ProMax: Exploring the Potential of LLM-derived Profiles with Distribution Shaping for Recommender Systems
New framework cuts semantic loss, boosting recommendations by 15% on benchmarks...
Deep Dive
Researchers introduced ProMax, a framework that uses LLM-derived user profiles to enhance recommender systems. By applying dense retrieval and dual distribution reshaping, ProMax guides models to learn preferences for unseen items without semantic loss. Tested on three public datasets with four classic methods, it outperformed existing LLM-based approaches. Accepted at SIGIR 2026.
Key Points
- ProMax uses LLM-derived profiles with dense retrieval to reduce semantic loss in recommendations.
- Dual distribution reshaping guides models to learn preferences for unseen items, improving recall by up to 12%.
- Tested on three public datasets with four classic methods, outperforming existing LLM-based approaches at SIGIR 2026.
Why It Matters
ProMax offers a lightweight, plug-and-play upgrade for recommenders, boosting accuracy without retraining large models.