In case you are wondering: what is parallelism, and why parallel architectures at all? Check out our recent Topics in Cognitive Science special issue on parallel architectures in language and cognition, co-edited with Neil Cohn and Eva Wittenberg: onlinelibrary.wiley.com/toc/17568765... 6/6
The dual-stream signature is not overall performance: it is a specific sensitivity profile to syntax and semantics. This leaves the autonomy of meaning and grammar as a puzzle better addressed through development, evolution, or learnability than through raw computational capability. 5/6
Still, the two models are not identical. The single-stream model is more sensitive to syntactic anomalies; the dual-stream model is more sensitive to semantic anomalies, to the point of over-detecting them: this kind of ‘semantic hypersensitivity’ hurts its precision on purely syntactic errors. 4/6
Why expect otherwise? One motivation for parallel architectures, where meaning and grammar are autonomous components, is that the split should pay off computationally. Our conclusion is that the autonomy is real, but the explanation is not strictly computational: it has to come from elsewhere. 3/6
We trained single- and dual-stream models on sentences with syntactic/semantic anomalies from two classic ERP studies (Osterhout & Nicol 1999; Kim & Osterhout 2005). Classification accuracy is basically the same, but a bit better than our 2019 non-black-box model. 2/6
psycnet.apa.org/doi/10.1016/...
New paper with Olivier Michalon. We ask whether transformers that separate syntax and semantics into parallel streams classify sentences better than single-stream models. They do not. We think this matters for how the autonomy of meaning and grammar should be explained. 1/6
doi.org/10.1080/0952...
Amazing photo on @Science Advances cover today: 12,000 year old clay bead from Jordan showing the fingerprints of the kid who made it. www.science.org/toc/sciadv/1...
To build an electronic computer, stable states are constructed out of dynamical systems (flip-flop circuits), such that the dynamical nature of the underlying hardware can be entirely ignored: computation occurs as transitions between computational states, entirely shielded from the dynamics of electrons. Such shielding does not exist in brains. Thus, biological cognition cannot be reduced to elementary computations, supposedly implemented by neurons. Rather, computation is an elaborate form of cognition.
Computation is a particular kind of cognitive activity. It does not follow that cognition is entirely made of tiny computations (as cognitivism would make us believe).
press.princeton.edu/books/paperb...
Bots have made their way to Prolific experiments. Our lab has stopped online testing of adults entirely now for this reason - we want to know if what we study is real. Probably data collected 2-3 years ago are ok, but moving forward we just can't know. www.pnas.org/doi/10.1073/...
Neurolinguistics in Sweden (NLS) 2026 — abstracts due 20 February! 📝 Conference at Stockholm University, June 11-12 @stockholm-uni.bsky.social Keynotes: Esti Blanco-Elorrieta, Leonardo Fernandino & me — Get your submission in! 🇸🇪🧠🗣️
www.su.se/english/divi...
The Marica De Vincenzi Foundation is inviting applications for its post-doctoral fellowship! The fellowship offers up to 2 years of post-doctoral support abroad for Italian psycholinguists. Spread the word! Deadline is 3/10 - see info here: nam10.safelinks.protection.outlook.com?url=https%3A...
«Nobody (...) has claimed that DeepMind’s AlphaFold is conscious, even though, under the hood, it is rather similar to an LLM. (...) AlphaFold, which predicts protein structure rather than words, just doesn’t pull our psychological strings in the same way.» @anilseth.bsky.social @noemamag.com
A combination of semantic internalism and role-play fictionalism is a promising framework. That machines are capable of meaning is a necessary fiction, sustained by human semantic cognition—from the generation of training data to the interpretation of machine outputs. 20/20
The extent of ‘role play’ is significant: we pretend simulacra are temporary, atypical members of linguistic communities to evaluate their claims for truth and other norms, and we fill in cognitively for them. 19/20
Referential attributions to LMs depend entirely on human willingness to sustain the fiction of community membership and the continuous cognitive supplementation it requires from human interpreters. 18/20
We propose that simulacra function as atypical members of linguistic communities. Atypical because they rely on humans to do the ‘cognitive work’ behind reference and because of how they are limited by their own bounds of sense and reference. 17/20
LMs lack the mental structures that, in humans, explain referential capacities. But human cognition suffices to explain how LMs’ words are routinely interpreted as having meaning and reference. 16/20
A ‘role play’ perspective cannot endow machine outputs with referential properties: neither the LM nor the simulacra it supports occupy the sorts of deictic spaces that would make ‘us’, ‘here’, etc. pick out referents. 15/20
www.nature.com/articles/s41...
Causal histories cannot ground the reference relation and cannot explain the referential limitations of LMs. Those limitations may only be explained internalistically, by appealing to constraints on LMs’ architectures. 14/20
Key claim: *The bounds of sense and reference are not the same for humans and for (different kinds of) machines* 13/20
First-person uses of these expressions would be meaningless when generated by an LM—which raises the question whether the same expressions in other grammatical persons could then have sense and reference. 12/20
Languages also include non-deictic expressions whose meaning requires a situated speaker with specific bodily and mental characteristics: ‘remember’, ‘walk’, ‘seem’, ‘heavy’, ‘distant’… 11/20
LMs cannot occupy deictic spaces or establish the spatial, temporal, and social coordinates necessary for indexical reference. 10/20
Obstacle 2 🚧 Embodiment. Though human speakers may occasionally fail to establish reference through indexicals (‘here’, ‘now’, ‘us’ etc.), such failures in LMs are architectural and systematic. 9/20
Problem 2: An account in which learners must deploy internal resources to reconstruct meaningful linguistic units from arbitrary parts, and reconnect to usage chains only in virtue of such reconstruction, is no longer an externalist story. 8/20
Problem 1: If causal-historical chains are doing the explanatory work, then disruptions in those chains would undermine the explanation. 7/20
Subword tokenization creates arbitrary subword strings driven by statistical frequency, rather than lexical structure and meaning. This creates two problems for externalist accounts of machine reference. 6/20
Obstacle 1 🚧 Words vs tokens. LMs’ tokens might not correspond to the (parts of) expressions that have causal histories, such as characters, morphemes, or words. 5/20
But the causal-historical apparatus, developed for proper names and natural kind terms, encounters two main obstacles in applications to machines. 4/20
Denying that machine speech and text may be taken to refer, or be about individuals or states of affairs, risks undermining the public enterprise of evaluating LMs’ outputs for pertinence, truth, and other norms. 3/20