When LLM judges agree, should we believe them? | Hacker News Reader

When LLM judges agree, should we believe them?

amazon.science

54 pointsBetelbuddy49 comments

A shared blind spot as noted at the bottom needs to be considered more often. In my day job, most of my coordination with others and now LLMs, is clarifying context and requirements. Claude is very happy to make assertions without the full picture in my experience, even when I give it as much context as I can.

While this is absolutely true - I'd hesitate to discount using similar agents for checking each other. Two agents will almost never hallucinate in the same way, regardless of their weights - and by having a second one (with a different context) check almost entirely eliminates the problem.

It depends what we're judging, doesn't it? If it's "is the formatting in this document compliant with our standards?" I think it's reasonable. If it's like, life-altering if it's wrong I'm less sanguine.

They have already shown algorithmic discrimination in predicting recidivism for brown people, as they are nonsensically overrepresented in the statistical data of US prison populations.

Folks should sue in a class-action lawsuit, any legal firm worth their beautiful walnut desks would seriously be happy take on that constitutionally backed mission. =3

> brown people, as they are nonsensically overrepresented in the statistical data of US prison populations

"Nonsensical?" They commit violent crimes, they get prosecuted for said violent crimes, and are serving prison sentences for those crimes. The algorithm picks up on this trend using the same logic that insurance actuaries use, which has also been largely neutered by critical theory.

What even is the argument here-- they're all innocent? Cops are ignoring piles of dead white people and their white murderers to only go patrol brown neighborhoods? We both know neither claim is true. The usual complaint is that cops avoid their neighborhoods and/or are lazy in investigating the crimes they report. The idea of overpolicing has always been a Marxist double-bind...nonsensical, I daresay.

Oh. Yes. Sure. If there is an error in the training data then they will obviously cite that error.

I'm mostly talking about random coding errors.

> and by having a second one (with a different context) check almost entirely eliminates the problem.

You solved one of the largest problems with current LLMs. How is it possible that nobody tried that before?

Because they do. There are already LLMs checking outputs of other LLMs, the bullshit answers that you see are the results of failures on that checks. If you remove all checks LLMs will create hallucinations even more often.

I'm sorry you're so upset, but it's true. Try it for yourself - get your LLM to hallucinate something, and then paste that text into another chat window and ask it to verify the result for you.

No modern LLM can tell how many eyes the magic card Pit Imp has. They all say 2. This is across all reasoning levels and paid Gemini, Claude, and GPT (Sol)

Drawing a line red to split up the image then has them answer correctly.

Their failure modes are highly correlated.

Yes - those failures (like strawberry) are. And those failures are very rare, which is why you had to reach for the Pit Imp MTG card, which I had to Google to understand your point.

But the failures are universal and correlated

Not really.

Cohen, Hamri, Geva & Globerson, "LM vs LM: Detecting Factual Errors via Cross Examination": Cross-examination "detects over 70% of the incorrect claims while maintaining a high precision of >80%".

> Two agents will almost never hallucinate in the same way, regardless of their weights

Citation needed

Personal experience using agents and seeing this happen frequently.

Try it yourself. Get one to hallucinate, then paste that text into a new window and ask it to verify the facts.

EDIT:

Also - Cohen, Hamri, Geva & Globerson, "LM vs LM: Detecting Factual Errors via Cross Examination": Cross-examination "detects over 70% of the incorrect claims while maintaining a high precision of >80%".

So 70% for ANY error, not just hallucinations.

By design, even the same LLM, when asked the same question multiple times, will almost never hallucinate in the same way. By that (flawed) logic, you could have the same LLM check itself.

Yes - exactly. Because they don't hallucinate in the same way - they are perfect for detecting each other's hallucinations.

But when you ask the SAME LLM (with the same context) they remember the hallucination so it doesn't work. You have to use a fresh one without the same context

You can have the same LLM check itself, especially if you start over with slightly different prompts or context. They sometimes get things right and sometimes don't. Although, if you genuinely have no way to know whether the answer is right or wrong, or the errors are severe, then these checks will not accomplish much. This checking can be better than nothing.

I disagree. When I started having my agents do an independent code review before allowing a commit - the code quality improved dramatically. I don't have numbers... but by and large the errors mostly went away. I went from struggling to get good code to just... getting good code.

Is it perfect? No. But it does what I asked it to do.

