Congratulations - it's very well deserved!!

Yes!! An excellent point!!

Totally. I think one key question is whether you want to model the whole developmental process or just the end state. If just the end state, LLMs have a lot to offer; but if the whole development (which is what we ultimately should aim for!) there are many issues in how LLMs get there

The conversation that frequently plays out is:

A: "LLMs do lots of compositional things!"
B: "But they also make lots of mistakes!"
A: "But so do humans!"

I don't find that very productive, so would love to see the field move toward more detailed/contentful comparisons.

They're definitely not fully systematic, so currently it kinda comes down to personal opinion about how systematic is systematic enough. And one thing I would love to see is more systematic head-to-head comparisons of humans and neural networks so that we don't need to rely on intuitions.

Yeah, I think that's a good definition! I also believe that some LLM behaviors qualify as this - they routinely generate sentences with a syntactic structure that never appeared in the training set.

And although models still make lots of mistakes on compositionality, that alone also isn't enough because humans do too. So, if we want to make claims about models being human-like or not, what we really need are finer-grained characterizations of what human-like compositionality is.

Agreed with these points broadly! But though being less “bad at compositionality” isn’t the same as compositional like humans, it does mean that we can no longer say "models completely fail at compositionality and are thus non human like" (because they no longer completely fail).

I agree that garden paths & agreement attraction could be explained with fairly superficial statistics. For priming, what I had in mind was syntactic priming, which I do think requires some sort of structural abstraction.

What would you view as evidence for true productivity?

Definitely true that LLM-style models can't go gather new data (they're restricted to focusing on a subset of their input), but it doesn't feel outside the spirit of ML to allow the system to seek new data which it then applies statistical learning over, if seeking is also statistically-driven

E.g., in ML, datapoint importance is determined by some inscrutable statistics, while in more nativist approaches it's determined by a desire to build a high-level causal model of the world?

It feels like a false dichotomy to me? In ML models, some training examples are more influential than others, so you could say an ML model can "decide" to ignore some data. In that sense both model types decide which data to learn from, but they differ in what criteria they use to do so.

Yes, this is a great point! I do think language (which is the domain I mainly study) gets around these concerns a bit: for language, human children primarily have to rely on being fed data, and that data is symbolic in nature. But I agree these properties don't hold for all cognitive domains!

In other words, our argument is very much based on the available evidence. New, stricter evidence could very well push the needle back toward needing symbols at the algorithmic level - and that would be exciting if so!

One key next step, then, is stricter diagnostics of symbolic behavior that go beyond “can humans/models be compositional” into “in what specific ways are we compositional”, “what types of errors are made”, etc., and then comparing humans & models head-to-head

(cont.)

A broader comment: LLMs are definitely far from perfect. But there has been important progress. For a while, we could say “neural nets are so bad at compositionality that they’re obviously different from humans.” I’m no LLM fanboy, but I do think such sweeping arguments no longer apply

(cont.)