Distinct dynamics of conceptual and referential disruptions in human reading and large language model processing
Authors: Rui He, Nihal Altay, Wolfram Hinzen
Organizations: Grammar and Cognition Lab, Department of Translation & Language Sciences, Universitat Pompeu Fabra, Barcelona, Spain · Institut Català de Recerca i Estudis Avançats (ICREA), Barcelona, Spain
Linguistic meaning is grounded in conceptual content, from which reference to particular entities emerges as words enter discourse. To examine the processing dynamics associated with these two dimensions of meaning, we selectively disrupted conceptual or referential information in short narratives and traced the resulting effects in human self-paced reading and in the predictive and representational processing of large language models. In human reading, conceptual disruptions produced a strong but localized processing cost, emerging immediately after the distorted word, reaching an early maximum, and then declining rapidly. Referential disruptions produced weaker effects, which decreased more gradually across subsequent words, and were more strongly modulated by sentence boundaries. In the language model, both disruptions emerged immediately at the manipulated word. Contextual model surprisal showed a pattern closely paralleling human reading: conceptual disruption produced a larger, more locally concentrated effect that decayed rapidly, whereas referential disruption produced a smaller and more gradual downstream effect. Output-layer representations showed a different pattern: referential disruption produced a larger initial displacement, while both distortions were subsequently characterized by power-law decay. Together, these results provide convergent evidence for distinguishable processing dynamics of two types of meaning: conceptual information imposes a more locally concentrated integration cost, whereas referential information engages a more distributed process of maintaining discourse-level identity.
Figures & tables
Figure 1: Comparability of stimulus versions on whole-story and lexical characteristics. Distributions of (A) whole-story perplexity, (B) mean lexical frequency (Zipf scale), and (C) mean word length across Original, conceptually distorted (CD), and referentially distorted (RD) versions of the 21 stories. Points represent individual stories, with grey connecting lines linking the three matched versions of each story. Pairwise differences were evaluated using Siegel–Castellan comparisons following the corresponding Friedman test. Significance labels indicate FDR-adjusted p values: n.s., p≥0.05 ; *, p<0.05 ; **, p<0.01 ; ***, p<0.001 .
Figure 2: Local effects of conceptual and referential distortions in human reading and LLM processing. (A) Estimated effects of conceptual distortion (CD) and referential distortion (RD), relative to the matched Original condition, on residualized log reading time. (B) Changes in contextual surprisal produced by CD and RD in Qwen3-4B-Base, calculated relative to the corresponding Original words. (C) Output-layer representation distance between distorted and Original versions, quantified as cosine distance between the corresponding final-layer hidden states. K denotes the final distorted token; K-3 to K-1 and K+1 to K+3 denote positions relative to K, and K_END denotes the sentence-final word. Points show estimated effects and error bars indicate 95% confidence intervals. The horizontal dashed line indicates zero effect. Asterisks indicate effects that remained significant after FDR correction across analysis positions, applied separately for each distortion type within each outcome (FDR-adjusted p<0.05 ).
Figure 3: Downstream reading-time dynamics following conceptual and referential distortions. (A–B) Distance-specific effects of conceptual distortion (CD) and referential distortion (RD), relative to matched Original observations, on residualized log reading time from K+1 to K+10. Points show mixed-effects-model estimates with 95% confidence intervals, and solid lines show the best-fitting functional forms selected by AICc. CD was best described by a log-normal trajectory, whereas RD was best described by a linear trajectory. The trajectories differed significantly across distance, χ2(9)=66.58,p<0.001 . (C–D) AICc values for the 11 candidate functional forms fitted to the CD and RD trajectories. Lower values indicate better model support, and stars mark the best-supported model. (E) Additional distortion effect associated with reaching the sentence boundary. Boundary modulation differed significantly between CD and RD ( p=0.035 ). Displayed p values above each estimate indicate the within-distortion boundary effects. (F) Model-estimated distortion effects averaged over an early downstream window (K+1 to K+3) and a late window (K+8 to K+10). The distance-window × distortion-type interaction was significant ( p=0.042 ), indicating that the early-to-late change differed between CD and RD. Error bars indicate 95% CI. Dashed lines indicate zero effect.
Figure 4: Downstream contextual surprisal dynamics following conceptual and referential distortions. (A–B) Distance-specific effects of conceptual distortion (CD) and referential distortion (RD), relative to matched Original observations, on contextual surprisal from K to K+10. Points show clustered-OLS estimates with 95% confidence intervals, and solid lines show the best-fitting functional forms selected by AICc. CD was best described by a log-Cauchy trajectory, whereas RD was best described by an exponential trajectory. The trajectories differed significantly across distance, F(10,122)=4.73,p<0.001 . (C–D) AICc values for the 11 candidate functional forms fitted to the CD and RD trajectories. Lower values indicate better model support, and stars mark the best-supported model. (E) Additional surprisal effect associated with reaching the sentence boundary. Boundary modulation differed significantly between CD and RD ( p=0.019 ). Displayed p values above each estimate indicate the within-distortion boundary effects. (F) Model-estimated surprisal effects averaged over an early downstream window (K+1 to K+3) and a late window (K+8–K+10). The distance-window × distortion-type interaction was significant ( p=0.017 ), indicating that the early-to-late change differed between CD and RD. Error bars indicate 95% CI. Dashed lines indicate zero effect.
Figure 5: Downstream representational dynamics following conceptual and referential distortions. (A–B) Distance-specific effects of conceptual distortion (CD) and referential distortion (RD), relative to matched Original representations, on output-layer representation distance from K to K+10. Representation distance was quantified as cosine distance between the corresponding final-layer hidden states. Points show clustered-OLS estimates with 95% confidence intervals, and solid lines show the best-fitting functional forms selected by AICc. Both CD and RD were best described by power-law trajectories. The trajectories differed significantly across distance, F(10,122)=7.70,p<0.001 . (C–D) AICc values for the 11 candidate functional forms fitted to the CD and RD trajectories. Lower values indicate better model support, and stars mark the best-supported model. (E) Additional representation-distance effect associated with reaching the sentence boundary. Boundary modulation did not differ significantly between CD and RD ( p=0.142 ). Displayed p values above each estimate indicate the within-distortion boundary effects. (F) Model-estimated representation distances averaged over an early downstream window (K+1 to K+3) and a late window (K+8 to K+10). The distance-window × distortion-type interaction was not significant ( p=0.443 ), indicating no evidence that the early-to-late change differed between CD and RD. Error bars indicate 95% CI. Dashed lines indicate zero effect.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1: Local effects of surprisal and representational distance with Llama 3.2 3B.
Figure S2: Effect decay of Llama 3.2 surprisal.
Figure S3: Effect decay of Llama 3.2 representational distance.
Department of Linguistics, Georgetown University, USA · Computing and Mathematical Sciences Division, MBZUAI, UAE · Center for Language AI Research, Tohoku University, Japan