Research map / Large language models / Language, safety and interpretability

Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization

AAAI 2025 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations · 3 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 a two-stage LLM-based recommender that tokenizes items into four codebook indices aligned with LLaMA's embedding space, then fine-tunes LLaMA-7B with sequential, text, query and negative-sampling tasks, and pre-caches top-K results for online serving.

Problem

Sparse collaborative item IDs are hard to integrate with dense LLM token representations; large item vocabularies make LLM-based sequential recommendation costly and misaligned with the LLM's semantic space.

Approach

An RQ-VAE-style cascaded codebook (N=4 levels, C=256 codes) built on DCCF behavior embeddings, with an added LLM-alignment loss pulling codes toward averaged LLM input embeddings of item text, producing item token indices for LLM fine-tuning plus a negative-behavior ("what will user hate next") task and beam-search top-K pre-caching.</203>

Key results

• On Games, the method reaches HR@10 0.1274 and NDCG@10 0.0724 versus LC-Rec's 0.1167 and 0.0651, with HR@1 improvement of +15.77% over LC-Rec.
• Ablation on Games: removing negative sampling drops HR@5 from 0.0858 to 0.0791; replacing the tokenizer with an encoder-decoder (ED) drops it to 0.0817.
• Using only behavior sequences on Games, the method achieves HR@5 0.0773 vs LC-Rec 0.0708, though textual information further improves to 0.0858.

Measured effects reported in the paper

Surprising findings

• More advanced LLM versions (Llama-2-7B) did not outperform Llama-7B, which the authors attribute to domain mismatch of extra pretraining corpora.

Limitations

• Pre-caching is constrained by storage and is ill-suited to frequently changing item pools.
• LLM training costs remain high relative to traditional recommenders.
• Only text and behavior modalities are used; scaling experiments limited to 13B models due to compute.

Experimental setup

Datasets: Amazon Games Amazon Arts Amazon Instruments

Benchmarks: Amazon Games Amazon Arts Amazon Instruments

Metrics: HR@1 HR@5 HR@10 NDCG@5 NDCG@10

Baselines: SASRec BERT4Rec FMLP DCCF P5 LC-Rec

Builds on: RQ-VAE LC-Rec DCCF P5

Closest papers by meaning

• Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.90)
• Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.89)
• LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.89)
• LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.89)
• LEARN: Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application (AAAI 2025 · similarity 0.89)
• CoRA: Collaborative Information Perception by Large Language Model’s Weights for Recommendation (AAAI 2025 · similarity 0.88)