MSN: Sparse Memory for Industrial Search Ranking¶

Core-mechanism reproduction. Product-key memory, sparse top-k reads, and gated backbone fusion execute.

Paper information¶

Field	Value
Paper	arXiv 2602.07526
Affiliations	ByteDance / Douyin Search
First public release	2026-02-07 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`msn`
Local code	`src/auto_research/reproductions/msn/`

Original paper¶

Background and contribution¶

Scale parameters without proportionally increasing ranking computation. Store capacity in two-axis product-key memory, retrieve a few slots per query, and gate with a dense backbone.

flowchart LR
  Q[Query] --> K[Score two key sets] --> S[Top-k slots] --> V[Sparse read] --> F[Gated fusion]
  Q --> B[Dense backbone] --> F

Key figure from the paper¶

Original Figure 3: sparse memory architecture. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ (i,j)\in\operatorname{TopK}(q^\top k_i^{(1)}+q^\top k_j^{(2)}),\quad h'=g(q)m_{ij}+(1-g(q))h. \]

Paper results¶

Douyin Search production reports active days +0.0503%, watch time +0.2958%, and finish rate +0.2071%.

Local reproduction¶

Build two axes from public genres and write interaction transitions as values. Read four slots. Against a dense transition/content ranker, gated top-k PKM gives NDCG@10 0.03540→0.04240 (+19.78%), in metrics/movielens-100k-seed42.json.

auto-research reproduce --paper msn --dataset-dir data --seed 42

Reproduction limits¶

This local public-data ranking comparison is not the production result.