Climber-Pilot: Non-myopic instruction-following retrieval¶
Core public-data retrieval mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | NetEase Cloud Music. |
| First public release | 2026-02-14 (arXiv v1) |
| Author code | No public author code found as of 2026-09-05. |
| Local adapter or method key | climber-pilot |
| Local reproduction code | src/auto_research/reproductions/climber_pilot/ |
Paper method and evidence¶
Time-aware multi-item prediction distills long-horizon consumption into a one-step generator. Condition-guided sparse attention injects category and business constraints without separate filtering or multi-step inference.
flowchart LR
A[Time-aware multi-item targets] --> B[Distill one-step generator]
B[Distill one-step generator] --> C[Apply constrained sparse attention]Original paper figure¶
Figure shows time-aware and condition-guided retrieval. From the original paper; copyright remains with the authors.
Core equation¶
\[
\mathcal L_{TAMIP}=-\sum_{k=1}^{K}w_k\log p(i_{t+k}|H_t),\qquad A'=A\odot M_c.
\]
Reported results¶
NetEase Cloud Music A/B reports a core business metric +4.24%.
Local reproduction¶
MovieLens-1M fixed catalog: NDCG@10 0.05401→0.09251 (+71.29%). Three-seed metrics, mean, SD and 95% CI. No private features, production traffic or serving. These local public-data results do not reproduce online lift.