Distinct dynamics of conceptual and referential disruptions in human reading and large language model processing

Temporal constraints on the neural signatures of narrative processing

How the brain constructs meaning across extended contexts remains poorly understood. While neural responses to words and sentences are well characterized, much less is known about the brain mechanisms supporting narrative comprehension. Sentence-level studies suggest that neural activation increases as word meanings are integrated into sentence meaning. At the discourse level, theories propose that narratives depend on situation models, possibly engaging networks beyond core language regions, including the default mode network. Because narrative comprehension unfolds over longer timescales, processing time may be a bottleneck. In this MEG study, we tested how representation size and presentation rate shape neural responses by varying linguistic structure (words, sentences, stories) and the speed of visual text in 1–4-word chunks. We found an early bilateral story effect in visual cortex, followed by a spatiotemporal progression of activity along the temporal lobes that culminated in a three-way contrast among word lists, sentence lists, and stories. Faster presentation altered this pattern: the left-lateralized story effect disappeared, and the right-lateralized effect became more spatially restricted. Under Fast presentation, significant effects were limited to left lateral language cortex distinguishing coherent inputs from word lists, and to two right-hemisphere story effects in extended language regions. We also observed a context effect in the Slow Story condition, with neural responses remaining constant as the narrative unfolded while they increased in the SentenceList and WordList conditions. This effect was absent under Fast presentation, suggesting story-specific comprehension that is temporally constrained. Together, the findings identify temporal constraints as a key determinant of the neural signatures of narrative processing.

Prior knowledge reveals two computational regimes for syntactic processing in the human brain

The human brain rapidly transforms continuous speech into structured, meaningful linguistic representations, yet how prior knowledge constrains this process remains unclear. To characterize this influence, we combined MEG recordings acquired during audiobook listening with corpus-derived transition probabilities over syntactic features defined within both phrase-structure and dependency-based grammars. Across grammatical formalisms, prior knowledge selectively sharpened the neural representation of memory-related features indexing syntactic structures that must remain open for future completion. This enhancement was strictly local: immediately preceding contexts improved neural decoding at both word onset and offset, whereas longer histories produced either a return to baseline at the word level or a deterioration in decoding performance. By contrast, integration-related features indexing the completion of syntactic operations showed no benefit from prior knowledge and were represented most strongly at word offset, consistent with their dependence on word-level structural resolution. These dissociable dynamics reveal two concurrent neural computational regimes for syntactic processing: a forward-looking, locally maintained predictive code for pending structure and an integrative code engaged when structure is resolved. More broadly, our findings impose a mechanistic constraint on neural theories of language processing and on accounts that equate prediction in human language comprehension with the comparatively unconstrained operations of large language models (LLMs).

Construal and variable discourse patterns explain the typology of visual perception verbs

The interplay of prominence-lending factors and language production in children

Several factors influence how adult speakers produce an utterance and which element they choose to start with. For example, German-speaking adults are more likely to begin an utterance with an element if it is (a) in their focus of attention, (b) animate rather than inanimate, and (c) on the left side of a picture rather than the right. However, while such prominence-lending factors affect adult language production, little is known about their effect on child language production. The current study addresses this issue by testing German preschoolers in picture-based language production tasks while measuring gaze patterns via eye-tracking. Specifically, we examined the effect of three factors known to affect adult language production—attentional cueing, animacy, and position in a visual display. To assess how children respond to these manipulations, we zoomed in on two syntactic structures for which children demonstrate varying degrees of proficiency: conjoined noun phrases and transitive sentences. Our results revealed no effects of animacy and position, but significant effects of attention on child language production. Children were more likely to first fixate on a character that had been visually cued and were also more likely to name that character first when producing conjoined noun phrases. Furthermore, initial fixations predicted the order in which children mentioned elements, suggesting a tight link between visual attention and language production. However, visual attention exclusively affected the linearization of nouns in simple conjoined noun phrases but not the selection of starting points in children's production of transitive sentences. These findings show that children's tendency to verbalize what is in their focus of attention is influenced by syntactic complexity. Specifically, even when children's focus of attention was on the patient character, they nonetheless adhered to their preferred active SVO sentence structure. Furthermore, children's sentence planning strategies as reflected by speech-onset-times and the number of disfluencies were likewise immune to prominence-lending factors. Thus, unlike adults in comparable experiments, our findings seem to suggest that children did not show a radically linear incremental word-by-word sentence planning strategy that proceeds from the element in the focus of attention.

The Primacy of Roles over Syntactic Structures: Mental Representation of Chinese Verbs in Argument Realization

Within the framework of argument realization theory, Semantic Role Lists, Participant Roles, and Predicate–Argument Structures represent competing models of verb semantics, each yielding distinct predictions regarding mental representation and processing complexity. This study formalizes these perspectives into two competing accounts: the Role-Quantity Hypothesis (H1), which posits that processing load is driven by the number of event participants, and the Structure-Quantity Hypothesis (H2), which attributes complexity to the multiplicity of syntactic templates. To evaluate these hypotheses, a lexical decision task was conducted on Chinese verbs. The results revealed a significant main effect of role quantity: two-role verbs elicited longer reaction times and lower accuracy than one-role verbs. Conversely, no significant differences were found between one-structure and two-structure verbs. These findings provide robust empirical support for H1, indicating that role-based representations possess greater psychological reality in the Chinese mental lexicon. We argue that for an isolating language like Chinese, verb processing is primarily event-driven, where role information serves as a predictive heuristic during early lexical access. This study offers new insights into the role-based nature of Chinese verb representation, its psychological reality in real-time processing, and the value of integrating argument realization theory with experimental psycholinguistics.

Verbalizable Representations Form a Global Workspace in Language Models

Out of everything the human brain processes, only a small fraction is consciously accessible, in the sense of being available for verbal report, deliberate control, and flexible reasoning. In this paper, we present evidence that an analogous functional distinction has emerged in large language models. Using a new interpretability technique, the Jacobian lens, we identify the representations a model is poised to verbalize at any point in its processing. These representations, which we collectively call the J-space, exhibit the functional properties characteristic of a global workspace: their contents can be reported, deliberately summoned and held, used to carry the intermediate steps of silent reasoning, and passed as arguments to arbitrary downstream computations, while automatic processing such as text parsing and routine inference proceeds without them. The J-space also has structural signatures that global workspace theory associates with conscious access: it carries coherent content only in an intermediate band of layers, holds on the order of tens of concepts at a time, and is broadcast by the model's weights more widely than other representations. These properties make it a practical window into a model's unspoken thinking. In alignment audits, it reveals strategic deliberation, evaluation awareness, and trained-in misaligned dispositions that never appear in the model's outputs. We find that post-training installs the Assistant's point of view in the workspace, and we introduce counterfactual reflection training, which improves behavior by training only what a model would say if interrupted and asked to reflect. These results indicate that language models maintain a small, privileged set of representations bearing some of the functional hallmarks of conscious access, and that decoding these representations sheds light on ongoing cognitive processes.