HSTA tracks technology diffusion across 30,000 arXiv preprints and USPTO patents, finding the highest semantic drift in Large Language Models (0.332) while paper-volume velocity fails to Granger-cause frontier compute surges

The paper proposes HSTA, an unsupervised pipeline that turns unstructured scientific text into comparable semantic-trajectory metrics: 384-dimensional sentence embeddings from all-MiniLM-L6-v2 are normalized onto a unit hypersphere, clustered with Spherical K-Means and projected with UMAP, and two metrics are formalized, Semantic Centroid Vector Drift (cosine distance between early and late sub-corpus centroids) and Commercialization Offset (the lag maximizing normalized cross-correlation of quarterly academic and patent document counts). Relative to conventional bibliometric approaches that rely on citation counts or pre-defined taxonomy codes, the pipeline operates directly on raw abstract text, avoiding human labeling bias and the administrative delays of patent granting and classification systems. The method runs on 30,000 document records, with the cluster count k=8 selected by scanning k=4 to 12 using the Mean Silhouette Coefficient and Davies-Bouldin Index (k=8 gives silhouette 0.342 and DB 1.18); robustness checks show the arXiv-versus-USPTO topological separation and the relative ordering of drift metrics persist across choices of k.

Empirically, semantic evolution differs sharply across sub-topics: the Large Language Models cluster shows the highest drift (0.332), Artificial Intelligence Systems and Computer Vision both 0.234, Foundational Model Design 0.174, while mature areas such as Device & Hardware Architecture (0.015) and Statistical Machine Learning (0.012) show minimal drift. The result converts the question of which technical directions are being restructured from a qualitative judgment into a continuously computable quantity, backed by checkable vocabulary evidence: early Large Language Models n-grams center on masked language modeling, BERT fine-tuning and contextual embeddings, while late n-grams shift to in-context learning, instruction tuning, RLHF and prompt engineering. Drift values come from a complete eight-cluster summary table (Table 1) and are qualitatively validated through TF-IDF vocabulary shifts in the Large Language Models cluster; robustness analysis shows language and generative-modeling clusters consistently display high drift while hardware device layers display low drift.

On cross-corpus alignment, all Commercialization Offsets are negative, meaning patent filing density peaks lead scientific preprint index windows within this sample, with lags ranging from -10 quarters for Statistical Machine Learning to -32 quarters for Neural Network Layers and Hardware Architectures. This offers a recomputable, text-density-based alternative measurement to the common intuition that scientific discovery precedes commercial filing, and it indicates that the academic-to-patent timing relationship is not uniform across sub-fields. Offsets are determined by the maximum of a normalized cross-correlation function over quarterly document counts (Equations 6 and 7), with all eight cluster values compiled in Table 2; the authors explicitly scope the result to density alignments 'within this specific streaming sample'.

After running bivariate VAR Granger tests linking quarterly paper volume velocity to Epoch AI frontier training compute (FLOPs), all cluster p-values sit above the 0.05 significance line (from 0.14 for Data Engineering & Processing to 0.81 for Large Language Models), indicating that publication volume velocity alone does not Granger-cause compute allocation surges. This negative result turns the question of whether textual signals can predict hardware capital expenditure into a testable proposition and indicates that text dynamics must be integrated with capital investment and hardware constraint models rather than treating publication volume as a leading indicator of compute investment. The tests use log-differenced quarterly series from 2016 to 2026 and a VAR model (Equations 8-10), with F-statistics and p-values for all eight clusters listed in Table 2; the authors restate in the conclusion that paper volume velocity alone does not predict compute capital allocation spikes.

The framework is aimed at readers who need high-frequency signals of technology diffusion: analysts in capital allocation, infrastructure planning and innovation policy, and science-of-science researchers who want to monitor sub-field evolution without relying on citation counts or taxonomy codes. It applies to a streaming sample spanning 2016-2026, drawing text from arXiv computer science and statistics preprints and USPTO patent abstracts, with Epoch AI frontier-model compute as the physical-constraint proxy; the drift metric can identify which sub-fields are being rapidly restructured, the commercialization offset can compare academic-to-patent timing across sub-fields, and the Granger tests can indicate when textual signals need to be modeled jointly with capital constraints.

The authors scope the negative Commercialization Offsets to density alignments 'within this specific streaming sample', so whether that directionality holds in other corpora or time windows remains an open question. The Granger tests cover only two series, quarterly paper volume velocity and frontier training compute, without including capital investment or hardware constraint variables, so the call to model textual signals jointly with physical capital constraints is currently a methodological proposal rather than a validated joint model. In addition, the 2026 volume expansion is attributed by the authors to indexing updates in open repository snapshots, and the effect of that time boundary on drift and offset estimates is worth continued observation.