# Differentiable Semantic ID for Generative Recommendation

arXiv: https://arxiv.org/abs/2601.19711 · HTML: https://mlsanity.com/paper/2601.19711/

## Key facts

- Cited by 19 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 27 Jan 2026.

## Built on by (most cited first, 19 of 19)

- [Understanding Semantic IDs: From Item Representation to Item Selection in Generative Recommendation](https://mlsanity.com/paper/2607.24995/) (2026)
- [Learning from Unreachable Rewards: Hint-Conditioned Reinforcement Learning for Generative Recommendation](https://mlsanity.com/paper/2608.11980/) (2026)
- [PRQ-KMeans: Projection Residual Quantization for Semantic ID Tokenization](https://mlsanity.com/paper/2608.24207/) (2026)
- [History-Conditioned Joint-Prefix Alignment for Generative Recommendation](https://mlsanity.com/paper/2608.29179/) (2026)
- [Recommendation Retrievers Need Verifiers: Universal Generative Reranking for Sequential Recommendations](https://mlsanity.com/paper/2609.12270/) (2026)
- [One Model, Two Markets: Bid-Aware Generative Recommendation](https://mlsanity.com/paper/2603.22231/) (2026)
- [Do We Really Need Multimodal Emotion Language Models Larger Than 1B Parameters?](https://mlsanity.com/paper/2607.12787/) (2026)
- [VaLiDRec: Variable-Length LLM-Aligned Semantic IDs for Generative Recommendation](https://mlsanity.com/paper/2607.25209/) (2026)
- [OneShot: Index-in-Ranking with Neural Scoring for Large-Scale Retrieval](https://mlsanity.com/paper/2607.27475/) (2026)
- [EchoRec: Multi-Item Prediction-Empowered Generative Recommendation via Cycle-Consistent Preference Alignment](https://mlsanity.com/paper/2608.14011/) (2026)
- [SIDScope: A Diagnostic Resource for Semantic-ID Interfaces in Generative Recommendation](https://mlsanity.com/paper/2608.18779/) (2026)
- [From a Static Multi-Level Small Semantic Codebook to a Dynamic Single-Level Large Semantic Codebook for Generative Recommendation](https://mlsanity.com/paper/2608.21012/) (2026)
- [Preference Shapes Relevance: Cross-component Hierarchical Semantic Alignment for Personalized Generative Retrieval](https://mlsanity.com/paper/2608.30553/) (2026)
- [Guiding the coarse levels of semantic IDs makes the fine levels learnable](https://mlsanity.com/paper/2609.22227/) (2026)
- [Advancing Model Research in AgentX: Long-Horizon Autonomy for Industrial Recommender Systems](https://mlsanity.com/paper/2609.30001/) (2026)
- [Neither Black nor White: Balancing Semantic and Collaborative Signals with Graph-Informed Semantic IDs (GrIS)](https://mlsanity.com/paper/2610.01533/) (2026)
- [Reasoning with Evidence, Not Merely Rationales: Verifiable Preference Proofs for LLM-Based Recommendation](https://mlsanity.com/paper/2610.02968/) (2026)
- [Beyond Semantic IDs: Finite-Beam Reachability and Continuous-Query Generation in Generative Recommendation](https://mlsanity.com/paper/2610.12977/) (2026)
- [WeGR: Unified Generative Retrieval for Industrial Search Systems in WeChat](https://mlsanity.com/paper/2610.16979/) (2026)

## Abstract (arXiv)

Generative recommendation provides a novel paradigm in which each item is represented by a discrete semantic ID (SID) learned from rich content. Most existing methods treat SIDs as predefined and train recommenders under static indexing. In practice, SIDs are typically optimized only for content reconstruction rather than recommendation accuracy. This leads to an objective mismatch: the system optimizes an indexing loss to learn the SID and a recommendation loss for interaction prediction, but because the tokenizer is trained independently, the recommendation loss cannot update it. A natural approach is to make semantic indexing differentiable so that recommendation gradients can directly influence SID learning, but this often causes codebook collapse, where only a few codes are used. We attribute this issue to early deterministic assignments that limit codebook exploration, resulting in imbalance and unstable optimization. In this paper, we propose DIGER (Differentiable Semantic ID for Generative Recommendation), a first step toward effective differentiable semantic IDs for generative recommendation. DIGER introduces Gumbel noise to explicitly encourage early-stage exploration over codes, mitigating codebook collapse and improving code utilization. To balance exploration and convergence, we further design two uncertainty decay strategies that gradually reduce the Gumbel noise, enabling a smooth transition from early exploration to exploitation of learned SIDs. Extensive experiments on multiple public datasets demonstrate consistent improvements from differentiable semantic IDs. These results confirm the effectiveness of aligning indexing and recommendation objectives through differentiable SIDs and highlight differentiable semantic indexing as a promising research direction. Our code is released under https://github.com/junchen-fu/DIGER.