Research map / Large language models / Language, safety and interpretability
Language Representations Can be What Recommenders Need: Findings and Potentials
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that linearly mapping LM title embeddings into a recommendation space yields strong CF performance, and proposes AlphaRec, an MLP + graph convolution + InfoNCE model built purely on language representations that beats ID-based CF baselines and shows zero-shot and intention-aware abilities.
Problem
Whether language models implicitly encode user preference (collaborative) information in their representation space, and whether recommenders can be built purely from language representations rather than ID-based embeddings.
Approach
Language representations of item titles, linearly (or with an MLP plus graph convolution) projected into a recommendation space trained with InfoNCE, achieve homomorphism-like performance that rivals or beats ID-based CF, implying collaborative signals are implicitly encoded in LMs.
Key results
- AlphaRec outperforms the best ID-based baseline (RLMRec / XSimGCL) on Recall@20 by 6.79% to 9.75% across Movies & TV, Video Games, and Books.
- In zero-shot transfer to Industrial & Scientific, MovieLens-1M and Book Crossing, AlphaRec improves Recall@20 over the best zero-shot baseline (VQ-Rec) by 157.09%, 66.67% and 101.55% respectively.
- In user intention capture on MovieLens-1M, AlphaRec with intention reaches HR@5 0.4704 versus 0.2738 for TEM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ID-based CF with AlphaRec using language representations and MLP, graph convolution, and InfoNCE | Movies & TV recommendation · Movies & TV | Recall@20: 0.1119 -> 0.1221 | Helped |
| replace ID-based CF with AlphaRec using language representations and MLP, graph convolution, and InfoNCE | Video Games recommendation · Video Games | Recall@20: 0.1384 -> 0.1519 | Helped |
| replace ID-based CF with AlphaRec using language representations and MLP, graph convolution, and InfoNCE | Books recommendation · Books | Recall@20: 0.0928 -> 0.0991 | Helped |
| train AlphaRec jointly on three source datasets and evaluate without target-dataset training | AlphaRec · Industrial & Scientific | Recall: 0.0645 -> 0.0913 | Helped |
| train AlphaRec jointly on three source datasets and evaluate without target-dataset training | AlphaRec · MovieLens-1M | Recall: 0.0804 -> 0.1486 | Helped |
| train AlphaRec jointly on three source datasets and evaluate without target-dataset training | AlphaRec · Book Crossing | Recall: 0.0485 -> 0.0660 | Helped |
| add user-intention representation at inference | AlphaRec · MovieLens-1M | HR@5: 0.0793 -> 0.4704 | Helped |
| add user-intention representation at inference | AlphaRec · Video Games | HR@5: 0.0663 -> 0.2569 | Helped |
| add random letters to the item title prompt (ablation) | Llama2-7B linear mapping · Movies & TV | Recall: 0.1027 -> 0.0952 | Hurt |
| add random letters to the item title prompt (ablation) | Llama2-7B linear mapping · Video Games | Recall: 0.1249 -> 0.1213 | Hurt |
| add random letters to the item title prompt (ablation) | Llama2-7B linear mapping · Books | Recall: 0.0662 -> 0.0632 | Hurt |
| replace XSimGCL ID embeddings with language representations | XSimGCL · Movies & TV, Video Games, and Books | Recall@20 | Mixed |
Surprising findings
- Linear mapping of BERT and RoBERTa title embeddings performs at or below MF, while LLM-based embeddings perform well, indicating collaborative signals emerge only in advanced LMs.
- Simply replacing ID embeddings in XSimGCL with language representations (XSimGCL_t) does not give stable gains, suggesting model design matters beyond input replacement.
Limitations
- Possible information leakage since LM pretraining data may contain the benchmark data (partly addressed with Amazon 2023).
- Zero-shot baselines were modified from sequential recommendation methods, so comparisons are approximate.
Experimental setup
Datasets: Amazon Movies & TV Amazon Video Games Amazon Books Amazon Industrial & Scientific MovieLens-1M Book Crossing
Benchmarks: Amazon Movies & TV Amazon Video Games Amazon Books Amazon Industrial & Scientific MovieLens-1M Book Crossing
Metrics: Recall@20 NDCG@20 HR@20 HR@5 NDCG@5
Baselines: MF MultVAE LightGCN SGL BC Loss XSimGCL KAR RLMRec ZESRec UniSRec VQ-Rec LLMRank
Builds on: LightGCN InfoNCE XSimGCL RLMRec KAR
Closest papers by meaning
- FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens (NeurIPS 2025 · similarity 0.90)
- Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.89)
- Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.88)
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.87)
- Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.87)
- Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.87)
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.