmarks have nothing to do with code. You're wrong. https://arxiv.org/abs/2210.07128

Code is often just a sequence of steps (sometimes with comments to indicate goals). As such, it is just another form of patterns of reasoning. Many chains of thought that you would utilize in code are useful skeletons to think about other things.

I don't see how this undermines my point.

Re: Language models still struggle with the concept of negation

#135

Post by School-Cotton »

petepete wrote: post #132
School-Cotton wrote: post #103

Earlier quoted context omitted.

The whole point of stock phrases, idioms, expressions, etc. is that their meaning is “non-compositional”; i.e., they mean something different from the literal combination of words that make them up. Including even extreme cases where they can mean the opposite, like this one. That doesn’t mean they’re “wrong”, it’s just a quirk of human language.

It does. "I couldn't care less" is a common term and "I could care less" is a misunderstanding. I'd hope more people would realise they're saying they _do care_ about something they're claiming to not care about. It sounds stupid to a native English speaker.

You’re just repeating the same point over and over rather than actually giving any argument.

> It sounds stupid to a native English speaker

You might be a native English speaker, but you’re not a native speaker of American English.

Re: Language models still struggle with the concept of negation

#136

Post by famouswaffles »

nathan_compton wrote: post #134
famouswaffles wrote: post #114

Earlier quoted context omitted.

LLMs trained on code reason better. Perform better on reasoning benchmarks even if the benchmarks have nothing to do with code. You're wrong. https://arxiv.org/abs/2210.07128

Code is often just a sequence of steps (sometimes with comments to indicate goals). As such, it is just another form of patterns of reasoning. Many chains of thought that you would utilize in code are useful skeletons to think about other things. I don't see how this undermines my point.

If code transfers to reasoning tasks that don't have anything to do with code then what is being "substituted" ? Ideas and concepts ?

Code and MMLU don't share similar "reasoning patterns" unless you're being extremely vague. In the, "they both require reasoning" sense.

Re: Language models still struggle with the concept of negation

#137

Post by foobarbecue »

byyyy wrote: post #123
foobarbecue wrote: post #122

Earlier quoted context omitted.

[flagged]

Birds don't have paws because they aren't quadrupeds according to websters. Bats walk on fours. They are quadrupeds therefore they have paws according to Merriam Webster. Why don't you address the point I brought up? I already completely understand your definition no need to reiterate it. However, there is a clear disconnect between your definition of paw and the definition from Merriam Webster. Please address it.

[flagged]

Re: Language models still struggle with the concept of negation

#138

Post by foobarbecue »

diceduckmonk wrote: post #73
hn_throwaway_99 wrote: post #11

Earlier quoted context omitted.

While the answer landed on bats, ChatGPT kind of trips itself up in its explanations: 1. "Most winged creatures either have paws (like bats) or lay eggs (like birds and insects)" - but bats don't have paws. 2. "However, if we consider the term "wings" more broadly" - why do we need to consider the term "wings" more broadly? Nobody is arguing there is some definition of "wing" that would exclude bats. Bat wings evolve…

> "there are some species of bats that give live birth" - all species of bats give live birth ChatGPT is being more logical to be honest. It’s true, some species of bats give live birth. To assume all bats give live birth might be committing the fallacy of induction.

That's true -- some bats never have kids. They just have lots of hobbies and are very fulfilled.

Re: Language models still struggle with the concept of negation

#139

Post by foobarbecue »

maxbond wrote: post #128
foobarbecue wrote: post #118

Earlier quoted context omitted.

I think we're getting totally taken for a ride... byyyy can't actually be serious right...? Anyway I'm enjoying imagining bats with little paws, it's an amusing image.

Well, if you're looking for my opinion, I don't think they're trolling. But I do think both of you were (intentionally or unintentionally) goading each other from a disagreement to an argument with your rhetoric (eg, where you said, "let's assume that's a reasonable source", or when they said "you obviously sourced this it just didn't match your point so you discarded it" - that brings up the temperature of the conve…

[flagged]

Re: Language models still struggle with the concept of negation

#140

Post by SkyPuncher »

pillefitz wrote: post #49
BoiledCabbage wrote: post #26

Earlier quoted context omitted.

Maybe I'm missing something, but this seems like a really poorly worded riddle. Or just a bad riddle. Given what you stated, the solution would be walk upstairs look at the 3 bulbs, walk downstairs switch one of the switches. Walk upstairs see which bulb changed on/off. Repeat for the other two switches. This takes roughly 5 mins rather than your solution which takes 3 years.

You're missing and getting the point at the same time. Your solution would be the correct one, while GPT pulls out the "only look once"-constraint out of nowhere. The riddle is perfectly fine.

This is the crux of the problem for GPT, enough of the text is from the “standard” riddle that is cannot break from it to actually answer the question.

I’m seeing this all of the time when trying to get GPT to perform certain answering problems. It’s heavily biased towards the “correct” answer - even when the prompt presents directly contradictory information.

Post reply on HN