Reformulating Transformers for LLMs: A Riemann Sphere Framework with Empirical Validation

Description

Complete mathematical framework and benchmarks for the PsiQRH Riemann Sphere language model. Hidden states are mapped to the Riemann sphere via stereographic projection, rotated by Mobius transformations in SL(2,C) (the double cover of the Lorentz group), and projected back. 96 trainable parameters (0.0003% of 32.1M total). Benchmarks across 5 domains (code, poetry, math, dialogue, science), GSM8K reasoning, latency scaling, and GPU memory, compared against GPT-2 (124M) and random baseline. Hardware: RTX 5060 Ti 15.5GB

Authors

DOI: 10.5281/zenodo.20753637

Publication Date: 2026-06-19

Back to publications list


About