{"slug": "jevrec-linguistic-encoding-with-jev-like-decision-modeling-for-llm-based", "title": "JevRec: Linguistic Encoding with Jev-like Decision Modeling for LLM-based Generative Recommendation", "summary": "Researchers proposed JevRec, a generative recommendation method that encodes items as variable-length words rather than assigning every item the same semantic identifier (SID) length, and reported that it outperformed all compared baselines across the evaluated recommendation metrics on three public datasets. JevRec keeps frequency-dependent semantic prefixes from shared codebooks, giving popular items shorter SIDs and long-tail items richer descriptions, and uses a Jev-like decision modeling component with a dedicated head to predict a length distribution and a separate head to generate content conditioned on that length while sharing the interaction-history representation. The authors report lower training and inference time than TIGER and reduced gradient conflict between length and content prediction.", "body_md": "arXiv:2610.36937v1 Announce Type: new \nAbstract: Generative recommendation (GR) represents items as semantic identifiers (SIDs) and autoregressively generates the next item's SID from interaction history. However, most methods assign the same SID length to all items despite highly unequal interaction frequencies. This uniform design repeatedly incurs the cost of fine-grained codes for popular items, while shortening codes to reduce this cost also compresses the semantic representations of long-tail items. Inspired by the compact forms of frequent expressions in natural language, we propose JevRec to encode items as variable-length words and generate the next item from interaction context. Linguistic encoding retains frequency-dependent semantic prefixes from shared codebooks, producing shorter SIDs for popular items and richer descriptions for long-tail items. Yet predicting a sequence-level length and generating token-level content impose different optimization requirements that may conflict under a shared output head. Our Jev-like decision modeling component therefore predicts a length distribution with a dedicated head and generates content conditioned on length through a separate head while sharing the interaction-history representation. Experiments on three public datasets show that JevRec outperforms all compared baselines across the evaluated recommendation metrics. Further analyses show lower training and inference time than TIGER and reduced gradient conflict between length and content prediction.", "url": "https://wpnews.pro/news/jevrec-linguistic-encoding-with-jev-like-decision-modeling-for-llm-based", "canonical_source": "https://www.machinebrief.com/news/jevrec-linguistic-encoding-with-jev-like-decision-modeling-f-70j5", "published_at": "2026-10-28 04:00:00+00:00", "updated_at": "2026-10-28 04:41:57.301144+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "generative-ai", "ai-research"], "entities": ["JevRec", "TIGER"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/jevrec-linguistic-encoding-with-jev-like-decision-modeling-for-llm-based", "markdown": "https://wpnews.pro/news/jevrec-linguistic-encoding-with-jev-like-decision-modeling-for-llm-based.md", "text": "https://wpnews.pro/news/jevrec-linguistic-encoding-with-jev-like-decision-modeling-for-llm-based.txt", "jsonld": "https://wpnews.pro/news/jevrec-linguistic-encoding-with-jev-like-decision-modeling-for-llm-based.jsonld"}}