A version 1 arXiv preprint dated 26 Aug 2026 reports a clear split in how two kinds of language disruption appeared after the critical word in human reading and language-model prediction. In an online word-by-word reading experiment, conceptual distortion (CD) was associated with a pronounced but short-lived rise in reading time, while referential distortion (RD) was associated with a smaller increase spread across later words. The model analysis showed a similar contrast in contextual surprisal: CD was more concentrated at the critical point, while RD declined more gradually.
The two conditions used different edits. CD replaced a noun that fit the context with a semantically incongruent alternative. RD changed a pronoun or determiner so that it conflicted with the intended discourse referent. Original versions supplied the undistorted baseline.
The reading-time pattern
The human part included 99 native English speakers recruited through Prolific. Each participant read 21 stories, one version per story, with seven Original, seven CD and seven RD stories in randomized order. Quality control left 1,972 of 2,079 possible participant-story trials: 664 Original, 657 CD and 651 RD, or 94.85% overall.
Neither distortion was associated with a significant reading-time effect at K, the manipulated word. Effects appeared from K+1 onward. The CD increase peaked around K+2 and then declined, while the smaller RD increase was more evenly distributed across later positions and extended to the sentence-final position.
Across K+1 to K+10, the human trajectories remained distinct. CD was best fit by a log-normal function, while RD was best fit by a linear function with an estimated decrease of 0.0038 log-reading-time units per additional word. At sentence boundaries, RD carried an additional reading-time cost of about 3.06%; CD boundary modulation was not significant. The distortion-by-position interaction was significant, with chi-square (9) = 66.58 and p < 0.001.
Inside the models
Researchers applied the same story versions to the pretrained Qwen3-4B-Base model, a 4-billion-parameter system run in evaluation mode without fine-tuning. They tracked contextual surprisal and the distance between output-layer representations. Across K to K+10, CD showed a large surprisal pattern at K followed by a sharp decline. RD started smaller and declined more gradually, with significant differences in the overall trajectory, sentence-boundary comparison and early-to-late change.
The output-layer representation measure showed a related but not identical pattern. It was largest at K in both conditions, and RD had the substantially larger initial displacement. Both trajectories were best fit by power-law decay. Representation distance was lower at sentence boundaries for both conditions, but the difference in boundary modulation was not significant (p = 0.142), nor was the early-to-late difference (p = 0.443). This measure is not equivalent to processing difficulty, so the larger RD displacement should not be read as evidence that RD was harder to process.
A replication in Llama-3.2-3B preserved the main model pattern. Both distortions showed immediate effects at the manipulated word; CD remained more concentrated in surprisal, while RD showed greater relative persistence. Final-layer representation distance followed power-law decay for both conditions. Boundary directions were similar to those in Qwen3-4B, but distortion-type boundary interactions were not significant.
A comparison with limits
The comparison used different linguistic edits: CD changed a noun, while RD changed a pronoun or determiner. Across whole stories, both distorted versions had higher perplexity than Original, but CD and RD did not differ significantly from each other; RD also showed a small lexical-frequency difference. The analysis therefore cannot isolate whether the trajectory split reflects the intended conceptual-versus-referential distinction alone.
Taken together, the human reading-time and model-surprisal results are consistent with a more concentrated downstream profile for CD and a more distributed, persistent profile for RD in these materials. The larger initial RD displacement is a change in an output-layer representation measure, not evidence that RD was harder to process. The study describes distinguishable propagation profiles within this experiment and these model measures.
De-identified reading-time data and preprocessing and analysis scripts are publicly available through the study’s listed Mendeley Data record and GitHub repository.
Paper data and sources
Original title: Distinct dynamics of conceptual and referential disruptions in human reading and large language model processing
Authors: Rui He, Nihal Altay, Wolfram Hinzen
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text