Answers are grounded in this paper and cite their sources.
Proposed digital-root frequency schedules improve encoding in neural networks, suggesting enhanced performance in long-context tasks.
Rotary Positional Encoding (RoPE; Su et al. 2021) uses a geometric frequency progression theta_i = 10000^(-2i/d) that, while effective, has no principled structure governing its period or extrapolation behavior. We propose Yare 3-6-9 RoPE, two frequency schedules tuned to digital-root structure: a vortex_369 schedule alternating between two poles, and a digital_root_doubling schedule walking the orbit of 1 under doubling modulo 9 — the cycle (1, 2, 4, 8, 7, 5). Both schedules produce closed-form period structure analyzable without empirical probing. The reference implementation is a non-destructive runtime patch over the HuggingFace RoPE inv_freq buffer, preserving the standard contract and fully revert-able. Unit tests verify norm preservation, idempotency, and determinism. We outline a long-context benchmark suite (LongBench, needle-in-a-haystack at 8k/32k/128k, passkey retrieval) comparing both Yare schedules against vanilla RoPE, NTK-aware scaling, YaRN, and LongRoPE. Empirical results are deferred to v1.1.
No takes yet. Share an insight, caveat, or question.
Weslyn Cory Whitehead (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: