OxygenREC-v2: Behavior-conditioned generative recommendation¶
Fidelity: core mechanism reproduction. Trains behavior instructions, weighted objectives, privileged future interactions and entropy-routed distillation. No JD production model or private funnel data.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | JD.COM |
| First public release | 2026-07-27 (arXiv v1) |
| Author code | No official author code identified (checked 2026-08-09). |
| Local adapter or method key | oxygenrec-v2 |
| Local reproduction code | src/auto_research/reproductions/oxygenrec_v2/ |
Paper method and evidence¶
A reranker cannot recover candidates the generator missed. Click/cart/order instructions enter the decoder prefix from the first SID token and weight pretraining. Post-training combines verifiable trajectories with privileged future-interaction self-distillation. Low-entropy positions get dense teacher supervision; high-entropy positions get forward KL to retain uncertainty. Inference never consumes future interactions.
flowchart LR
A[Behavior instruction and history] --> B[Weighted SID generation]
B[Weighted SID generation] --> C[Future-interaction teacher]
C[Future-interaction teacher] --> D[Entropy-routed distillation]Original paper figure¶
The paper's architecture shows instruction-controlled generation and post-training. From the original paper; copyright remains with the authors.
Core equation¶
Reported results¶
Matched backbone HR@512 goes 43.24%→44.14%, Recall@512 34.95%→36.39%. In six JD production settings at 5%–20% traffic for 5–8 days, feed/floor UCTCVR rises 1.61%–4.44%, homepage GMV +21.21% and unified product-detail GMV +4.25%.
Local reproduction¶
MovieLens-1M subset: 240 users, 400 items, matched full-catalog candidates, steps and seed. Behavior-weighted generation uses 20% high-entropy routing. Popularity terciles stand in for missing behavior labels; inference marginalizes over all three behaviors to avoid leaking test targets. PT-only Hit@10/NDCG@10 is 0.0625/0.02861; treatment 0.0250/0.01313 (NDCG −54.09%). Popularity is not a reliable buying-intent proxy; this is a negative local transfer, not a rebuttal of private-data gains. Stable metrics · Cross-domain summary. No private multibehavior logs, production three-level SIDs, 3B-A1B MoE or serving.