4 ms·

> let me put it to you this way: what is the LLM doing differently when it generates tokens that are "wrong" compared to when the tokens are "right"? It is con

by gwern 2y ago

> let me put it to you this way: what is the LLM doing differently when it generates tokens that are "wrong" compared to when the tokens are "right"? It is conditioning on latents about truth, falsity, reliability, and calibration. All of these inferred latents have been shown to exist inside LLMs, as they need to exist for LLMs to do their jobs in accurately predicting the next token. (Imagine trying to predict tokens in, say, discussions about people arguing or critiques of fictional stories, or discussing mistakes made by people, and not having latents like that!) LLMs also model other things: for example, you can use them to predict demographic information about the authors of texts (https://arxiv.org/abs/2310.07298 https://arxiv.org/abs/2310.07298), even though this is something that pretty much never exists IRL, a piece of text with a demographic label like "written by a 28yo"; it is simply a latent that the LLM has learned for its usefulness, and can be tapped into. This is why a LLM can generate text that it thinks was written by a Valley girl in the 1980s, or text which is 'wrong', or text which is 'right', and this is why you see things like in Codex, they found that if the prompted code had subtle bugs, the completions tend to have subtle bugs - because the model knows there's 'good' and 'bad' code, and 'bad' code will be followed by more bad code, and so on. This should all come as no surprise - what else did you think would happen? - but pointing out that for it to be possible, the LLM has to be inferring hidden properties of the text like the nature of its author, seems to surprise people.

timr 2y ago
> It is conditioning on latents about truth, falsity, reliability, and calibration. All of these inferred latents have been shown to exist inside LLMs, as they need to exist for LLMs to do their jobs in accurately predicting the next token. No, it isn't, and no, they haven't [1], and no, they don't. The only thing that "needs to exist" for an LLM to generate the next token is a whole bunch of training data containing that token, so that it can condition based on context. You can stare at your navel and claim that these higher-level concepts end up encoded in the bajillions of free parameters of the model -- and hey, maybe they do -- but that's not the same thing as "conditioning on latents". There's no explicit representation of "truth" in an LLM, just like there's no explicit representation of a dog in Stable Diffusion. Do the thought exercise: if you trained an LLM on nothing but nonsense text, would it produce "truth"? LLMs "hallucinate" precisely because they have no idea what truth means. It's just a weird emergent outcome that when you train them on the entire internet, they generate something close to enough to truthy, most of the time. But it's all tokens to the model. [1] I have no idea how you could make the claim that something like a latent conceptualization of truth is "proven" to exist, given that proving any non-trivial statement true or false is basically impossible. How would you even evaluate this capability?
- rustcleaner 2y ago
  My sense is an LLM is like Broca's area. It might not reason well, but it'll make good sounding bullshit. What's missing are other systems to put boundaries and tests on this component. We do the same thing too: hallucinate up ideas reliably, calling it remembering, and we do one additional thing: we (or at least the rational) have a truth-testing loop. People forget that people are not actually rational, only their models of people are.
- kromem 2y ago
  This was AFAIK the first paper to show linear representations of truthiness in LLMs: https://arxiv.org/abs/2310.06824 https://arxiv.org/abs/2310.06824 But what you should really read over is Anthropic's most recent interpretability paper.
  - timr 2y ago
    > In this work, we curate high-quality datasets of true/false statements and use them to study in detail the structure of LLM representations of truth, drawing on three lines of evidence: 1. Visualizations of LLM true/false statement representations, which reveal clear linear structure. 2. Transfer experiments in which probes trained on one dataset generalize to different datasets. 3. Causal evidence obtained by surgically intervening in a LLM's forward pass, causing it to treat false statements as true and vice versa. Overall, we present evidence that language models linearly represent the truth or falsehood of factual statements. You can debate whether the 3 experiments cited back the claim (I don't believe they do), but they certainly don't prove what OP claimed. Even if you demonstrated that an LLM has a "linear structure" when validating true/false statements, that's whole universe away from having a concept of truth that generalizes, for example, to knowing when nonsense is being generated based on conceptual models that can be evaluated to be true or false. It's also very different to ask a model to evaluate the veracity of a nonsense statement, vs. avoiding the generation of a nonsense statement. The former is easier than the latter, and probably could have been done with earlier generations of classifiers. Colloquially, we've got LLMs telling people to put glue on pizza. It's obvious from direct experience that they're incapable of knowing true and false in a general sense.
kromem 2y ago
One of the surprising results in research lately was the theory of mind paper the other week that found around half of humans failed the transparent boxes version of the theory of mind questions - something previously assumed to be uniquely a LLM failure case. I suspect over the next few years we're going to see more and more behaviors in LLMs that turn out to be predictive of human features.

