Researchers propose FLEXRec to boost compact LLMs for recommendations
FLEXRec achieves SOTA accuracy with Qwen 3 1.7B and Llama 3.2 3B models...
A team of researchers from the University of Queensland, Macquarie University, and Beihang University has developed FLEXRec (Fusion of Layer-wise Exits for Sequential Recommendation), a novel framework designed to improve the performance of compact large language models (LLMs) in recommendation tasks without sacrificing efficiency. The approach addresses a key challenge in deploying LLMs for real-world applications: computational scalability. Traditional LLM-based recommenders often struggle with high computational costs, making them impractical for many use cases.
FLEXRec introduces a discriminative framework that enhances compact LLMs while retaining scalable full-corpus ranking capabilities. The method works by inserting prediction heads (exits) at multiple transformer layers and adaptively fusing their score distributions using an adaptive continuous router (AC-Router). This dynamic selection process, regulated by a novel target-k hinge loss, ensures optimal performance by selecting both the number and identity of exits for each user sequence. Experiments on three real-world datasets with Qwen 3 1.7B and Llama 3.2 3B models demonstrate that FLEXRec achieves state-of-the-art accuracy among compact-backbone methods while remaining highly efficient.
- FLEXRec inserts prediction heads at multiple transformer layers and dynamically fuses their outputs using an AC-Router
- Achieves SOTA accuracy on three real-world datasets with Qwen 3 1.7B and Llama 3.2 3B models
- Dynamic selection of exits optimizes performance for each user sequence, balancing accuracy and efficiency
Why It Matters
Enables scalable, accurate recommendation systems using compact LLMs, reducing computational costs while maintaining performance for real-world deployments.