Asymmetric Generative Recommendation via Multi-Expert Projection and Multi-Faceted Hierarchical Quantization

Abstract: Generative Recommendation (GenRec) models reformulate recommendation as a sequence generation task, representing items as discrete Semantic IDs used symmetrically as both inputs and prediction targets. We identify a critical dual-stage information bottleneck in this design: (1) the Input Bottleneck, where lossy quantization degrades fine-grained semantics, while popularity bias skews the learned representations toward frequent items, and (2) the Output Bottleneck, where imprecise discrete targets limit supervision quality. To address these issues, we propose AsymRec, an asymmetric continuous-discrete framework that decouples input and output representations. Specifically, Multi-expert Semantic Projection (MSP) maps continuous embeddings into the Transformer's hidden space via expert-specialized projections, preserving semantic richness and improving generalization to infrequent items. Multi-faceted Hierarchical Quantization (MHQ) constructs high-capacity, structured discrete targets through multi-view and multi-level quantization with semantic regularization, preventing dimensional collapse while retaining fine-grained distinctions. Extensive experiments demonstrate that AsymRec consistently outperforms state-of-the-art generative recommenders by an average of 15.8 %. The code will be released.

• The paper introduces AsymRec, which combines Multi-Expert Semantic Projection for continuous inputs with Multi-faceted Hierarchical Quantization for discrete outputs, improving average NDCG@10 by 15.8% over the strongest baseline across four Amazon datasets.
• AsymRec preserves fine-grained and long-tail item semantics while using structured discrete targets to prevent representation collapse, achieving an effective rank of 178.1 versus 99.5 for continuous-output regression and outperforming standard product quantization with fewer tokens.
• The method also shows practical deployment value, delivering a 1.4% increase in total consumption and a 1.9% GMV uplift in a seven-day, 1% online A/B test, although broader cold-start and fusion studies remain open.

Motivation and problem statement

Generative Recommendation (GenRec) systems reformulate sequential recommendation as autoregressive sequence generation over discrete Semantic IDs (SIDs), following the paradigm established by TIGER and extended by systems such as HSTU, RPG, and OneRec. The paper under review identifies a structural weakness in this dominant design: input and output representations are symmetric, both derived from a single lossy quantization of continuous semantic embeddings. The authors decompose this into a dual-stage information bottleneck. On the input side, quantization discards fine-grained semantics irrecoverably, and lookup-table ID embeddings are updated disproportionately for frequent items, inducing popularity bias and poor cold-item generalization. On the output side, codebook collisions and high reconstruction error yield noisy supervision targets. The authors also argue—counterintuitively but with empirical support—that simply regressing continuous embeddings at the output is not a remedy, because it induces dimensional collapse.

The proposed framework, AsymRec, breaks this symmetry: inputs are continuous, outputs are discrete, with each side optimized independently.

Method

Multi-expert Semantic Projection (MSP)

MSP replaces SID lookup entirely at the input stage. Each item's raw embedding $x_i$ is mapped into the Transformer's hidden space via a Mixture-of-Experts composition:

$$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$

where each expert $f_e$ is a 2-layer MLP and the gating network $g$ produces normalized weights. Because the mapping is continuous and learned directly from the encoder embedding, the topological structure of the item space is preserved, allowing semantically similar items—including infrequent ones—to remain nearby in representation space. With $E=3$ experts, MSP is lightweight relative to the overall model.

Multi-faceted Hierarchical Quantization (MHQ)

MHQ constructs the output-side discrete targets by combining product quantization's multi-facet decomposition with residual quantization's hierarchical depth. An embedding is first projected via a learnable matrix $W_P$, partitioned into $M$ disjoint subspaces, and each subspace is quantized with $L$ levels of residual quantization using EMA codebook updates ($\gamma = 0.99$). The resulting SID has length $M \times L$. Training uses three losses: reconstruction error $$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$0, a subspace energy-balance loss penalizing uneven variance allocation across facets, and an orthogonality regularizer $$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$1 on the projection. Notably, MHQ is trained as a separate stage; its loss does not participate in recommendation-model training.

