Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Temporal Straightening as a Predictive Mechanism in Human Language Processing

Jiaming Xu¹, Jerry Tang¹, Alexander Huth², Robbe L. T. Goris; ¹University of Texas at Austin, ²The University of Texas at Austin

Presenter: Jiaming Xu

Predicting what comes next is central to how humans process language, and to how artificial language systems, trained on next-word prediction, learn flexible representations that support diverse tasks. Despite the importance of prediction in both systems, the neural mechanisms underlying prediction in the human brain remain poorly understood. Inspired by the temporal straightening hypothesis from vision neuroscience, we investigated predictive representations in language processing from a geometric perspective. This hypothesis proposes that the brain transforms complex inputs to follow straighter temporal trajectories in representational space, enabling prediction through linear extrapolation. Here, we tested whether a similar principle applies to the human language system. Using fMRI data from subjects listening to a natural spoken narrative, we estimated representational timescale as a proxy for trajectory straightness across regions in the language processing hierarchy. We found that timescale increased in higher-order regions, indicating that neural trajectories become progressively straighter along the hierarchy. These findings offer a new perspective on predictive mechanisms in language, suggesting that temporal straightening may serve as a general organizing principle across different systems.

Topic Area: Language & Communication

Extended Abstract: Full Text PDF