OneTrans-V2: Shared User Context for Retrieval, Pre-Ranking, and Ranking¶
Concept demonstration; L1 mechanism validation on full MovieLens-1M. All three stages and their losses ran. MovieLens lacks transactions, ad supply, and a real exposure funnel, so proxy metrics are not the paper's business gains.
Paper information¶
| Field | Value |
|---|---|
| Paper | arXiv 2609.28589 |
| Organization | ByteDance |
| First public release | 2026-09-23 (arXiv v1) |
| Authors' code | No official author release found as of 2026-09-27. |
| Adapter | onetrans-v2 |
| Local code | src/auto_research/reproductions/onetrans_v2/ |
Original paper¶
Background and contribution¶
Industrial recommendation retrieves from a large pool, pre-ranks to narrow it, and ranks a shortlist. Separate models repeatedly encode the same history and make fine-ranker feedback harder to transfer upstream. OneTrans-V2 retains each stage's inputs and objectives but shares a causal Transformer user context. History cannot read stage tokens. Each stage token sees only events before the request and not other stage tokens, reusing encoding without future leakage.
flowchart LR
H[Behavior history] --> C[Shared causal Transformer]
C --> R[Retrieval token: decision prefix then three-level SID]
C --> P[Pre-ranking token: lightweight candidate features]
C --> F[Ranking token: richer candidate features]
F -->|Stop-gradient distillation| PKey figure from the paper¶
Original Figure 2: Architecture, modules, and connections from the paper; copyright belongs to the authors.
Mechanisms and equations¶
- Decision-conditioned generative retrieval (DCGR): predict a decision prefix then autoregressively predict three semantic ID levels. Business offsets reorder decisions without overwriting the item distribution conditional on a decision. The paper's decisions include purchases, value, novelty, and ad supply. Section 4.2.1
- Pre-ranking and ranking: lightweight and richer candidate features predict CTR/CVR, respectively. Detach, center, and temperature-scale fine-ranking logits to teach the pre-ranker. Section 4.2.2
With decision prefix \(z\) and SID \(c_{1:3}\), \(P(z,c_{1:3}\mid H)=P(z\mid H)\prod_{\ell=1}^{3}P(c_\ell\mid H,z,c_{<\ell})\). The local run supervises two observable rating/genre proxies, not the paper's complete purchase, value, discovery, and ad-supply decisions.
The scaled joint objective is \(\mathcal L=0.1(\mathcal L_{\rm decision}+\mathcal L_{\rm SID})+\mathcal L_{\rm pre}+\mathcal L_{\rm fine}+\mathcal L_{\rm KD}\). Detached teacher logits are batch-centered, temperature-scaled at \(0.5\), and passed through sigmoid; the student is the pre-ranker. This preserves the loss connection, not the original CTR/CVR labels.
Paper results¶
Section 6.3 reports a user-level 50/50 live A/B with +9.74% GMV/user and 3.2× three-stage end-to-end QPS at the same hardware budget. These come from private commerce systems. Public proxy scores do not measure GMV, ad revenue, a true funnel, or serving-cluster throughput.
Local reproduction¶
There is no independent three-stage cascade baseline. Use full MovieLens-1M chronological histories: the last two events are validation/test, with all training prefixes earlier. Behavior encoding is causal and stage tokens cross-attend only to history. Retrieval fits three-level residual k-means SIDs to item genres, with two proxy decisions and three-level autoregressive cross-entropy. Offsets modify only the two decision logits, not ranking or conditional SID probabilities directly. Pre-ranking reads item IDs; fine ranking adds genres. Both use binary rating ≥3 and ≥4 objectives, with detached ranking-to-pre-ranking distillation. Defaults: seeds 42/43/44, 200 updates each. Evaluate real held-out ratings without sampled negatives; report stage AUC/Brier and decision/SID exactness. Offset testing is a mechanism check, not business impact.
| Three-seed MovieLens-1M mean | Validation | Test |
|---|---|---|
| Pre-ranking AUC, rating ≥3 | 0.55085 | 0.54883 |
| Ranking AUC, rating ≥3 | 0.57562 | 0.57336 |
| Pre-ranking AUC, rating ≥4 | 0.55006 | 0.54108 |
| Ranking AUC, rating ≥4 | 0.56393 | 0.56426 |
| Retrieval three-level SID precision | 0.05456 | 0.05553 |
Results vary across seeds and some are low. Ratings are not clicks/conversions and genre novelty is not full discovery. The concept_demo label does not claim the paper's GMV, QPS, or full DCGR control.
See per-seed metrics and data fingerprints. This path is CPU-only as advertised; there is no A100/A30 receipt.
Reproduction limits¶
Executed shared causal encoding, stage isolation, autoregressive three-level SID, observable decisions, both multitask BCE stages, and ranking-to-pre-ranking distillation. Not executed: real purchases/value/ad decisions, industrial item encoding, sparse MoE, Sequence-Native Training amortization across impressions, serving, or live A/B.