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

CoRA: Collaborative Information Perception by Large Language Model’s Weights for Recommendation

AAAI 2025 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems · 0 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 CoRA, which converts collaborative filtering embeddings into low-rank LoRA-style weight updates merged into a frozen LLM, so collaborative information is perceived in parameter space rather than via input tokens.

Problem

Adapting LLMs for recommendation by injecting collaborative features into the input space disrupts the LLM's general knowledge and the semantics of textual prompts, yielding sub-optimal recommendation performance.

Approach

A collaborative query generator (Q-Former-style, with learnable queries cross-attending to CF user/item embeddings) outputs low-rank collaborative weights that are added to the LLM's attention and output weights, so the frozen LLM perceives collaborative signals without fine-tuning or extra prompt tokens.

Key results

• CoRA-MF reaches AUC 0.8179 and UAUC 0.6262 on Amazon-Book, above the best baseline BinLLM (0.8157 / 0.5724).
• On ML-1M, CoRA-MF reaches AUC 0.7361 and UAUC 0.6884, versus 0.7164 AUC for the best CoLLM variant.
• Ablation shows inserting collaborative weights into query, key, value and output projections (qkvo) performs best; adding feed-forward weights did not help.

Measured effects reported in the paper

Surprising findings

• Fine-tuning Vicuna-7B on recommendation data degraded its performance on general benchmarks (WikiFact, SocialIQA, XSum) and on its own recommendation-text tasks, as shown in the pilot study.
• CoLLM's performance on ML-1M declined after adding item text, indicating input-space alignment interferes with textual semantics.

Limitations

• Evaluated only on Vicuna-7B as the LLM backbone.
• Evaluated only on two datasets (ML-1M and Amazon-Book) with AUC/UAUC metrics.
• Relies on a pre-trained CF model whose embeddings are used as input features.

Experimental setup

Datasets: Amazon-Book ML-1M WikiFact SocialIQA XSum

Benchmarks: Amazon-Book ML-1M

Metrics: AUC UAUC

Baselines: MF LightGCN SASRec ICL Prompt4NR TALLRec PersonPrompt CoLLM BinLLM

Builds on: VLoRA LoRA BLIP-2 Q-Former TALLRec CoLLM

Closest papers by meaning

• Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.90)
• LEARN: Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application (AAAI 2025 · similarity 0.89)
• Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.88)
• Can LLMs Outshine Conventional Recommenders? A Comparative Evaluation (NeurIPS 2025 · similarity 0.88)
• Customizing Language Models with Instance-wise LoRA for Sequential Recommendation (NeurIPS 2024 · similarity 0.88)
• Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.88)