PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
May 15, 20260 citationsOpen Access

Yare 3-6-9 Rotary Positional Encoding: A Digital-Root Frequency Schedule for RoPE

View Full Paper
WWWeslyn Cory Whitehead

Key Points

  • The aim is to improve Rotary Positional Encoding by introducing structured frequency schedules based on digital-root principles.
  • Proposed two new frequency schedules: vortex_369 and digital_root_doubling.
  • Implemented a non-destructive runtime patch over the HuggingFace RoPE inv_freq buffer.
  • Conducted a benchmark suite (LongBench) to compare new schedules with existing methods.
  • Identified closed-form period structure for both new schedules without empirical validation.
  • Implemented norm preservation and determinism in the reference implementation.
  • Empirical results to be presented in future work (v1.1).

Abstract

Rotary Positional Encoding (RoPE; Su et al. 2021) uses a geometric frequency progression thetaᵢ = 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 vortex₃69 schedule alternating between two poles, and a digitalᵣootdoubling 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 invfreq 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.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Weslyn Cory Whitehead (2026) studied this question.

synapsesocial.com/papers/6a06b940e7dec685947abd9fhttps://doi.org/10.5281/zenodo.20150178
Ask AI
Helpful
Bookmark
Share
View Full Paper