Research map / Large language models / Language, safety and interpretability
Catalog-Native LLM: Speaking Item-ID dialect with Less Entanglement for Recommendation
Reza Shirkavand, Xiaokai Wei, Chen Wang, Zheng Hui, Heng Huang, Michelle Gong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add text-derived bias to ID-only model | ID-only Qwen2.5-0.5B · Amazon Arts | HR@10: +42.8% | Improved |
| add explicit attribute text tokens to ID-only model | ID-only Qwen2.5-0.5B · Amazon Arts | HR@10: +24.6% | Improved |
| use IDIOMoE token-type expert split instead of plain ID-only | ID-only Qwen2.5-0.5B · Industrial dataset | HR@10: +22.7% | Improved |
| add MoE item/text expert separation over Item-LLM (ablation) | Item-LLM · Industrial dataset (relative to SASRec) | NDCG@10 improvement (%): +23.5% -> +27.1% | Improved |
| use item expert shrink factor 4 on item FFN (ablation) | IDIOMoE · Amazon Beauty | NDCG@10 / HR@10: +41.8% NDCG@10, +26.6% HR@10 | Improved |
| shrink item expert aggressively (shrink=8) (ablation) | IDIOMoE · Amazon Beauty | NDCG@10 / HR@10 | Hurt |
| shrink item expert capacity (ablation) | IDIOMoE · Industrial dataset | NDCG@10 / HR@10 | Hurt |
| place MoE layers in last 8 layers (ablation) | IDIOMoE · Amazon Arts | HR@10 / NDCG@10: +27.6% HR@10 and +28.4% NDCG@10 | Improved |
| place MoE only in first 8 layers (ablation) | IDIOMoE · Amazon Arts | HR@10 / NDCG@10 | Hurt |
| replace static token-type routing with switch-style dynamic gating (ablation) | IDIOMoE · Amazon recommendation | recommendation quality | Hurt |
| replace MoE with append-blocks or prepend-blocks capacity (ablation) | IDIOMoE · Amazon Beauty and industrial | recommendation accuracy | Hurt |
| replace MoE with LoRA adapters on all layers (ablation) | IDIOMoE · Industrial dataset | recommendation accuracy | Hurt |
Closest papers by meaning
- Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.90)
- Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.89)
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.89)
- CoRA: Collaborative Information Perception by Large Language Model’s Weights for Recommendation (AAAI 2025 · similarity 0.89)
- LEARN: Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application (AAAI 2025 · similarity 0.89)
- Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
