Questions for Damián Blasi' Talk on "LLM as theories of human language: the view from linguistic diversity"

[xqrue-49]

Questions for Damián Blasi' Talk on "LLM as theories of human language: the view from linguistic diversity"

Abstract
The prowess of LLMs to acquire complex linguistic representations has inspired a new wave of LLM-centric theories of human language. In this view, the inductive biases LLMs bring to play are not a mere artifact of the model’s architecture but a substantive claim about the very nature of possible and probable languages, such that those characteristics that are easier to LLM-model are also more represented across the world’s languages. In this presentation, I will show that - in spite of the intuitive nature of the claim - the role of functional forces in shaping linguistic diversity is, at best, modest.

[dario-marino]

Is it possible to observe the main differences of language theory when trained with a different subset (or only one) languages/language?

[jstanleyi]

If functional pressures play only a modest role in shaping linguistic diversity, what alternative mechanisms do you think primarily account for the patterns we observe across the world’s languages?

[BaileyMeche]

How do your cross-linguistic findings on the limited role of functional pressures in shaping linguistic diversity challenge the claim that LLM inductive biases reflect universal constraints on possible human languages?

[mikezheng2001]

If functional forces play only a modest role in shaping linguistic diversity, what do you consider to be the primary drivers instead? And as a follow-up, are there specific linguistic features across the world's languages that current LLMs fundamentally struggle to capture because they lack these non-functional mechanisms?

[ralphluo11]

You conclude that functional pressures in shaping linguistic diversity are “at best, modest.” What empirical signal drives that—small effect sizes overall, or effects that largely disappear once you control for genealogy, geography, and contact?

[ZheruiLU]

How do you define and measure “functional forces” in a way that cleanly distinguishes them from genealogy, areal contact, and sociolinguistic context, and what is the key empirical pattern that makes you conclude their role is modest once those factors are accounted for?

[Charlottefox]

Looking forward to your talk.

I agree that theories of human language must re-center socio-historical contingencies, such as indexicality and contiguity. But given that LLMs can adapt to user characteristics and prompt history (the in-context learning), how would you argue the fundamental gap between this co-text and human interaction in the actual context? Thank you.

[amritapathak1]

If functional pressures only modestly explain linguistic diversity, what other forces do you think play the largest role in shaping cross linguistic patterns, and do current LLMs capture any of those forces?

[bkehoechicago]

You argue that functional forces play only a modest role in shaping linguistic diversity. How do you distinguish the influence of those functional pressures from other factors when evaluating why certain patterns are widespread across languages?

[ben-knell]

If functional forces only have a minor effect on language, are influences cultural or something else then in your view? How can researchers distinguish what types of forces have historically affected languages?

[larissaxcn]

If LLM learnability does not strongly predict cross-linguistic distribution, what kinds of forces do you think play the larger role in shaping linguistic diversity?

[dorisduowang]

If the functional forces that make some patterns easier for LLMs are only modestly related to linguistic diversity, what other forces do you think most strongly explain why certain language features are widespread across the world’s languages?

[ruining-he]

How does the linguistic diversity influenced by LLMs interact with other kinds of diversity: cultural homogeneity/diversity, scientific homogeneity/innovation?

[carrieeehuang]

You argue that functional forces shaping linguistic diversity are modest. How do you disentangle “functional pressures are weak” from “the modeling setup under-measures them,” for example because of the training data mix, tokenization, or the dominance of a few written standard languages? Put differently, could functional forces show up more strongly in spoken interaction, morphology-heavy languages, or low-resource settings that current LLM pipelines represent poorly?

[yilinx-10]

How do you think human–LLM collaboration in text generation might influence the future evolution of language?