Architecture

AsymRec uses a shallow Transformer decoder ($$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$2 layers, hidden dimension 448) over MSP-projected inputs with positional encodings. The final hidden state feeds $$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$3 parallel classification heads, each predicting one codeword index via cross-entropy averaged over all heads. Inference uses graph-constrained decoding to guarantee valid codeword sequences.

Experimental results

Experiments use four Amazon Review categories (Sports, Beauty, Toys, CDs) with text-embedding-3-large encoders ($$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$4) and leave-last-out evaluation. AsymRec ranks first on all metrics across all four datasets, outperforming the strongest baseline (RPG) by an average of 15.8% on NDCG@10. For example, on Beauty it attains NDCG@10 of 0.0516 versus RPG's 0.0464; on CDs, Recall@5 of 0.0614 versus 0.0498.

Ablations isolate each component:

Variant	NDCG@10
Full AsymRec	0.0516
Discrete SIDs as inputs	0.0491
Single expert (capacity-matched)	0.0508
Continuous embeddings as outputs	0.0406
Standard PQ instead of MHQ	0.0494

Three findings deserve emphasis. First, replacing continuous inputs with discrete codes degrades performance, and a frequency-binned retrieval analysis shows the discrete-input variant performs well only on the highest-frequency bin while AsymRec maintains balanced recall across the long tail—an unusually direct demonstration that symmetric quantized inputs cause popularity bias. Second, the fully continuous pipeline suffers the largest degradation (NDCG@10 of 0.0406). The authors attribute this to representation collapse and quantify it via Effective Rank: the continuous-output variant achieves an effective rank of only 99.5, versus 178.1 for discrete SID outputs, whose singular spectrum decays far more gradually. This supports the claim that discrete classification targets act as a regularizer preventing "lazy" mean-like predictions. Third, MHQ dominates standard PQ on a token-efficiency basis: with $$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$5 (24 tokens) it reaches NDCG@10 of 0.0514, exceeding the best PQ configuration ($$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$6, 64 tokens) at 0.0494. Gains saturate beyond $$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$7.

An interesting secondary result: Reciprocal Rank Fusion of the continuous-input and discrete-input variants yields NDCG@10 of 0.0540, exceeding either alone—evidence that the two representations are complementary, though the authors leave systematic fusion optimization unexplored.

Industrial deployment

Beyond offline benchmarks, the authors deploy the method in a production pCVR system on a large advertising platform, integrating SIDs derived from cross-domain latent factors and multimodal LLM embeddings as categorical features, trained end-to-end with a joint pCVR plus reconstruction objective. A seven-day A/B test on 1% of traffic reports a 1.4% lift in total consumption and 1.9% GMV uplift, statistically significant against the production baseline. This is a meaningful validation that asymmetric representation learning transfers from academic benchmarks to industrial ranking.

Limitations and open questions

Several caveats are worth noting. The offline evaluation relies exclusively on Amazon benchmarks with 5-core filtering, which truncates the long tail that the cold-item claims address; the frequency-binned analysis mitigates but does not eliminate this concern. The online deployment applies MHQ-derived SIDs as features within a discriminative ranker rather than as generative targets, so the A/B results validate the tokenization more than the full asymmetric generation pipeline. The RRF fusion result suggests the discrete-input variant retains value that AsymRec discards, and the authors explicitly defer comprehensive fusion optimization to future work. Finally, the choice of hyperparameters ($$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$8, $$h_i = \sum_{e=1}^{E} \alpha_{i,e}\, f_e(x_i), \quad \boldsymbol{\alpha}_i = g(x_i)$$9, saturation at $f_e$0) is empirical; no theoretical account is given for why these configurations suffice, nor for when continuous-output regression might be stabilized by alternative regularization.

Conclusion

AsymRec makes a focused architectural argument: the symmetry between input and output representations in GenRec is a source of avoidable information loss. Continuous expert-specialized projection preserves input-side semantics and improves tail generalization, while structured hierarchical quantization provides higher-capacity, collapse-resistant supervision than either pure PQ or continuous regression. The combination of consistent benchmark gains (15.8% average NDCG@10 improvement), mechanistic evidence via effective-rank analysis, and statistically significant production A/B results constitutes a well-supported case for asymmetric representation design in generative recommenders.