I'm confused. How does this disagree with me? You can do the same process with the same model, and that is what I was talking about doing.

Observation: A panel of Judges (multiple Judges), whether it's multiple AI's or not, is fundamentally -- a Jury!

Since the baseline LLM architecture is similar with each other, maybe already there are relations between their each opinion. Of course, this is an assumption and the explicit training seems effective in this case. But, I'm curious whether the approach would be effective in other cases (in terms of generalization?)

They would all agree raspberry has two Rs

But still refuse to answer "How many strings does a bass play with in water?" correctly, perhaps the chat monitors in the third world data entry centers will manually patch the nonsense for a more rational answer someday. lol =3

This was the first time I'd heard that gotcha question. I just threw it at Opus 5:

    None — a bass in water is a fish, and fish are notoriously bad at music.
    
    The instrument version plays four strings as standard (five and six-string basses exist for players who want to go lower or higher), and it prefers to stay dry.

Seems like a pretty good answer to me!

Indeed, giving a definite answer to an ambiguous nonsense question is still incorrect.

A fish can play with as many strings as it finds, but only one when on a hook. Yet this too is an incorrect answer, as it again ignores the ambiguity in the phrasing. =3

How do you know the fish is playing? Is he happy, enjoying it? =3

I agree, part of the ambiguity is also unfairly projecting our own subjective experience onto hapless creatures. =3

This feels like an xkcd 169 situation, to be honest.

It was actually a trivial allusion to a rather old poetic parable, and highlights a foundational flaw in LLM inference model statistical salience.

If a LLM based chat bot does ever answer it correctly, than you know with a fair degree of certainty it was content moderators stepping into the chat. Have a wonderful day. =3

https://en.wikisource.org/wiki/The_Poems_of_John_Godfrey_Sax...

https://en.wikipedia.org/wiki/Blind_men_and_an_elephant

OK, I still have no idea what you actually think the "correct" answer is.

I think they want the correct answer to be "your question doesn't really make sense, so I'm not going to answer it". (But I also think Opus's answer is better.)

All the answers I've seen so far are assuredly not inaccurate (the elephants trunk is like a snake), but never fully correct (an elephant is not a snake).

While the LLM spits out each ambiguous context search result, it never answers the actual query without a human cheaters help. =3

https://en.wikipedia.org/wiki/Pareidolia

I kinda find it funny when I use the advisor on claude code and it agrees with the ideas that the previous model did.

For info: the advisor(s) available are higher end models. For example: you use sonnet, the available advisors are opus and fable. If you use Haiku, the advisor are sonnet, opus and fable.

Kinda weird to generalize "LLM". Every lab, every model is different. Has its own biases, reward functions etc.

I believe it depends on the LLM itself. Like what model as each model has diff weights and diff data trained onn

I would recommend to read the article, it’s actually more nuanced than the title

Without reading the article (doesn't matter if it's pro or contra): no, of course not.

It shouldn't even be a debatable question.

I think you should have read the article first, at minimum the subheader

> Discounting the opinions of LLM judges with highly correlated outputs ensures that panels of judges reflect a true diversity of perspectives.

The quoted sentence still leads to the same answer: no.

Because there's no "discounting of opinions". They are running a separate LLM to "score" opinions. And the result is still "no" regardless of "lineages" or "sources".

And the end of the article leads me to believe that the entire article and approach is LLM-induced garbage:

--- start quote ---

When LLM judges agree, we should ask why. Sometimes agreement is independent evidence. Sometimes it is a shared blind spot. A good aggregation method should be able to tell the difference.

--- end quote ---

This doesn't even make logical sense. Actual judges have highly correlated outputs. This would literally be actively sculpting the range of opinions you want to see.

It's like the idea of political districting that thinks that the aim should be to balance each district between "the two" political parties. You're not doing anything but institutionalizing two political parties and constant conflict. You're setting the range of acceptable opinions, then choosing at random between them. Even more relevantly: when both institutionalized parties have the same opinion, it's considered the correct opinion no matter how much or how little public support it has.

> the aim should be to balance each district between "the two" political parties.

I've never heard anyone explicitly advocating for gerrymandering in favor of conflict/variance before? Is that a real thing?

Great point this will be interesting how this develops.

3 tasks?

Betteridges law of headlines says "NO"

Reply on news.ycombinator.com