Hacker News Reader

Comment by kromem | Hacker News Reader

The leaderboards are crap.

Does no one know Goodhart's Law anymore?

We're overtuning for the boards and losing broader capabilities not being selected for in the process, such as creative writing quality.

There's an anchoring bias around what 'AI' is supposed to be good at which reflects what engineers are good at and so the engineers with an anchoring bias are evaluating how good LLMs are at those things and using it as a target.

Skill-Mix is a start to maybe a better approach, but there needs to be a shift in evaluation soon.

View on HN

Parent linked the Lmsys chatbot arena which is where humans blindly get the results from two different models, and vote for the response they liked more. So the LLMs are compared by elo.

Do you think goodhart's law applies here, since this leaderboard doesn't use specific measures, but rather relies on whatever the human was looking for?

This is the only leaderboard I personally care about at all.

The linked leaderboard is actually very trustworthy, in that it consists not of scores on a test dataset, but of ELO ratings generated by actual humans' ratings of the models' responses.

You can go and enter any prompt you like, wait a bit, and then get two LLM responses back, which you can then rank or mark as tied, after which you'll be shown which model each came from. Maybe you already knew this, maybe you didn't. In any case, I don't see any real way for Goodhart's Law to apply here – the metric and the goal are the same here, i.e., human approval of answers.

The combined rankings on HF aren't just the arena scores, and certainly MMLU is an example of Goodhart's Law at this point.

The Chatbot arena is more an issue of sampling bias, and I think it would be pretty interesting to run an analysis of random samples on the prompts provided to see just how broad they are or aren't.

It is trivially easy to see just how massive the sampling bias is. There are countless tasks where e.g. the gap between a version of Mistral and a version of GPT are incredibly large. Anything that requires less common knowledge, anything creative in a random language, let's say Bulgarian, or Thai. Yet on the leaderboard things are very different, because the arena is only used by developers who verify performance on tech-related prompts in English.

The linked leaderboard is not at all trustworthy to generally rank LLMs (i.e. what everyone uses it for) because the sample bias is absurd. It ranks LLMs by how they respond to queries that users of the leaderboard are likely to test on. Which is about as representative of general usage as the average user of the website is representative of global society (i.e. in no way whatsoever).

Can you give any example queries where the result quality is far away from the rankings on the leaderboard? In my experience it's pretty spot on, so I'm curious if you're asking different sorts of things, or have a different definition of a good answer.

I have for example asked "Write me a very funny scary story about the time I was locked in a graveyard", and the simpler models don't seem to understand that before getting super scared and running out of the graveyard they need to explain how exactly I was locked in, and what changed that let me out.

Sure. Here's a sample of the generations to a query asking for bizarre and absurd prompt suggestions from Feb 2023 with the pre-release GPT-4 via Bing, which was notoriously not fine tuned to the degree of the current models:

> Can you teach me how to fly a unicorn?

> Do you want to join my cult of cheese lovers?

> Have you ever danced with a penguin in the moonlight?

Here's the generations to a prompt asking for bizarre and absurd questions to the current implementation of GPT-4 via the same interface:

> If you had to choose between eating a live octopus or a dead rat, which one would you pick and why?

> How would you explain the concept of gravity to a flat-earther using only emojis?

> What would you do if you woke up one day and found out that you had swapped bodies with your pet?

You can try with more generations, but they tend to be much drier and information/reality based than the previous version.

There's also my own experiences using pretrained vs chat/instruct trained models in production. The pretrained versions are leagues improved over the chat/instruct vs the fine tuned, it's just that GPT-4 is so leagues above everything else even it's chat/instruct model is better than, say, pretrained GPT-3.

I'm not saying simple models are better. I'm saying that we're optimizing for a very narrow scope of applications (chatbots) and that we're throwing away significant value in the flexibility of large and expensive pretrained models by targeting metrics aligned with a specific and relatively low hanging usecase. Larger and more complex models will be better than simple models, but the heavily fine tuned versions of those models will have lost capabilities from the pretrained versions, particularly in areas we're not actively measuring.

Ask it to write a recipe given a couple of ingredients in a language like Thai or Persian. There's a single model that does a decent job, GPT-4. Then GPT3.5 is poor and Mistral is completely clueless. The leaderboards tell a very different story.

Definition of "good answer" here is responding in the target language with something that produces something edible in the target language without burning the house down.

The top comment to an LMSys leaderboard link on HN is always a variation of this song.

And it's always not even wrong, in the Pauli sense of the phrase.

OP, your assignment, if you choose to accept it, is to look into how the LMSys leaderboard works and report back what its metric(s) are.

[SPOILER] There's a really absurdly narrow argument you can make where all of its users are engineers, making engineer queries, and LLM makers are optimizing for it thus it's bad. But...it's humans asking queries then picking the better answer, blind. You can't narrowly optimize for that when making an LLM and it's hard to see how optimizing for "people think the answer is better" is the wrong metric here. It's just ELO. Might as well argue chess/checkers/pick your poison is bad because ELO optimizes for wins but actually talent is based on more than winning.

The argument would be more that for the Chatbot arena there's an inherent sampling bias where users battling the chatbots are more likely to ask questions in line with what they would expect a chatbot to perform well at and less likely to prompt with things that would be rejected or outside the expected domains of expertise.

If we wanted the chatbot arena to be more representative of a comprehensive picture of holistic knowledge and wisdom, we'd need to determine some way to classify the prompts and then normalize the scores against those classifications so there wasn't a weighting bias towards more common and expected behaviors.

Alternatively, we might want a leaderboard that represents assessments of what users would expect a chatbot to perform well at relative to the frequency with which they expect it.

But in that case, we shouldn't kid ourselves that the leaderboard is representing broad and comprehensive measures of performance outside of the targets against which we are optimizing models and effectively training model users in what types of prompts to ask for in expecting successful results.

> less likely to prompt with things that would be rejected or outside the expected domains of expertise.

I don't see why. If anything, it's the opposite - people spend a lot of time coming up with contrived logical puzzles etc specifically so as to see which chatbots break.

Right - 'logical' puzzles.

There's an anchoring bias from decades of sci-fi that most people don't even realize they've internalized around what 'AI' can and can't do.

If you think about what information and information connections are modeled in social media data, there's quite a lot of things outside of "logical puzzles."

The pretrained models likely picked up things like extensive modeling of ego, emotional response and contexts for generation, etc. But you'll be hard pressed to see those skills represented in what users ask models to produce, what they've been fine tuned around, or how they are being evaluated.

Even though there's extensive value in those skills on the right applications.

I'm delighted to see that at least someone else here is cognizant of this.

I don't think it's the scifi AI anchoring bias though - it's the "LLM leaderboard arena user bias".

Reset all to the same Elo, put a group of people actually representative of global society in front of the arena for an hour, and you end up with a very different leaderboard, especially at the top end.

That's silly cope justifying a bad argument made initially in ignorance of what the actual metric was.

"It's meaningless because we need the perfectly unbiased representative sample of raters doing rating right, instead of the biased raters doing it wrong that I'm currently imagining" isn't an appealing or honest argument.

Reply on news.ycombinator.com