Comment by dTal | Hacker News Reader

Comment by dTal | Hacker News Reader

Parent Full thread

The whole point is the quantified output. If you just ask an LLM to type out its confidence "manually", it'll make up some nonsense. The logprob numbers are more reliable.

I got this technique to work extremely reliably last year. However there were a bunch of caveats: 1) Firstly, you must institute a check that the multiple choice tokens dominate the output distribution. They should sum to 95% or more, ideally 99%, or the LLM is not following instructions properly. This is also the problem with constrained decoding - if the LLM really doesn't want to output a valid answer, the one you extract will not be high quality. 2) You need to ask it multiple times, permuting which option corresponds to which letter, and average the results. LLMs are surprisingly biased towards picking "A", especially if they're otherwise not sure. 3) For the same reason, performance improves if you frame the prompt as if it were the middle of a quiz. "Question 1" carries baggage that "Question 12" doesn't. 4) You must be exceedingly careful with tokenization.

But when all was said and done, I got a general purpose A/B classifier that gave high resolution quantitative output for the cost of a couple dozen tokens ingested and a couple inference passes.

>The whole point is the quantified output. If you just ask an LLM to type out its confidence "manually", it'll make up some nonsense. The logprob numbers are more reliable.

The whole point of my argument is that neither is good, but from a technical perspective logprobs is probably the worst unless you train a model on specific outputs. In which case you'd throw out the generality again, so when I think about it more, it's actually the worst overall. In my experiments, having the model simply assign "high" or "low" probability in a structured output generally performs best. You can try numbers, but you will never get anything close to what you could expect from traditional ML. And most certainly not from logprobs.

Yes, for vision classifiers we have in prod, I've seen a huge difference between A, B, C, 1, 2, 3 style answers and emitting a string. Even from just base model behavior pre-sft/rl. It was one of those obvious in retrospect moments.

> LLMs are surprisingly biased towards picking "A"

GP pointed at a causal explanation for this: almost every sentence in English that's a statement will start with "A" or "An", so "biased towards picking ''A''" will include most attempts at saying anything long-form for any reason.

I don't think that's the source of the bias I saw. I am confident that my prompting strategy eliminated attempts to generate long form content - specifically, I took care to wrap (A) and (B) in parentheses, so the completion looked like "Answer: (" - with this scheme an LLM is very unlikely to want to write "Answer: (A sentence goes here...". I know this, quantitatively, because I reliably got 99% distribution coverage with only A+B - that is, no inclination to write "The" or other common sentence starter. That's the beauty of the scheme - you can pretty directly and quantitatively validate how well the LLM understood the instructions. You expect it to only output A or B - so does it?

Meanwhile, the bias could be as much as 70% in favor of A in ambiguous cases - a signal completely drowning the <1% inclination to violate the format.

What about switching to numbers or just some random Unicode character like smiley faces. Could be interesting if someone tested what LLMs like to say on a "cold start" lol.

I would also note that models aren't people and don't think like people, so it's also possible that (at least for autoregressive ones) it could just be more likely to say "A" than "B" at that point, not necessarily because of "want" or "reason" but simply because that's what it was trained to do (such as in English writing).

> LLMs are surprisingly biased towards picking "A", especially if they're otherwise not sure.

Not nearly as sophisticated as myself who would mutter "When in doubt - Charlie out" before marking C.

Sure, it was high resolution (precise), how was accuracy compared to Jev (or existing open source implementations of the same concept, like laya)?

Also, Jev/laya do it in one forward pass, for multiple questions about the same state, rather than multiple passes for one question about that state. Well, for the usual multilingual configuration, two forward passes through different small models for laya, but that's because one is the router which chooses which model should do the real work, but still.

I never claimed that what I did was comparable to these modern options - I didn't validate it in more than an ad-hoc way anyway, and it used an off-the-shelf LLM rather than something specially trained. I didn't consider it worth releasing or making a big fuss of.

I contribute my experience here only because I've seen a lot of chatter lately about doing exactly this sort of thing, and I thought I'd share how I made it work for me. There are a lot of ways it can silently fail and give bad numbers if you aren't careful, and I wouldn't want people to think it doesn't work just because they used a vibe coded GitHub project from the last 48 hours that doesn't take these things into account.

A colleague of mine compared Jev with existing classification models, here the results: https://www.timetoact-group.at/en/techblog/techblog/jev-for-...

Reply on news.ycombinator.com