Elliot Murphy

Vittoria Dentella, Fritz Guenther, Elliot Murphy et al.

Large Language Models (LLMs) are recruited in applications that span from clinical assistance and legal support to question answering and education. Their success in specialized tasks has led to the claim that they possess human-like linguistic capabilities related to compositional understanding and reasoning. Yet, reverse-engineering is bound by Moravec's Paradox, according to which easy skills are hard. We systematically assess 7 state-of-the-art models on a novel benchmark. Models answered a series of comprehension questions, each prompted multiple times in two settings, permitting one-word or open-length replies. Each question targets a short text featuring high-frequency linguistic constructions. To establish a baseline for achieving human-like performance, we tested 400 humans on the same prompts. Based on a dataset of n=26,680 datapoints, we discovered that LLMs perform at chance accuracy and waver considerably in their answers. Quantitatively, the tested models are outperformed by humans, and qualitatively their answers showcase distinctly non-human errors in language understanding. We interpret this evidence as suggesting that, despite their usefulness in various tasks, current AI models fall short of understanding language in a way that matches humans, and we argue that this may be due to their lack of a compositional operator for regulating grammatical and semantic information.

Evelina Leivada, Elliot Murphy, Gary Marcus

Machine intelligence is increasingly being linked to claims about sentience, language processing, and an ability to comprehend and transform natural language into a range of stimuli. We systematically analyze the ability of DALL-E 2 to capture 8 grammatical phenomena pertaining to compositionality that are widely discussed in linguistics and pervasive in human language: binding principles and coreference, passives, word order, coordination, comparatives, negation, ellipsis, and structural ambiguity. Whereas young children routinely master these phenomena, learning systematic mappings between syntax and semantics, DALL-E 2 is unable to reliably infer meanings that are consistent with the syntax. These results challenge recent claims concerning the capacity of such systems to understand of human language. We make available the full set of test materials as a benchmark for future testing.

Elliot Murphy, Emma Holmes, Karl Friston

Natural language syntax yields an unbounded array of hierarchically structured expressions. We claim that these are used in the service of active inference in accord with the free-energy principle (FEP). While conceptual advances alongside modelling and simulation work have attempted to connect speech segmentation and linguistic communication with the FEP, we extend this program to the underlying computations responsible for generating syntactic objects. We argue that recently proposed principles of economy in language design - such as "minimal search" criteria from theoretical syntax - adhere to the FEP. This affords a greater degree of explanatory power to the FEP - with respect to higher language functions - and offers linguistics a grounding in first principles with respect to computability. We show how both tree-geometric depth and a Kolmogorov complexity estimate (recruiting a Lempel-Ziv compression algorithm) can be used to accurately predict legal operations on syntactic workspaces, directly in line with formulations of variational free energy minimization. This is used to motivate a general principle of language design that we term Turing-Chomsky Compression (TCC). We use TCC to align concerns of linguists with the normative account of self-organization furnished by the FEP, by marshalling evidence from theoretical linguistics and psycholinguistics to ground core principles of efficient syntactic computation within active inference.

Evelina Leivada, Gary Marcus, Fritz Günther et al.

Modern Artificial Intelligence applications show great potential for language-related tasks that rely on next-word prediction. The current generation of Large Language Models (LLMs) have been linked to claims about human-like linguistic performance and their applications are hailed both as a step towards artificial general intelligence and as a major advance in understanding the cognitive, and even neural basis of human language. To assess these claims, first we analyze the contribution of LLMs as theoretically informative representations of a target cognitive system vs. atheoretical mechanistic tools. Second, we evaluate the models' ability to see the bigger picture, through top-down feedback from higher levels of processing, which requires grounding in previous expectations and past world experience. We hypothesize that since models lack grounded cognition, they cannot take advantage of these features and instead solely rely on fixed associations between represented words and word vectors. To assess this, we designed and ran a novel 'leet task' (l33t t4sk), which requires decoding sentences in which letters are systematically replaced by numbers. The results suggest that humans excel in this task whereas models struggle, confirming our hypothesis. We interpret the results by identifying the key abilities that are still missing from the current state of development of these models, which require solutions that go beyond increased system scaling.

Evelina Leivada, Vittoria Dentella, Elliot Murphy

In the field of Artificial (General) Intelligence (AI), the several recent advancements in Natural language processing (NLP) activities relying on Large Language Models (LLMs) have come to encourage the adoption of LLMs as scientific models of language. While the terminology employed for the characterization of LLMs favors their embracing as such, it is not clear that they are in a place to offer insights into the target system they seek to represent. After identifying the most important theoretical and empirical risks brought about by the adoption of scientific models that lack transparency, we discuss LLMs relating them to every scientific model's fundamental components: the object, the medium, the meaning and the user. We conclude that, at their current stage of development, LLMs hardly offer any explanations for language, and then we provide an outlook for more informative future research directions on this topic.

