PIT: Dynamic personalized item tokenization¶
Core public-data mechanism, not private production recommendation.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2602.08530 |
| Institution | Beijing University of Posts and Telecommunications. |
| First public release | 2026-02-09 (arXiv v1) |
| Author code | No public author code found as of 2026-09-05. |
| Local adapter or method key | pit |
| Local reproduction code | src/auto_research/reproductions/pit/ |
Paper method and evidence¶
Jointly train tokenizer and generative recommender, selecting the candidate SID with the smallest user-conditioned loss. A one-to-many beam index preserves alternatives under streaming distribution shifts.
flowchart LR
I[Candidate item SIDs] --> U[User-guided selection]
U --> R[Train generative recommender]
I --> B[Retain alternatives in beam index]Original paper figure¶
The original figure shows jointly trained tokenization and retrieval. From the original paper; copyright remains with the authors.
Core equation¶
\[
z_i^*(u)=\arg\min_{z\in\mathcal B(i)}\mathcal L_{rec}(u,z),\qquad \min_{\theta,\phi}\mathcal L_{token}+\mathcal L_{rec}.
\]
Reported results¶
The paper reports a core online business metric +0.402%; this is not local evidence.
Local reproduction¶
Same MovieLens-100K full catalog: NDCG@10 0.05401→0.05305 (-1.78%). Three-seed metrics, mean, SD and 95% CI. Fixed subset and proxy objective verify the computation only, not private logs, production models, bucketing or serving.