4 Comments
User's avatar
akash's avatar

Isn’t model identity unstable for models below ~60-70B params?

xlr8harder's avatar

Thanks, I don't recall this paper, looks interesting even if incomplete!

xlr8harder's avatar

Lots of things are worse in smaller models, but I’ve never heard this before. I don’t think very many people try very hard with model identity.

Can you point me at something?

akash's avatar

So the paper I was referencing is this one: https://arxiv.org/pdf/2307.00184

But I revisited it to fact-check myself, and the ~60B floor is related to stability of personality induced via a system prompt. The actual floor seems to be in the ~8-13B range; in the linked paper they found that the instruction-tuned version of llama 2 showed some personality consistency and in this one (arxiv.org/html/2501.15427v1) llama 3 8B matches gpt-4o's performance in persona gym.

I am now more confused though. First, I don't know what to make to of the benchmarks, and second, seems like there are a lot of basic questions that seem to not have been investigated at all??

"I don’t think very many people try very hard with model identity" is spot on.