AI
Diffusion Models Boost Recommendation Systems with Bidirectional Semantic ID Generation
Researchers introduce DiffGRM, a diffusion-based approach that outperforms traditional autoregressive methods by generating item IDs in parallel rather than sequentially.
Recommendation systems have increasingly turned to generative approaches that encode items as multi-digit semantic identifiers (SIDs) and predict the next item by generating these codes sequentially. However, left-to-right autoregressive generation introduces two key limitations: it prevents the model from using bidirectional context to validate whether all digits form a coherent item, and it treats all digits equally despite their varying difficulty and semantic importance.
The Problem with Sequential Generation#
Current generative recommendation models use autoregressively-masked architectures that process SID digits one at a time from left to right. This sequential constraint creates two bottlenecks. First, each digit only receives information from earlier digits in the sequence, blocking the model from leveraging later-position context to refine earlier decisions—a critical handicap when all n digits must jointly identify a single coherent item. Second, the uniform next-token objective applies identical training pressure to every digit, regardless of whether a digit carries semantic granularity or prediction difficulty, leading to overtraining on easy digits and undertraining on hard ones.
DiffGRM: A Parallel Alternative#
To overcome these limitations, researchers propose DiffGRM (Diffusion-based Generative Recommendation Model), which replaces the autoregressive decoder with a masked discrete diffusion model. This shift enables the system to generate all SID digits in parallel while exploiting bidirectional context, and to concentrate training effort on the digits that matter most. The approach improves NDCG@10 scores by 6.9% to 15.5% over existing generative and discriminative baselines across multiple datasets.
The model operates across three coordinated stages:
- Tokenization via Parallel Semantic Encoding (PSE): decouples digits so information is distributed evenly rather than cascading in causal chains
- Training via On-policy Coherent Noising (OCN): masks uncertain digits strategically to prioritize high-value supervision signals and balance learning across semantic levels
- Inference via Confidence-guided Parallel Denoising (CPD): prioritizes denoising higher-confidence digits first and generates multiple Top-K candidate items with controlled diversity
