Zenith: Scale ranking with Prime Tokens¶
Core public-data mechanism, not production serving.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2601.21285 |
| Institution | North Carolina State University. |
| First public release | 2026-01-29 (arXiv v1) |
| Author code | No public author code found as of 2026-09-05. |
| Local adapter or method key | zenith |
| Local reproduction code | src/auto_research/reproductions/zenith/ |
Paper method and evidence¶
Token Fusion compresses heterogeneous sparse features into fewer high-dimensional Prime Tokens. RSA/TMHSA maintain token diversity, while TSwiGLU/TSMoE independently enhance tokens for better compute/size scaling.
flowchart LR
E[Heterogeneous sparse features] --> F[Compress to Prime Tokens]
F --> A[Preserve token differences]
A --> B[Token-wise boosting]Original paper figure¶
The original figure shows token fusion and boosting. From the original paper; copyright remains with the authors.
Core equation¶
\[
P=\mathrm{TokenFusion}(E),\qquad P'=P+\mathrm{TokenBoost}(P).
\]
Reported results¶
TikTok Live A/B: Quality Watch Session/User +9.93%, Quality Watch Duration/User +8.11%, CTR AUC +1.05%. Paper evidence only.
Local reproduction¶
Same MovieLens-100K full-catalog protocol: NDCG@10 0.05401→0.04623 (-14.40%). Three-seed measurements, mean, SD and 95% CI. The public proxy does not reproduce private logs, foundation models, bucketing, serving or online lift.