I think this article pushes the premise farther than is reasonable.
The best anyone expects from an LLM "truth vector" is that it would encode the model's belief about whether the statement is true. Of course a perfect truth oracle is impossible.
> The best anyone expects from an LLM "truth vector" is that it would encode the model's belief about whether the statement is true.
I think even that's too-optimistic: The model is a document-extender, so its real "truth" is whether a token seems like it would would fit-next in a partial document, based on prior documents. (This is usually not the kind of analytic truth we're interested in.)
Instead, if we peek at vectors and weights, we'll probably end up measuring the mood and tone of earnestness and conviction for subsequent output tokens, such as dialogue being emitted for a fictional character or narrator. It'd be like measuring "mockery" or "mournfulness" or "uncertainty".
If our document described the character as Yoda, then "the force" is probably truthy. Described as Peter Pan, flying with pixie dust is truthy. If described as helplessly naive and credulous, and almost anything "said" can be truthy.
LLMs are not optimized only for truth they are optimized for a more complicated objective that includes e.g. humans liking their output. It is a universal truth that to get humans to like you, you have to lie to them.
Fun, though as hinted at the end, the point of LLM "truth" probes is to measure the model's internal judgment of truthfulness. There's no reason this judgment, even if measured with 100% accuracy, couldn't be mistaken or logically inconsistent.
A basic course in statistics will inform you of why a 99.99% accurate test should be looked at with skepticism when diagnosing a rare disease. Yet we see the fancy 9s and think somehow this many 9s is enough.
It looks like Zach Weinersmith predicted this exact line of research 11 years ago [1], when he suggested testing the liar sentence using fMRI.
The same analysis applies: the probe tells us what the LLM thinks about the truth value if the sentence, not the truth value of the sentence. I don't think anyone claimed that these probes were truth oracles.
Aren't there definitions of Truth that are not the negation of Falsehood? Can't there be a function True(x) that is not equal to !False(x)? Can't there be a third function Paradox(x) such that these counterexamples can be considered paradoxes and therefore outside of the truth?
I'm admittedly not a logician and don't formally study paradoxes, but I never quite understood the whole category of "this sentence is false" paradoxes, I feel that the sentence is paradox and therefore it's neither false nor true.
I do agree that it's a truth vector sounds like a silly panacea fantasy, though. But more logical formality is not the counter argument that would convince me of it, rather I believe that there's less formal and rigorous ways to get closer to truth.
I’m not a logician either but believe this is what Tarski’s definition of truth solves for. In order to make a statement about the statement itself, you have to introduce a new meta language. Then a statement in the meta language is only true if the underlying statement is true.
Interesting point about "self-referential sentences". I tend to agree. In my view a sentence saying something like "This sentence ..." does not have valid semantic meaning. It says nothing because, what "This" in "This sentence" means is ill-defined.
If terms we use are not well-defined, then sentences using such terms can not have meaning.
But for the sake of argument let's explore, what could the "this" in (so called) "self-referential" sentences refer to?
Do they refer to a specific encoding of the sentence you are reading, as some bits in computer memory perhaps?
That would require that those bits somehow have a unique "identity" and the "this" in a self-referential sentence would have to refer to those bits in specific addresses of a specific memory-chip.
But of course the "this" does not specify which memory chip, which specific (concrete) encoding of its (purported) meaning it is referring to. And if it did, then it would be talking about that specific set of bits in that specific memory-chip, not of "itself".
The fallacy is that what we perceive as a "sentence we read" is somehow "speaking" of something. But no, the sentence is not a subject, a sentence can not speak, and THEREFORE it specifically can not speak of itself.
A sentence can not speak, only actors, only subjects, like humans and AI, can "speak". And their speech must be encoded in some physical medium. A written sentence like "This sentence is ..." gives us the false impressions that somehow the SENTENCE IS SPEAKING of itself!
But speech can not speak, speech is the product of speaking.
Hence, in my view, "self-referential sentences" do not have any meaning and whatever paradoxes they might seem to create are results of confusion between ontological levels of "Subject" vs. "Speech".
As the article notes, the sentence "This sentence is written in English" is well understood, and true. "This sentence is written in French" is also well understood and false.
The best anyone expects from an LLM "truth vector" is that it would encode the model's belief about whether the statement is true. Of course a perfect truth oracle is impossible.