Elliot Murphy

A comprehensive model of natural language processing in the brain must accommodate four components: representations, operations, structures and encoding. It further requires a principled account of how these components mechanistically, and causally, relate to each another. While previous models have isolated regions of interest for structure-building and lexical access, many gaps remain with respect to bridging distinct scales of neural complexity. By expanding existing accounts of how neural oscillations can index various linguistic processes, this article proposes a neurocomputational architecture for syntax, termed the ROSE model (Representation, Operation, Structure, Encoding). Under ROSE, the basic data structures of syntax are atomic features, types of mental representations (R), and are coded at the single-unit and ensemble level. Elementary computations (O) that transform these units into manipulable objects accessible to subsequent structure-building levels are coded via high frequency gamma activity. Low frequency synchronization and cross-frequency coupling code for recursive categorial inferences (S). Distinct forms of low frequency coupling and phase-amplitude coupling (delta-theta coupling via pSTS-IFG; theta-gamma coupling via IFG to conceptual hubs) then encode these structures onto distinct workspaces (E). Causally connecting R to O is spike-phase/LFP coupling; connecting O to S is phase-amplitude coupling; connecting S to E is a system of frontotemporal traveling oscillations; connecting E to lower levels is low-frequency phase resetting of spike-LFP coupling. ROSE is reliant on neurophysiologically plausible mechanisms, is supported at all four levels by a range of recent empirical research, and provides an anatomically precise and falsifiable grounding for the basic property of natural language syntax: hierarchical, recursive structure-building.

Elliot Murphy

Natural language syntax can serve as a major test for how to integrate two infamously distinct frameworks: symbolic representations and connectionist neural networks. Building on a recent neurocomputational architecture for syntax (ROSE), I discuss the prospects of reconciling the neural code for hierarchical 'vertical' syntax with linear and predictive 'horizontal' processes via a hybrid neurosymbolic model. I argue that the former can be accounted for via the higher levels of ROSE in terms of vertical phrase structure representations, while the latter can explain horizontal forms of linguistic information via the tuning of the lower levels to statistical and perceptual inferences. One prediction of this is that artificial language models will contribute to the cognitive neuroscience of horizontal morphosyntax, but much less so to hierarchically compositional structures. I claim that this perspective helps resolve many current tensions in the literature. Options for integrating these two neural codes are discussed, with particular emphasis on how predictive coding mechanisms can serve as interfaces between symbolic oscillatory phase codes and population codes for the statistics of linearized aspects of syntax. Lastly, I provide a neurosymbolic mathematical model for how to inject symbolic representations into a neural regime encoding lexico-semantic statistical features.

Elliot Murphy, Jill de Villiers, Sofia Lucero Morales

In this study we compared how well DALL-E 2 visually represented the meaning of linguistic prompts also given to young children in comprehension tests. Sentences representing fundamental components of grammatical knowledge were selected from assessment tests used with several hundred English-speaking children aged 2-7 years for whom we had collected original item-level data. DALL-E 2 was given these prompts five times to generate 20 cartoons per item, for 9 adult judges to score. Results revealed no conditions in which DALL-E 2-generated images that matched the semantic accuracy of children, even at the youngest age (2 years). DALL-E 2 failed to assign the appropriate roles in reversible forms; it failed on negation despite an easier contrastive prompt than the children received; it often assigned the adjective to the wrong noun; it ignored implicit agents in passives. This work points to a clear absence of compositional sentence representations for DALL-E 2.

Elliot Murphy, Evelina Leivada, Vittoria Dentella et al.

A core component of a successful artificial general intelligence would be the rapid creation and manipulation of grounded compositional abstractions and the demonstration of expertise in the family of recursive hierarchical syntactic objects necessary for the creative use of human language. We evaluated the recently released o3 model (OpenAI; o3-mini-high) and discovered that while it succeeds on some basic linguistic tests relying on linear, surface statistics (e.g., the Strawberry Test), it fails to generalize basic phrase structure rules; it fails with comparative sentences involving semantically illegal cardinality comparisons ('Escher sentences'); its fails to correctly rate and explain acceptability dynamics; and it fails to distinguish between instructions to generate unacceptable semantic vs. unacceptable syntactic outputs. When tasked with generating simple violations of grammatical rules, it is seemingly incapable of representing multiple parses to evaluate against various possible semantic interpretations. In stark contrast to many recent claims that artificial language models are on the verge of replacing the field of linguistics, our results suggest not only that deep learning is hitting a wall with respect to compositionality (Marcus 2022), but that it is hitting [a [stubbornly [resilient wall]]] that cannot readily be surmounted to reach human-like compositional reasoning simply through more compute.

Elliot Murphy

In order to design strong paradigms for isolating lexical access and semantics, we need to know what a word is. Surprisingly few linguists and philosophers have a clear model of what a word is, even though words impact basically every aspect of human life. Researchers that regularly publish academic papers about language often rely on outdated, or inaccurate, assumptions about wordhood. This short pedagogical document outlines what the lexicon is most certainly not (though is often mistakenly taken to be), what it might be (based on current good theories), and what some implications for experimental design are.