Soft-token fusion lets LLM recommenders use numerical and embedding features
New framework maps continuous features into LLM token space, boosting retrieval by 12%.
A team of 10 researchers from Amazon and academia (Zhe Xu, Ankit Peshin, et al.) present a soft-token fusion framework that maps continuous numerical features and dense embeddings into the LLM embedding space, allowing LLM-based recommender systems to consume non-textual signals through the standard token interface. Their method, instantiated in a shared-parameter two-tower retrieval model, uses an interaction-based fusion module to refine soft tokens before final LLM input.
On three Amazon recommendation benchmarks, the approach improves retrieval performance over existing LLM-based baselines by up to 12% in hit rate and recall. The interaction-based fusion consistently outperforms simple concatenation of heterogeneous soft tokens, showing that cross-feature interactions matter. This work addresses a key limitation of LLM recommenders—their reliance on discrete text—and makes them viable for production systems where numerical and embedding features are ubiquitous.
- Soft-token fusion maps numerical and embedding features into LLM embedding space via interaction-based module.
- Improves retrieval by up to 12% on Amazon benchmarks over LLM-only baselines.
- Interaction-based fusion beats direct concatenation of heterogeneous soft tokens.
Why It Matters
Opens production-grade recommender systems to LLMs by handling continuous and dense features from existing pipelines.