BARGE: Tencent's new AI fixes generative recommendation's structural gaps

BARGE (Bridging the structural gap for Autoregressive Recommendation GEneration) tackles a key limitation in generative recommendation models. Prior work encodes items into hierarchical semantic IDs via residual quantization and predicts them token by token, but this flattens multi-token IDs into a single sequence, destroying item-level structure. Furthermore, training and inference inconsistencies over hierarchical codebooks cause semantic drift. BARGE introduces Item Context-Aware Attention (ICA) to restore item-level structure during encoding, and combines Hierarchical Path Reranking (HPR) with Dual-Path Decoding (DPD) to suppress semantic drift from two complementary angles during decoding. The paper demonstrates superior performance on public benchmarks and a large-scale offline test.

BARGE's practical value was confirmed in an online A/B test on a Tencent platform, yielding a 0.60% increase in click-through rate, 1.34% increase in click unique visitors, and 1.70% increase in total reading time. These results show that addressing structural gaps in generative recommendation can lead to measurable business improvements. The paper (arXiv:2607.21028) is 14 pages with 15 figures and was submitted in July 2026. By bridging the structural gap, BARGE offers a more coherent and effective way to apply autoregressive generation to recommendation systems.