Fair comparison of generative vs. dense retrieval paradigms

Determine, under equivalent input information and experimental conditions, how sequential generative retrieval based on semantic IDs compares to sequential dense retrieval in terms of recommendation performance and the associated computation and storage trade-offs.

Background

The paper contrasts two paradigms for sequential recommendation: dense retrieval, which learns and stores a unique embedding per item and performs maximum inner product search over all items, and generative retrieval, which predicts item indices (semantic IDs) with a generative model using beam search. While generative retrieval reduces storage and inference costs, prior work lacks fair, controlled comparisons against dense retrieval using matched inputs and setups.

The authors highlight that, despite generative retrieval’s promise, questions remain regarding comparative performance and practical trade-offs. They conduct experiments and propose a hybrid model (LIGER) to explore these questions in small-scale academic benchmarks, but the general comparative understanding itself is stated to be open.

References

Despite its potential for large-scale applications, a comprehensive comparison between generative retrieval and sequential dense retrieval under fair conditions is still lacking, leaving open questions regarding performance, and computation trade-offs.

— Unifying Generative and Dense Retrieval for Sequential Recommendation  (2411.18814 - Yang et al., 2024) in Abstract

Full-catalog CE costs $O(N)$ per update and has only been evaluated on the reported catalog sizes; billion-item training, ANN latency, memory, and dynamic-catalog accuracy remain open.

— Beyond Semantic IDs: Finite-Beam Reachability and Continuous-Query Generation in Generative Recommendation  (2610.12977 - Sun et al., 9 Oct 2026) in Section Discussion and Limitations, paragraph “Limitations”