Yes, of the languages I know, all the LLM's get consistently "stupider" as the languages get "smaller" (in terms of available data on the net). Even if the LLM can translate fairly well into the language and speak in the language, the reasoning skills are simply not as good as in English, and progressively worse as languages get smaller, and frequency of hallucinations increases substantially.
There's a core problem with LLMs: they learn sentences, not facts.
So an LLM may learn a ton of English-language sentences about cats, and much fewer Spanish sentences about gatos. And it even learns that cat-gato is a correct translation. But it does not ever figure out that cats and gatos are the same thing. And in particular, a true English-language fact about cats is still true if you translate it into Spanish. So the LLM might be a genius if you ask it about cats in English, but in Spanish it might tell you "gatos tienen tres patas" simply because OpenAI didn't include enough Spanish biology books. These machines are just unfathomably dumb.
> but in Spanish it might tell you "gatos tienen tres patas"
Have you actually had a State of the art LLM do something like this?
Because this
>But it does not ever figure out that cats and gatos are the same thing. And in particular, a true English-language fact about cats is still true if you translate it into Spanish
is just untrue. You can definitely query knowledge only learnt in one language in other languages.
I wasn't talking about "state of the art LLMs," I am aware that commercial offerings are much better trained in Spanish. This was a thought experiment based on comments from people testing GPT-3.5 with Swahili.
> You can definitely query knowledge only learnt in one language in other languages.
Do you have a source on that? I believe this is simply not true, unless maybe the pretraining data has enough context-specific "bridge translations." And I am not sure how on earth you would verify that any major LLM only learnt something in one language. What if the pretraining data includes machine translations?
Frustratingly, just a few months ago I read a paper describing how LLMs excessively rely on English-language representations of ideas, but now I can't find it. So I can't really criticize you if you don't have a source :) The argument was essentially what I said above: since LLMs associate tokens by related tokens, not ideas by related ideas, the emergent conceptual relations formed around the token "cat" do not have any means of transferring to conceptual relations around the token "gato."
>I wasn't talking about "state of the art LLMs," I am aware that commercial offerings are much better trained in Spanish. This was a thought experiment based on comments from people testing GPT-3.5 with Swahili.
A thought experiment from other people comments on another language. So...No. Fabricating failure modes from their personally constructed ideas about how LLMs work seems to be a frustratingly common occurrence in these kinds of discussions.
>Frustratingly, just few months ago I read a paper describing how LLMs excessively rely on English-language representations of ideas, but now I can't find it.
That the models are as proficient as they are is evidence enough of knowledge transfer clearly happening. https://arxiv.org/abs/2108.13349. If you trained a model on the Catalan tokens GPT-3 was trained on alone, you'd just get a GPT-2 level gibberish model at best. I don't doubt you, i just don't think it means what you think it means.
As for papers, these are some interesting ones.
How do languages influence each other? Studying cross-lingual data sharing during LLM fine-tuning - https://arxiv.org/pdf/2305.13286
It's not like there is perfect transfer but the idea that there's none at all seemed so ridiculous to me (and why i asked the first question). Models would be utterly useless in multilingual settings if that were really the case.
I agree with you, and we seem to hold a minority opinion. LLMs contain a LOT of information and are very articulate - they are language models after all. So they seem answer questions well, but fall down on thinking/reasoning about the information they contain.
But then they can play chess. I'm not sure what to make of that. Such an odd mix of capability and uselessness, but the distinction is always related to something like understanding.
LLMs trained on just human text are indeed limited, imitative models. But when they train from the environment they can surpass humans, like AlphaZero or AlphaGeometry
Of the open source LLMs I've tried, all suck in non-English. I imagine it's difficult to make an LLM work in tens of languages on a consumer computer.