Join the community of Machine Learners and AI enthusiasts.
Sign Up
I think the emphasis on data quality is well placed. Smaller models seem much less forgiving of noisy or inconsistent training data, so improvements in curation may end up having a larger impact than simply adding more parameters.
That raises an interesting question: do you think the looping behavior is primarily caused by the training data, or do distillation and quantization play a larger role as models get smaller?
I think is the second thing. A small model that enters a strange context may produce a relatively flat distribution. It does not have a strong idea of what should come next but still it has to pick something up. Anti-repetition behavior from data consumes capacity, the less capacity, the more likely your model will suffer from this bias.
That's a simplified view I have on self-reinforcing degeneration mechanism which you can read about here btw: https://arxiv.org/abs/2109.08705
Half of "small reasoning models are overrated" might be an unpinned decoding claim.
The doom loop is a property of a (weights, sampling config) pair, not of the weights. The same checkpoint at repetition penalty 1.0 and at 1.15 are two different systems as far as a leaderboard is concerned, and almost no benchmark records the decoding config next to the score. So the number can be honest and still not reproducible.
On the capacity story, I went and read the paper you linked. The part I keep coming back to is that they identify the mistakes made BEFORE degeneration starts, then go look at hidden states to explain why those mistakes amplify. That pre-onset window is exactly where your data-vs-capacity question lives.
So the cheap version of their analysis, on your models: log next-token entropy over the few tokens right before the loop starts. If entropy is already low at onset, the model was confident and wrong, which points at data. If entropy is high at onset and collapses after the first repeat, that is your flat-distribution story, and the decoder is what locks it in.
Have you logged entropy at loop onset on the ones that degenerate, or is it inferred from the output?