by gwern 2y ago

timr 2y ago

> It is conditioning on latents about truth, falsity, reliability, and calibration. All of these inferred latents have been shown to exist inside LLMs, as they need to exist for LLMs to do their jobs in accurately predicting the next token. No, it isn't, and no, they haven't [1], and no, they don't. The only thing that "needs to exist" for an LLM to generate the next token is a whole bunch of training data containing that token, so that it can condition based on context. You can stare at your navel and claim that these higher-level concepts end up encoded in the bajillions of free parameters of the model -- and hey, maybe they do -- but that's not the same thing as "conditioning on latents". There's no explicit representation of "truth" in an LLM, just like there's no explicit representation of a dog in Stable Diffusion. Do the thought exercise: if you trained an LLM on nothing but nonsense text, would it produce "truth"? LLMs "hallucinate" precisely because they have no idea what truth means. It's just a weird emergent outcome that when you train them on the entire internet, they generate something close to enough to truthy, most of the time. But it's all tokens to the model. [1] I have no idea how you could make the claim that something like a latent conceptualization of truth is "proven" to exist, given that proving any non-trivial statement true or false is basically impossible. How would you even evaluate this capability?

rustcleaner 2y ago
My sense is an LLM is like Broca's area. It might not reason well, but it'll make good sounding bullshit. What's missing are other systems to put boundaries and tests on this component. We do the same thing too: hallucinate up ideas reliably, calling it remembering, and we do one additional thing: we (or at least the rational) have a truth-testing loop. People forget that people are not actually rational, only their models of people are.
kromem 2y ago
This was AFAIK the first paper to show linear representations of truthiness in LLMs: https://arxiv.org/abs/2310.06824 https://arxiv.org/abs/2310.06824 But what you should really read over is Anthropic's most recent interpretability paper.
- timr 2y ago
  > In this work, we curate high-quality datasets of true/false statements and use them to study in detail the structure of LLM representations of truth, drawing on three lines of evidence: 1. Visualizations of LLM true/false statement representations, which reveal clear linear structure. 2. Transfer experiments in which probes trained on one dataset generalize to different datasets. 3. Causal evidence obtained by surgically intervening in a LLM's forward pass, causing it to treat false statements as true and vice versa. Overall, we present evidence that language models linearly represent the truth or falsehood of factual statements. You can debate whether the 3 experiments cited back the claim (I don't believe they do), but they certainly don't prove what OP claimed. Even if you demonstrated that an LLM has a "linear structure" when validating true/false statements, that's whole universe away from having a concept of truth that generalizes, for example, to knowing when nonsense is being generated based on conceptual models that can be evaluated to be true or false. It's also very different to ask a model to evaluate the veracity of a nonsense statement, vs. avoiding the generation of a nonsense statement. The former is easier than the latter, and probably could have been done with earlier generations of classifiers. Colloquially, we've got LLMs telling people to put glue on pizza. It's obvious from direct experience that they're incapable of knowing true and false in a general sense.

kromem 2y ago

One of the surprising results in research lately was the theory of mind paper the other week that found around half of humans failed the transparent boxes version of the theory of mind questions - something previously assumed to be uniquely a LLM failure case. I suspect over the next few years we're going to see more and more behaviors in LLMs that turn out to be predictive of human features.