WHALE: A unified Wukong–HSTU recommender¶
Actual Wukong, HSTU, and per-layer exchange; no production kernels or private features.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Meta Platforms, Inc. |
| First public release | 2026-07-19 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-09. |
| Local adapter or method key | whale |
| Local reproduction code | src/auto_research/reproductions/whale/ |
Paper method and evidence¶
A Wukong branch recursively models higher-order feature interactions and an HSTU branch models behavior sequences. Bidirectional cross-attention exchanges states each layer rather than scaling separate towers independently.
flowchart LR
A[Update Wukong and HSTU separately] --> B[Exchange states by cross-attention]
B[Exchange states by cross-attention] --> C[Repeat each layer]Original paper figure¶
The original figure shows progressive branch exchange. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
Meta 14-day A/B: primary metric +0.113%, Metric 1 +0.824%, Metric 2 +1.820%; inference QPS about 5% lower, reflecting a serving tradeoff.
Local reproduction¶
Train multiplicative Wukong interactions, gated causal HSTU and layerwise cross-attention. Match late-fusion control width/depth: seed-42 NDCG@10 0.00429→0.00072 (-83.20%). Metrics. No private ultra-sparse features, original scale or Triton/AOTInductor kernels. The local decline is retained.