BARGE: Repairing item structure and semantic drift in generative recommendation

简体中文

Fidelity: Core mechanism. Trains dual semantic OSQ paths, ICA, HPR, two decoders, and item-space OR fusion on scaled-down public data and serving.

Paper information

Field	Value
Paper link	arXiv 2607.21028
Organization	Tencent
First published	2026-07-23 (arXiv v1)
Original author code	No official or author code was provided in the paper (checked 2026-08-09).
Adapter	barge
Local reproduction code	src/auto_research/reproductions/barge/

Original paper summary

Background and main changes

RQ-VAE expands an item into several Semantic ID tokens, obscuring item boundaries in a standard Transformer. With one quantization path, early beam errors can compound across levels. BARGE pools tokens within each item and writes back a gated ICA context. HPR contrastively reranks cumulative semantic paths at every level. An orthogonal split produces two complementary decoding paths.

flowchart LR
  E["item embedding"] --> O["Householder rotation + orthogonal split"]
  O --> A["channel A residual codes"]
  O --> B["channel B residual codes"]
  H["history dual IDs"] --> I["ICA item pooling + gated fusion"]
  I --> S["shared encoder"]
  S --> DA["decoder A + HPR"]
  S --> DB["decoder B + HPR"]
  DA --> F["item-space OR fusion"]
  DB --> F

Key figure from the paper

Original Figure 2 (overview): The overall design and key components. Image from the original paper; copyright belongs to the authors. Click the image for its source.

Core equation

ICA first obtains item context, then applies a gated writeback:

\[ z^{(i)}=\operatorname{LN}(\operatorname{CrossAttn}(q,X^{(i)},X^{(i)})), \qquad \hat x_l^{(i)}=x_l^{(i)}+\sigma(W_g[x_l^{(i)}\Vert\hat z^{(i)}])\odot\hat z^{(i)}. \]

HPR applies symmetric InfoNCE to cumulative path \(p^{(l)}=\sum_{j=1}^{l}e_{c_j}\) and user context. Inference adds generation log-probability and reranker score. DPD enforces \(R^\top R=I\), splits \(Rz\) into two orthogonal subspaces, and quantizes each separately.

Offline and online results in the paper

The paper reports gains on Amazon Beauty, Sports, Toys, Yelp, and private Tencent offline data. Tencent A/B reports CTR +0.60%, unique clickers +1.34%, and total reading duration +1.70%.

Local reproduction

Control: Single-path RQ-VAE/TIGER with matching width, layers, rows, optimizer, and 140 steps. Adding OSQ, ICA, HPR, and DPD increases NDCG@10 from 0.01147 to 0.02131, or +85.77% relative.

On 180 MovieLens-100K users and 300 items, Hit@10 rises +60.00%, but head share also rises +165.02%, so the accuracy gain has a trade-off. OSQ reconstruction loss decreases from 0.05969 to 0.01947, with orthogonality error 2.98e-7. See metrics/movielens-100k-seed42.json.

auto-research reproduce --paper barge --dataset-dir data --seed 42

Reproduction limits

MovieLens genres substitute for Tencent item embeddings. Full-catalog teacher-forced path scoring substitutes for production beam kernels. Private impression negatives and online serving are absent. OSQ, ICA, HPR, and DPD are actually trained, but local percentages are not online lifts.