简体中文 · English
PinRec: Outcome-conditioned multi-token generative retrieval¶
Fidelity: full core pipeline. Runs a causal decoder, outcome embeddings, normalized ANN outputs, and multi-token contrastive learning over an unordered future window. MovieLens rating buckets replace Pinterest's cross-surface actions.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2504.10507 |
| Institution | |
| First public release | 2025-04-09 (arXiv v1) |
| Author code | No official or author-released code found in the paper (checked 2026-08-09). |
| Local adapter or method key | pinrec |
| Local reproduction code | src/auto_research/reproductions/pinrec/ |
Original paper summary¶
Background and main changes¶
Conventional next-item prediction tends to continue a user's most common behavior, and strict ordering does not fit feed sessions. PinRec injects the desired outcome into the output head and generates several ANN-compatible item vectors at once for the same future window.
flowchart LR
H["Heterogeneous behavior history"] --> T["Causal Transformer"]
O["Desired outcome"] --> G["Outcome-conditioned heads"]
T --> G
G --> M["Windowed multi-token vectors"]
M --> A["ANN retrieval"]Key figure from the paper¶
Figure 1: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.
Core equations¶
Offline and online results¶
Paper Homefeed unordered Recall@10: UC 0.608, OC 0.625, MT+OC 0.676. At 16 outputs, latency is about 10× lower than OC, recall +16%, and diversity +21.3%. Online MT+OC results are Fulfilled Sessions +0.28%, Time Spent +0.55%, and Homefeed Grid Clicks +3.33%. OC alone reaches up to +4.01% Grid Clicks.
Local reproduction¶
Local comparison: An unconditioned next-token baseline is compared with an outcome-conditioned three-token window. Unordered Recall@10 falls 27.78% and NDCG@10 34.82%. This ablates outcome conditioning and multi-token generation, not a comparison against DIN.
| Model | Unordered Recall@10 | NDCG@10 |
|---|---|---|
| UC next-token | 0.00318 ± 0.00184 | 0.00132 ± 0.00085 |
| OC + 3-token window | 0.00230 ± 0.00061 | 0.00086 ± 0.00034 |
After one protocol correction and an increase from 160 to 200 steps, OC+MT still falls by 27.78% and 34.82% respectively. MovieLens supplies only sparse rating outcomes and two held-out future events, not Pinterest's multi-behavior window supervision. The negative result is not hidden by tuning. See metrics/movielens-100k-seeds42-44.json.