GraphLoRA embeds graph networks in LoRA for better LLM recommendations
New method merges collaborative signals with LLM tuning, beating SOTA on multiple benchmarks.
Large Language Models (LLMs) have shown promise for recommendation systems, but aligning their textual understanding with user-item collaborative signals remains difficult. Existing approaches either convert collaborative data into text prompts or inject pre-trained embeddings, both of which fail to capture high-order relational dependencies. To address this, researchers from Anhui University, Hefei University of Technology, and Shenzhen University propose GraphLoRA, a structure-aware low-rank adaptation method.
GraphLoRA embeds a trainable graph message-passing network within the low-rank adaptation (LoRA) pathway. This allows structural signals like user-item interactions to actively guide parameter updates during fine-tuning, rather than being treated as static input. The method enables deep integration between graph topology and textual semantics. Experiments on multiple recommendation benchmarks show that GraphLoRA outperforms state-of-the-art LLM-based recommendation methods, achieving superior generalization while maintaining computational efficiency. The code is publicly available. The paper has been accepted as a findings paper at ACL 2026.
- GraphLoRA embeds a trainable graph message-passing network inside the LoRA adaptation pathway for LLMs.
- It enables collaborative topology to actively guide parameter updates, capturing high-order relational dependencies.
- Outperforms state-of-the-art LLMRec methods across multiple benchmarks, with code open-sourced and accepted at ACL 2026.
Why It Matters
Better LLM-based recommendations that understand both user behavior semantics and collaborative signals for more accurate personalization.