Researchers describe how to tell if ChatGPT is confabulating(arstechnica.com) |
Researchers describe how to tell if ChatGPT is confabulating(arstechnica.com) |
Why is it compelled to provide one, anyway?
Which is to say, why is the output of each model layer a raw softmax — thus discarding knowledge of the confidence each layer of the model had in its output?
Why not instead have the output of each layer be e.g. softmax but rescaled by min(max(pre-softmax vector), 1.0)? Such that layers that would output higher than 1.0 just get softmax'ed normally; but layers that would output all "low-confidence" results (a vector all lower than 1.0) preserve the low-confidence in the output — allowing later decoder layers to use that info to build I-refuse-to-answer-because-I-don't-know text?
1. An LLM's mathematical "confidence" of having a clear best-scoring candidate for the predicted next token when given a list of tokens.
2. A not-yet-invented AI that models the idea of different entities interacting, the concept of questions and answers, the concept of logical conflicts, and it's "confidence" that a proposition is compatible with other "true" propositions and incompatible with false ones.
To help illustrate the difference, suppose you trained an LLM on texts where a particular question was always answered with "I don't know, I have zero confidence in anything anymore." Later the LLM will regurgitate similarly nihilistic text, and by all objective internal measures it will be extremely "confident" as it does so.
> Why is it compelled to provide one, anyway
It's following the patterns in its training data, which probably reflects a whole lot more people trying to provide answers (sometimes even deliberately wrong ones) as opposed to admitting uncertainty.
This is especially true if developers put their thumb on the scale by injecting primer-text like "You are an intelligent computer eager to provide answers", as opposed to "behave like Socrates and help people understand that nothing is truly knowable."
After all, what I'm describing is something that even a classical Bayesian spam-filter classifier RNN can pull off — where a hidden layer near the output layer can notice that either:
1. the preceding layers have generated a confidence for both the "spam" or "ham" classifications that is not differentiable from 0 by at least epsilon, or
2. the preceding layers have generated a confidence for both the (mutually exclusive) "spam" and "ham" categories that are indistinguishable (not at least epsilon apart post-softmax)
...and in those cases will output "I DUNNO (TRY GREYLISTING IT)" rather than "SPAM (BLOCK IT)" or "HAM (PASS IT THROUGH)".
What I'm expecting to accomplish with a rescaled softmax output (or by other embeddings as long as they propagate/multiply confidences of each successive layer, allowing confidence to approach 0), is to allow some attention-head at some late layer in the model, to develop an overriding-output strategy that reacts to "not differentiable from epsilon" residual confidence in the previous layer's output vector (= the current layer's Q vector), by giving high confidence to an "I don't know the first thing about what you're saying; I didn't really 'get' what you wrote" concept in the current layer's output vector (so high that it overrides any other response at that layer.) This then just gets produced as a response by the same machinery that generates well-embedded responses from concepts at other layers. (Think alignment, not hard-trained fixed outputs.)
---
Though, thinking more carefully about it, something else is missing too. Since LLMs already have all the info available in later layers to recognize condition #2 above (as even under a pure Transformer decoder mask+add+norm+softmax kernel, it's still possible to do math that recognizes when the first N top-P-ranked elements of a vector of mutually-exclusive concepts are not differentiable by at least epsilon, and develop a special reaction for that case) — but they still don't tend to learn this.
I think the concept missing here, is a training technique that supervised training of simpler classifiers has done for forever, but which doesn't seem to come up at all in Transformer training frameworks. And that's dynamically generating the training label for an example input, based on aggregate statistical information output through a side-channel while running inference on the example input. I.e. training the model to have a specific reaction to its own internal state in response to an input, rather than to the input itself.
Let's say you want to use an LLM as a spam-classifier — given an input, have it output a classification {SPAM, HAM, DUNNO}. It's easy enough, just with a dataset of labelled exampels, to take any LLM and do a single fine-tune that results in a {SPAM, HAM} classifier. But you don't want a static dataset of DUNNO examples — because you don't want the classifier to output DUNNO when you aren't sure. You want the classifier to output DUNNO when it isn't sure.
