4 ms·
Why not "simply" multigen every (important) query and take the statistical average? Hallucinations are random, the truth isn't. This is absurdly expensive wit
by dontupvoteme 3y ago
Why not "simply" multigen every (important) query and take the statistical average?
Hallucinations are random, the truth isn't.
This is absurdly expensive with GPT4, cheaper with 3, and dirt cheap locally with LLaMA
- xmprt 3y agoThat only works if the generated outputs are completely independent and not correlated. I'd be interested in research that shows whether multigen actually reduces hallucination rates.
- dontupvoteme 3y agoTrue, I'm just throwing multigen out there as a wild ass solution However you could do multigen across different models, e.g. GPT/Claude/LLaMA which should not correlate entirely
- thatjoeoverthr 3y agoJust try it!
- dontupvoteme 3y agoI'll give this a shot after work I think. The question I have is what's a prompt which reliably hallucinates but still produces the correct answer some of the time? I know it gets some python functions "wrong", but i think they were actually "right" in the version it was trained on, so software seems out.
- jiggawatts 3y agoThat’s an established technique with papers written on the topic and everything. Anecdotally I tested this by having GPT4 translate Acadian cuneiform — which it can just barely do. I had it do this four times and it returned four gibberish answers. I then prompted it with the source plus the four attempts and asked for a merged result. It did it better than the human archeologists did! More readable and consistent. I compared it with the human version and it matched the meaning. Expensive now… soon to be standard?
- dontupvoteme 3y agoMakes sense, it's too obvious to not have already been studied :) Do you know what I might search for to find info on it? >Expensive now… soon to be standard? With translation, better safe than sorry? It's a very important field and preserves human history so, why not?
- jiggawatts 3y agoIt's trivial to test, just use ChatGPT yourself and ask it to solve the same problem several times in new sessions. Then paste in all attempts and ask for a combined result. The main issue is context length: if you use 4 attempts you have to have to fit in the original question, four temporary answers, and the final answer. So that's 6 roughly equal sized chunks of text. With GPT4's 8K limit that's just 1300 tokens per chunk, or about 900 words. That's not a lot! The LLMs with longer context windows are not as intelligent, and tend to miss details or they don't follow instructions as accurately. Right now this is just a gimmick that demonstrates that more intelligence can be squeezed out of even existing LLMs...
- intended 3y agoI don’t know if I will sound rude, but your example itself illustrates the crux of problem. The only way you could know that the output was wrong was because you could verify it in the first place. You can’t verify answers for questions in unfamiliar domains - or even for novel questions in your own domain. Hah, it feels like a weird version of P!=NP.
- intended 3y agoHow? Truth is a human thing. Statistically averaging out 4, 5, 6, N text generations from an LLM will not converge to any “truth”. You have essentially stated that outputs from a text generator are normally distributed around “Facts”. May I gently suggest, that and older quote about an infinite number of simians, typewriters and the works of Shakespeare, is more appropriate ?
- marcosdumay 3y agoWhat makes you say that neural networks hallucinations are random? There's absolutely no reason for them to be.