Wander Around Preprints · home

Wander Around Preprints

Open preprints from the Prometheus7 Institute on the substrate-augmented language-modeling research programme. Updated 2026-05-06.

arXiv:wander.2605.0004 [cs.CL · cs.LG]

Wander Around: A Walkable Cartography of the Long-Tail Web

psiloceyeben · Prometheus7 Institute

Cartography platform for ~7M long-tail websites with a substrate-augmented spatial AI.

arXiv:wander.2605.0002 [cs.CL · cs.LG · stat.ML]

Empirical Demonstration of Compositional-Substrate Data Efficiency in a 27-Million-Parameter Language Model

psiloceyeben · Prometheus7 Institute

13× to 18× perplexity reduction on cross-corpus evaluation; HRR-initialized 27M-param model matches behavior expected of substantially larger randomly-initialized baselines.

arXiv:wander.2605.0001 [cs.CL · cs.LG · cs.NE]

Compositional Substrate as a Data-Efficiency Substitute for Parameter Scale: A Holographic Reduced Representation Approach to Small Language Models

psiloceyeben · Prometheus7 Institute

Algebraic binding via Plate-style HRR replaces gradient-discovered compositional structure, yielding 3–5 orders of magnitude data efficiency.

arXiv:wander.2605.0003 [cs.LG · cs.AI · cs.CL]

Frame-Relative Parameter Accounting for Substrate-Augmented Models

psiloceyeben · Prometheus7 Institute

"Parameter count" is a frame-dependent quantity. Three accounting frames (gradient, substrate, composition) yield three different counts, all valid; reporting only one obscures the others.