JevRec: Linguistic Encoding with Jev-like Decision Modeling for LLM-based Generative Recommendation

Most generative recommendation methods assign the same semantic identifier length to every item even though interaction frequencies are highly unequal. Uniform length repeatedly pays the cost of fine-grained codes for popular items, while shortening codes to cut that cost compresses the semantic representations of long-tail items.

The Jev-like decision modeling component predicts a length distribution with a dedicated head and generates content conditioned on that length through a separate head, while both heads share the interaction-history representation.

Predicting a sequence-level length and generating token-level content impose different optimization requirements that can conflict under a shared output head. The paper reports reduced gradient conflict between length and content prediction with the split design.

The paper reports experiments on three public datasets and states that JevRec outperforms all compared baselines across the evaluated recommendation metrics. The digest does not name the individual datasets.

The paper reports lower training and inference time than TIGER, alongside reduced gradient conflict between length and content prediction and better results than the compared baselines on the evaluated recommendation metrics.

Short codes reduce the cost of fine-grained codes for popular items but compress the semantic representations of long-tail items. JevRec instead gives popular items shorter identifiers and long-tail items richer descriptions from shared codebooks.

The provided material describes JevRec as an arXiv paper covering the method and its experiments; it contains no installation instructions, repository link, or model release details, so the paper itself is the resource described here.