Research map / Label-efficient and robust learning / Recognition and long-tail

Rethinking Semantic ID Construction for Generative Recommendation: SimHash with Parallel Decoding and Semantic Alignment

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Recommendation representation learning

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Revisits SimHash semantic IDs for generative recommendation and shows that parallel multi-token decoding plus a cosine alignment loss to frozen LLM embeddings lets training-free hashing match or beat learned quantizers on four Amazon datasets.

Problem

Semantic ID-based generative recommendation usually relies on learned quantizers (RQ-VAE, OPQ); training-free hashing such as SimHash is believed to be inferior, partly because autoregressive decoding mismatches its unordered codes and hashing discards fine-grained semantics.

Approach

Pair training-free SimHash codes over text embeddings with parallel (multi-token) prediction heads, which match the unordered code structure, and add a negative-cosine alignment loss that anchors item representations to frozen LLM embeddings to compensate for hashing information loss.

Key results

• FLASH achieves the best R@5/N@5/R@10/N@10 on all four datasets, e.g. on Toys it improves over RPG by 5.18% R@10 and 11.02% N@10.
• Alignment loss improves R@10 on Beauty by 13.06% for SASRec, 9.57% for TIGER and 7.49% for FLASH.
• SimHash tokenization takes under 0.5 s on CPU across datasets versus tens to hundreds of seconds for learned quantizers (e.g. RQ-VAE 67-147 s on GPU).

Surprising findings

• Random code assignment with parallel decoding and alignment reaches much lower accuracy, but SimHash matches or beats learned quantizers (RQ-VAE, RQ-KMeans, OPQ) on most datasets, contradicting the prior TIGER ablation conclusion.
• Semantic alignment improves performance consistently across SASRec, TIGER and FLASH, suggesting it is a model-agnostic regularizer.

Limitations

• Depends on quality of pre-computed text embeddings; degrades with sparse or noisy text.
• Hash collisions may grow at larger scales or different codebook settings.
• Evaluated only on public Amazon benchmarks; no industrial-scale validation and text-only (no images).

Experimental setup

Datasets: Amazon Beauty Amazon Toys and Games Amazon Sports and Outdoors Amazon CDs and Vinyl

Benchmarks: Amazon Beauty Amazon Toys and Games Amazon Sports and Outdoors Amazon CDs and Vinyl

Metrics: Recall@5 NDCG@5 Recall@10 NDCG@10

Baselines: Caser GRU4Rec HGN BERT4Rec SASRec S3-Rec TIGER LETTER DIGER RecJPQ VQ-Rec RPG

Builds on: TIGER RPG SASRec LSH

Closest papers by meaning

• Language Models as Semantic Indexers (ICML 2024 · similarity 0.87)
• Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.87)
• Efficient Inference for Large Language Model-based Generative Recommendation (ICLR 2025 · similarity 0.86)
• HashAttention: Semantic Sparsity for Faster Inference (ICML 2025 · similarity 0.85)
• Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.85)
• Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations (NeurIPS 2025 · similarity 0.85)