MPFormer: Adaptive Framework for Industrial Multi-Task Personalized Sequential Retriever¶
Core public multi-task retrieval mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Kuaishou Technology. |
| First public release | 2025-08-28 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-09. |
| Local adapter or method key | mpformer |
| Local reproduction code | src/auto_research/reproductions/mpformer/ |
Paper method and evidence¶
Objective tokens drive one sequential retriever across behaviors. Allocate capacity and retrieval quota dynamically by task difficulty.
flowchart LR
A[Objective tokens and history] --> B[Adaptive shared retriever]
B[Adaptive shared retriever] --> C[Allocate capacity and candidates]Original paper figure¶
Figure shows objective-driven retrieval. From the original paper; copyright remains with the authors.
Core equation¶
\[
z_m=F([e_m;h]),\quad\mathcal L=\sum_m q_m(h)\mathcal L_m,\quad\sum_mq_m=1.
\]
Reported results¶
Kuaishou reports watch time +0.426%, training resources -60%, serving resources -66.7%.
Local reproduction¶
MovieLens-100K full catalog, 260 users/420 items, seed 42: shared transition+content baseline versus MPFormer, NDCG@10 0.0354→0.0329 (-7.15%). Fixed metrics. No private logs, traffic, checkpoints or production serving; local results do not reproduce paper lift.