So let's say you do two fine-tunes instead. The first one acts as an encoder, outputting a two-element vector (SPAM confidence, HAM confidence). And the second one acts as a decoder, turning those into categories.
What you actually need to achieve "correct" DUNNO outputs, is to train the decoder fine-tune not on labelled training examples, but by taking your existing (labelled!) training dataset; running it through the model with the decoder not connected, to get the raw confidences; applying a confidence-gating measure to them; and then, for any example that doesn't pass that measure, training the decoder (as a standalone LoRA) to output DUNNO.
(disclaimer: I'm not an ML expert, maybe this is just a me problem, but I find nets extremely sensitive to stuff like changing activations, adding normalisation, layer initialisation, layer sizes, all this stuff that seems kinda arbitrary to a non-experts like me)
So I suppose if you want to stay on the right side of those people - or you are one - you call it confabulation instead.
To fill in gaps in one's memory with fabrications that one believes to be facts.
…is much more accurate.
Since we’re talking about a technical process it helps to be more precise in our use of language.
> Here we develop new methods grounded in statistics, proposing entropy-based uncertainty estimators for LLMs to detect a subset of hallucinations—confabulations—which are arbitrary and incorrect generations.
A figure from the paper shows this better than my TL;DR: https://www.nature.com/articles/s41586-024-07421-0/figures/1
(See also: Androids that resemble men but aren't, asteroids that resemble stars but aren't, meteoroids that resemble meteors but aren't...)
It's also interesting to see what temperature value they use (1.0, 0.1 in some cases?)... I have a feeling using the actual raw probability estimates (if available) would provide a lot of information without having to rerun the LLM or sample quite as heavily.
You could use the same technique that this paper describes to compare the answers each LLM gave. LLMs don’t have to be in opposition to traditional NLP techniques
We have focused on the inherent lack of input context, leading to wrong conclusions, but what about that 90B+ parameters universe, plenty of room for multiple contexts to associate any input to surprising pathways.
In the olden days of MLPs we had the same problem with softmax basically squeezing N output scores into a normalized “probability”, where each output neuron actually was the sum of multiple weighted paths, which one winning the softmax made up the “true” answer, but there may as well have been two equally likely outcomes, with just the internal “context” as difference. In physics we have the path integral interpretation and I dare say, we humans too, may provide outputs that are shaped by our inner context.
This article seems rather contrived. They present this totally broken idea of how LLMs work (that they are trained from the outset for accuracy on facts) and then proceed to present this research as it is a discovery that LLMs don't work like that.
If it's sure it won't confirm it both ways.
(I still don't understand why everyone seems happy to conflate "intelligence" with fact-retrieval?)
Because it's useful and impossible till very recently.
This assertion in the article doesn't seem right at all. When LLMs weren't trained for accuracy, we had "random story generators" like GPT-2 or GPT-3. The whole breakthrough with RLHF was that we started training them for accuracy - or the appearance of it, as rated by human reviewers.
This step both made the models a lot more useful and willing to stick to instructions, and also a lot better at... well, sounding authoritative when they shouldn't.
Is it plausible that LLM’s get so smart that we can’t understand them. Do we spend like years trying to validate scientific theories confabulated by AI?
In the run up to super-intelligence, it seems like we’ll have to tweak the creativity knobs up, like the whole goal will be to find novel patterns humans don’t find, is there a way to tweak those knobs that get us super genius and not super conspiracy theorist? Is there even a difference? Part of this might depend on whether or not we think we can feed LLM’s “all” the information.
But in fact, assuming that Silicon Valley CEO’s are some of the smartest people in the world, I might argue that confabulation of a possible future is in fact their primary value. Not being allowed to confabulate is incredibly limiting.
It's just that usually the words it generates are accurate enough for my needs.
This feedback is used for training.