Fair points, especially on GSM8K saturation and Qwen possibly already sitting close to the solution. That said, even if this is mostly "last-mile alignment", the fact that it can be done with such a tiny signal is still interesting, it suggests the gap between capability and behavior might be much smaller (and cheaper to bridge) than we assume.

> the gap between capability and behavior might be much smaller

Can you elaborate a bit on what you mean with the gap?

Yeah, my big problem with the paper is it just might be an artifact of qwen's training process.

In all fairness most of the unique stuff I can do is probably an artifact of my training process, so it seems unfair to deny an LLM the same accomodation.

How much did your training cost society?

This got me thinking, and it might actually even be a comparable amount. Let's estimate 12 years of schooling run at minimum $100,000 per student, at least in the US [1], and then add onto that number whatever else you may do after that, i.e. a bunch more money if paid (college) or "unpaid" (self-taught skills and improvements) education, and then the likely biggest portion for white-collar workers, yet hard-to-quantify, in experience and "value" professional work will equip one with.

Now divide the average SOTA LLM's training cost (or a guess, since these numbers aren't always published as far as I'm aware) by the number of users, or if you wanted to be more strict, the number of people it's proven to be useful for (what else would training be for), and it might not be so far off anymore?

Of course, whether it makes sense to divide and spread out the LLMs' costs across users in order to calculate an "average utility" is debatable.

[1] https://www.publicschoolreview.com/average-spending-student-...

They're using the truncated SVD, not the full variant, that's computationally cheaper.

I've done a lot of exploratory work with Stable Diffusion LoRAs, and I actually do buy that there's some juice here, though it's almost certainly not nearly as good as other techniques can be. In particular, this technique will likely avoid the intruder dimension problem which plagues naive LoRA. SVD is expensive, but you only have to do it once at the beginning of training.

I haven't done much research lately, but when I was working on it, I was having substantial success training an adapter of the form U_k @ P @ A, where U_k was the top k left singular vectors of the underlying weight, and then P and A were your typical LoRA projection matrices.

The 13 parameters are kind of misleading here; the real juice is going to be in the P_i fixed random matrices. My suspicion is that they are overfitting to the benchmark, but they almost certainly are observing a real gain in model capacity that is largely due to avoiding the intruder dimension problem.