BONSAI framework boosts LLM-based recommendations by 21.6%

arXiv paper 2607.16633 from researchers at Snap (Jiliang Tang, Liam Collins, Tong Zhao, Neil Shah, Mingxuan Ju) and Michigan State University (Jingzhe Liu, Hanbing Wang) tackles a critical but overlooked aspect of LLM-based generative recommendation: the structural design of the decoding trie. While existing work focuses on generating more comprehensive term IDs (e.g., titles or keywords) to describe items, the trie formed by these terms is often poorly suited to constrained beam search, degrading performance. The authors identify two key properties for an effective trie: adaptive and variable ID length (allowing items with varying semantic richness to use IDs of appropriate lengths) and constrained branching factors at shallow levels (drastically improving beam search success).

To realize these properties, BONSAI extracts recommendation-informative words from item metadata and uses a minimum set cover formulation to recursively build the trie. Experiments demonstrate up to a 21.6% relative improvement over state-of-the-art baselines. Further analysis confirms the generalizability of these properties — they can be applied to enhance other term ID methods as well. This work is particularly relevant as generative recommendation becomes an increasingly popular paradigm, using LLMs as autoregressive backbones to predict next-item term IDs in a constrained search space.