Persistent Hodge Laplacian Analysis of Anaphoric Structure in Large Language Model Representations
IEEE Access · 2026 · 14 · 112698-112713
DOI: 10.1109/access.2026.3713829 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Vladimir A. Lazovskiy und 3 weiteren Autor:innen. Die Arbeit nennt 55 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Federal Assignment of the Ministry of Science and Higher Education of the Russian Federation,.
Abstract (Originalsprache)
We introduce persistent Hodge eigenchains as a framework for probing relational structure in language model representations. By decomposing the simplicial complex built over token-pair cosine similarities via the Hodge Laplacian, the method separates representational geometry into curl, gradient, and harmonic components and tracks their persistence across a filtration—providing a richer geometric lens than linear probes or global topological summaries. Applying the framework to anaphora resolution across five controlled datasets and two architecturally distinct models (Mistral-7B, RWKV-6), we find that anaphoric pairs produce longer-lived eigenchains than non-anaphoric pairs in a majority of spectral components and layers for grammatical anaphora datasets (D1–D4), with the harmonic component consistently strongest for structurally complex sentences. Grammatical disambiguation (D4) is detected with high sensitivity (up to 28/32 BH-significant layers for Mistral-7B), while semantic disambiguation (D5) yields a clean null result: neither eigenchain persistence nor eigenvalue position distinguishes the correct from the distractor antecedent, establishing a principled boundary for the method. A cosine-similarity baseline shows that the topological analysis adds discriminative value over raw pairwise distances specifically for grammatical disambiguation (Mistral-7B: Hodge 28/32 vs. cosine 19/32 BH-significant layers), while largely recapitulating the cosine signal for simple anaphora. Results vary across models and spectral components; not all datasets and components yield significant effects, and the findings should be interpreted accordingly.
Förderung
- Federal Assignment of the Ministry of Science and Higher Education of the Russian Federation,
Zitieren
APA
Lazovskiy, V. A., Khismatullin, R. K., Kazantsev, V. B., & Stasenko, S. V. (2026). Persistent Hodge Laplacian Analysis of Anaphoric Structure in Large Language Model Representations. IEEE Access, 14, 112698-112713. https://doi.org/10.1109/access.2026.3713829
Vancouver
Lazovskiy VA, Khismatullin RK, Kazantsev VB, Stasenko SV. Persistent Hodge Laplacian Analysis of Anaphoric Structure in Large Language Model Representations. IEEE Access. 2026;14:112698-112713. doi:10.1109/access.2026.3713829
BibTeX
@article{Lazovskiy2026,
author = {Lazovskiy, Vladimir A. and Khismatullin, Roman K. and Kazantsev, Victor B. and Stasenko, Sergey V.},
title = {Persistent Hodge Laplacian Analysis of Anaphoric Structure in Large Language Model Representations},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {112698--112713},
doi = {10.1109/access.2026.3713829}
}
EUVIMED