Bridging the Structural Gap: Adapting Autoregressive Generation for Recommendation

Generative Recommendation (GR) models typically represent items as hierarchical semantic IDs (token sequences) and predict them autoregressively. However, this approach suffers from two structural gaps: the loss of item-level boundaries when flattening tokens into a sequence, and the tendency for decoding to drift into incorrect subtrees of the hierarchical codebook (semantic drift). The authors aim to bridge these gaps to improve recommendation accuracy.

The authors propose BARGE, which introduces three lightweight modules:

• Item Context-Aware Attention (ICA): This module aggregates token embeddings into an item-level context and fuses it back into each token via a gated residual mechanism, ensuring the encoder retains item-level structural information.
• Hierarchical Path Reranking (HPR): This module augments beam search by scoring the semantic compatibility between the decoder's initial hidden state and the cumulative path embedding at each layer, using a dual-tower contrastive objective to suppress intra-path drift.
• Dual-Path Decoding (DPD): This module utilizes an Orthogonal Split-and-Quantize (OSQ-VAE) tokenizer to generate two complementary semantic ID channels. By decoding these channels in parallel and fusing the results, the model provides a cross-channel recovery mechanism for items missed by a single path.

BARGE demonstrates superior performance across public benchmarks compared to existing GR baselines. In a large-scale industrial A/B test on a Tencent media platform, the model achieved a 0.60% increase in click-through rate, 1.34% in click unique visitors, and 1.70% in total reading time, confirming its practical effectiveness in real-world, high-scale recommendation scenarios.