Research map / Large language models / Instruction and fine-tuning

FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens

NeurIPS 2025 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 1 citations (OpenAlex)

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.

Proposes FACE, a model-agnostic framework that maps collaborative filtering embeddings into pre-trained LLM tokens (descriptors) via a disentangled, vector-quantized autoencoder with a frozen LLM-vocabulary codebook and contrastive alignment to text summaries, improving several CF backbones without fine-tuning the LLM.

Problem

LLMs cannot interpret the latent, non-semantic embeddings produced by collaborative filtering models, which limits recommendation effectiveness and interpretability when integrating LLMs with CF.

Approach

Decompose CF embeddings into concept-specific vectors via a multi-projector plus transformer, quantize them with a residual quantizer whose shared codebook is a frozen, filtered subset of pretrained LLM token embeddings (projected by a trainable matrix), and align the resulting descriptor sentences with user/item text summaries via contrastive learning, so the LLM can read CF embeddings as tokens

Key results

• FACE improves backbones with maximum gains of 7.31% on Amazon-book, 11.55% on Yelp, and 8.00% on Steam, per the paper's claims.
• On Amazon-book, LightGCN R@20 rises from 0.1563 to 0.1622 with FACE, and NDCG@20 from 0.0961 to 0.1009.
• Real-user and LLM ranking of interaction explanations gives FACE descriptors a slightly better (lower) average rank than RLMRec profiles (LLM: 1.700 vs 1.800).

Measured effects reported in the paper

Surprising findings

• Qwen2-based FACE performs marginally worse, which the authors attribute to its specialization in query-document matching rather than semantic similarity.
• Smaller embedding models (MiniLM) perform comparably to LLaMA2-7B when used as the FACE backbone.

Limitations

• Only relatively small LLMs (LLaMA2-7B, MiniLM-L6, Qwen2-7B) were tested; behavior with larger LLMs is unverified.
• Applicability is limited to collaborative filtering; sequence recommendation is unexplored.
• Potential LLM biases carry over to generated descriptors and explanations.

Experimental setup

Datasets: Amazon-book Yelp Steam COCA

Benchmarks: Amazon-book Yelp Steam

Metrics: Recall@5 Recall@20 NDCG@5 NDCG@20 I@10 I@50

Baselines: GMF LightGCN SimGCL LightGCL RLMRec CTRL KAR AlphaRec

Builds on: VQ-VAE RQ-VAE SimVQ RLMRec LightGCN

Closest papers by meaning

• Language Representations Can be What Recommenders Need: Findings and Potentials (ICLR 2025 · similarity 0.90)
• LLMEmb: Large Language Model Can Be a Good Embedding Generator for Sequential Recommendation (AAAI 2025 · similarity 0.88)
• Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.88)
• 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 0.88)
• On the Embedding Collapse when Scaling up Recommendation Models (ICML 2024 · similarity 0.87)