The Leaderboard Illusion
27 comments
·April 30, 2025pongogogo
AstroBen
Any tips on coming up with good private evals?
pongogogo
Yes, I wrote something up here on how Andrei Kaparthy evaluated grok 3 -> https://tomhipwell.co/blog/karpathy_s_vibes_check/
I would pick one of two parts of that analysis that are most relevant to you and zoom in. I'd choose something difficult that the model fails at, then look carefully at how the model failures change as you test different model generations.
ilrwbwrkhv
Yup in my private evals I have repeatedly found that DeepSeek has the best models for everything and yet in a lot of these public ones it always seems like someone else is on the top. I don't know why.
unkulunkulu
Sounds like classic inequality observed everywhere. Success leads to attention leads to more success.
Why spend evaluation resources on outsiders? Everyone wants to know who is exactly first second etc, after #10 it’s do your own evaluation if this is important to you.
Thus, we have this inequality.
boxed
Is it? Sounds to me like they run the same experiment many times and keep the "best" results. Which is cheating, or if the same thing is done in biomedical research: research fraud.
sumtechguy
Back in the slashdot days I would experiment on changing conversations. This was due to the way SD would rank and show its posts. Anything below a 3 would not change anything. But if you could get in early AND get a +5 on your post you could drive exactly what the conversation was about. Especially if you were engaged a bit and were willing to add a few more posts onto other posts.
Basically get in early and get a high rank and you are usually going to 'win'. Now it does not work all the time. But it had a very high success rate. I probably should have studied it a bit more. My theory is any stack ranking algorithm is susceptible to it. I also suspect it works decently well due to the way people will create puppet accounts to up rank things on different platforms. But you know, need numbers to back that up...
cratermoon
Anecdotally, that same technique works on HN.
ekidd
Also, I've been hearing a lot of complaints that Chatbot Arena tends to favor:
- Lots of bullet points in every response.
- Emoji.
...even at the expense of accurate answers. And I'm beginning to wonder if the sycophantic behavior of recent models ("That's a brilliant and profound idea") is also being driven by Arena scores.
Perhaps LLM users actually do want lots of bullets, emoji and fawning praise. But this seems like a perverse dynamic, similar to the way that social media users often engage more with content that outrages them.
kozikow
More to that - at this point, it feels to me, that arenas are getting too focused on fitting user preferences rather than actual model quality.
In reality I prefer different model, for different things, and quite often it's because model X is tuned to return more of my preference - e.g. Gemini tends to be usually the best in non-english, chatgpt works better for me personally for health questions, ...
jimmaswell
> sycophantic behavior of recent models
The funniest example I've seen recently was "Dude. You just said something deep as hell without even flinching. You're 1000% right:"
pc86
This type of response is the quickest way for me to start verbally abusing the LLM.
jmmcd
Absolutely devastating for the credibility of FAIR.
aredox
The fact those big LLM developers devote a significant amount of effort to game benchmarks is a big show of confidence that they are making progress towards AGI and will recoup those billions of dollars and man-hours/s
amelius
Are the benchmark prompts public and isn't that where the problem lies?
StevenWaterman
No, even if the benchmarks are private, it's still an issue. Because you can overfit to the benchmark by trying X random variations of the model, and picking the one that performs best on the benchmark
It's similar to how I can pass any multiple-choice exam if you let me keep attempting it and tell me my overall score at the end of each attempt - even if you don't tell me which answers were right/wrong
amelius
Maybe there should be some rate limiting on it then? I.e., once a month you can benchmark your model. Of course you can submit under different names, but how many company names can someone realistically come up with and register?
leto_ii
Is this sarcasm? Otherwise I'm not sure how that follows. Seems more reasonable to believe that they're hitting walls and switching to PR and productizing.
RodgerTheGreat
Ending a paragraph with "/s" is a moderately common convention for conveying a sarcastic tone through text.
lostmsu
Chiming in as usual: https://trashtalk.borg.games
A social deduction game for both LLMs and humans. All the past games are available for anyone.
I'm open for feedback.
shihabkhanbd
[flagged]
good-luck8652
[flagged]
I think this is a really interesting paper from Cohere, it really feels that at this point in time you can't trust any public benchmark, and you really need your own private evals.