Ted Underwood 🦋: A team centered at McGill and Harvard argues that LLM "hallucination" is better understood as confabulation, and resembles human habits of using narrative for sense-making. #machinelearning 🤖 arxiv.org/abs/2406.04175

A team centered at McGill and Harvard argues that LLM "hallucination" is better understood as confabulation, and resembles human habits of using narrative for sense-making. #machinelearning 🤖 arxiv.org/abs/2406.04175

Confabulation: The Surprising Value of Large Language Model Hallucinationsarxiv.org This paper presents a systematic defense of large language model (LLM) hallucinations or 'confabulations' as a potential resource instead of a categorically negative pitfall. The standard view is...

Ironically, people who decry LLM and their hallucinations the hardest seem more prone to confabulations themselves. Sadly, the similarity seems unlikely to breed understanding.

Nice to see this. The confabulation term came to me from my mom’s hospital stay when she broke her hip Dec 2022, though Rodney Brooks also used it long before in a different context arstechnica.com/information-...

I vaguely remembered someone had made this case early on, in journalism or on a blog. So are you the source? I’m attuned to this kind of thing because I blog (and microblog!) enough that I suspect I’m an unacknowledged source for several now-common metaphors :)

Yes I think I coined using "confabulation" as an alternative to hallucination with LLMs. It spread widely because Wikipedia cited my article, and it is now in at least one dictionary Others likely used it too here and there, but my use seems to have pushed it into the mainstream. Here's the story:

Worth noting that "confabulation theory" was a thing in computational neuroscience before, but it's not the same thing www.scholarpedia.org/article/Conf... And Rodney Brooks appears to have used "confabulate" in a similar way related to AI in 2007: www.bcs.org/articles-opi...

Confabulation theory (computational intelligence) - Scholarpediawww.scholarpedia.org

It's also telling to me that all the citations in that article for using "confabulation" related to LLMs occur in 2023 after my article 😀 but sadly none seem to cite my article itself. Though I found some other academic papers that do cite my piece

Oh cool, here's one aclanthology.org/2023.emnlp-m...

I see this as insightful and helpful. I imagine that this behavior is similar to what we see from articulate children. We might say that they are fanciful or imaginative. Yet, when we get this same behavior from adults, we describe them as liars and charlatans.

Are we really doing "Harvard researchers say" now?

It's four people. This is how I commonly describe collective authorship while staying under a character limit. 1/2 bsky.app/profile/tedu...

Need to adjust language models to reflect a particular year or period? Researchers at UW show that you can do this in a ridiculously straightforward way. Generate “time vectors”; then interpolate or (um… ) extrapolate them! huggingface.co/papers/2312.... #MLSky

2/2 Though I admit this still trades on the cultural capital of Aarhus. bsky.app/profile/tedu...

Good and important work from researchers at Uppsala and Aarhus. For more than a decade, researchers have been struggling to connect literary canonicity to measures of information density or surprise. The argument is rarely 100% slam-dunk, but this gets close. + aclanthology.org/2024.latechc...

aclanthology.org

"Sui et al." vs. 18 characters to list all their surnames vs. ~30 for "a team centered at McGill and Harvard"

McGill is actually the relevant detail here for people who know DH. Including McGill is a hint that humanists were involved in the project, without having to enumerate four surnames. (And it's not always only four that you have to enumerate). Idgaf about Harvard.

🤷 I guess as a general practice, I don't like naming the universities for credit: People did the work. Sometimes I pick a subset of most relevant people (e.g. a friend or the main author). (I make exceptions when it's the main point: e.g. George Mason about economics or Liberty about free speech.)

In this case, it was the main point (or a main point) if you know the field. I'm writing for an audience of people who know what labs are where. I will happily defer to your rhetorical practices re: the sociology of science, and also on the general principle that I let ppl write what they want.

Post