I noticed long ago that this is true of statistical machine translation too --- giving output with the exact opposite meaning to the input if there were any negationary words or phrases in there.
Language models still struggle with the concept of negation
41–50 of 172 posts
Re: Language models still struggle with the concept of negation
#42What animals don’t have paws or lay eggs, but have wings? Work it out, step by step. ChatGPT (GPT-4): Sure, let's go through this step by step: First, you asked for animals that don't have paws. This would exclude mammals such as dogs, cats, bears, and so on, as these species have paws. Second, you specified animals that don't lay eggs. This eliminates a wide range of animals including all birds, most reptiles, and s…
Wow, some species of bats don't lay eggs. TIL.
?
Bats give live birth.
Re: Language models still struggle with the concept of negation
#43Earlier quoted context omitted.
It's telling that it didn't pick up on the fact that the whole "once you look you can't touch the switches anymore" isn't in this version of the riddle. I mean the obvious strategy in this case is turn on the first switch, go upstairs, look at the bulbs. Go back downstairs and try the second switch. You've not got them all mapped out in about 30 seconds. Using GTP4 I asked if there was a way to do it in less than 3 y…
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/tools-not-needed/
Re: Language models still struggle with the concept of negation
#44The title of the article is kinda the answer at the same time. Chatbots don't know what stuff is. They have no ability to gain knowledge out of learned text, just counting occurences of words in texts and giving them a weight, depending on the relationship in that text. They are just putting combinations of text together. And the concept of negating something related to something else kinda needs an understanding of…
> They have no ability to gain knowledge out of learned text, just counting occurences of words in texts and giving them a weight, depending on the relationship in that text. They are just putting combinations of text together.
No, they use deep neural networks to build a hierarchical semantic model. They are not simple occurrence counters.
Also the current state of the art of LLMs handles negation easily. This article is outdated.
Here's an example from https://openai.com/research/language-models-can-explain-neur...
"Seriously, you guys. I think I found the Mobile Leprechaun from '06. He's been hiding right in front of our eyes."
Token: hiding
layer 0: “verbs in gerund form (ending in 'ing')”
layer 2: “words related to hiding, concealment, or enclosed spaces”
layer 4: “words related to mental states, particularly anxiety and stress”
layer 17: “words and phrases related to silence or quietness”
Re: Language models still struggle with the concept of negation
#45At this point, any article that makes claims about "LLMs" rather than specific model versions lacks credibility. The current version of GPT-4 is very different from most existing LLMs (including the previous version of ChatGPT). The next version release will also be different. Google just released PaLM 2 publicly. It is significantly better than what people saw with the initial Bard versions. They have a code-generat…
Did you read the article or just the title? They mention the specific models the researchers were testing and note that increasing model size did not seem to offer much improvement on this metric. It also ends with a discussion of research into methods for improving performance on queries involving negation.
Re: Language models still struggle with the concept of negation
#46This is one of my favorite test cases: Prompt: Here is a riddle. It is a common riddle but with some changes that make it more difficult. You are an alien that will live for at least 10000 years. You have no sense of temperature. you can not feel hot or cold. you have eyes and can see. you are in a house. downstairs are 3 light switches that control 3 light bulbs that are upstairs. Each light bulb will last for exact…
The reply is funny. I told ChatGPT that it can do multiple trips and that it can see both switches and light bulbs at the same time.
ChatGPT keeps insisting that you have to turn on a switch, wait a few minutes , flip switch to off, turn other switch to on, then go and look and feel the switches. The classic solution to the riddle.
But if you can see both switches and light bulbs, you turn on one switch, you see which bulb turn on. You then turn on the 2nd switch and see which light turn on. You are done. No wait needs to happen ;-)Re: Language models still struggle with the concept of negation
#47Earlier 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's a sample-and-poll approach that I've used with factual question-and-answer based tools that can help with this kind of thing as well. First, you'll notice the added "work it out step by step" to encourage what is called chain-of-thought in the literature. Second, generate say, five responses from the LLM, then take all five responses and uses them in another prompt where you ask the LLM to reflect on all of t…
Then, if that's not accurate, take that entire process and repeat that 5 times, looking for the most common answer.
Re: Language models still struggle with the concept of negation
#48The stop words bit is a bit wtf. Someone didn't do their homework
Yep, that bothered me too. They are called stop words because, but only if, you filter them out. If you don't, they aren't stop words. There is no inherent "stop-ness" there.
Re: Language models still struggle with the concept of negation
#49This is one of my favorite test cases: Prompt: Here is a riddle. It is a common riddle but with some changes that make it more difficult. You are an alien that will live for at least 10000 years. You have no sense of temperature. you can not feel hot or cold. you have eyes and can see. you are in a house. downstairs are 3 light switches that control 3 light bulbs that are upstairs. Each light bulb will last for exact…
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.
Re: Language models still struggle with the concept of negation
#50Earlier 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.
Ah thanks, that clears it up. The formatting made me miss that that was the ChatGPT response. I thought that was the user sharing their solution that they look for.