What the study found
The study found that user values, which are more stable than short-term interests, can be extracted from historical interactions and used to improve recommendation systems. The authors present ZOOM, a zero-shot multi-LLM collaborative framework, as an approach for doing this.
Why the authors say this matters
The authors suggest that adding user values to recommender systems may help make recommendations more stable and better aligned with users’ latent preferences, meaning underlying preferences that are not directly observed. They also argue that this addresses the difficulty and cost of collecting user values directly.
What the researchers tested
The researchers built ZOOM, a framework that uses large language models, or LLMs, to extract user values from historical interactions in a zero-shot setting, meaning without task-specific training examples. They used text summarization to shorten item content, then assigned two agent roles: evaluators to generate initial values and supervisors to refine them through debate. They also tested fusion methods, including direct concatenation and contrastive learning, to add the extracted values to recommendation models.
What worked and what didn't
The abstract reports that experiments on two recommendation datasets and two state-of-the-art recommendation models showed the framework was effective for automatic user value mining and improved recommendation performance. It also says the approach helped address the input-length problem from long histories and item content, and aimed to reduce hallucinations from LLMs through the evaluator-supervisor process.
What to keep in mind
The available summary does not provide detailed quantitative results, and it does not describe limitations beyond the two challenges the authors sought to address: long inputs and hallucinations. The abstract also does not specify which fusion method worked best.
Key points
- The paper argues that user values are more stable than short-term interests in recommendation settings.
- ZOOM uses large language models to extract user values from historical interactions without task-specific training examples.
- Text summarization is used to shorten item content before extraction.
- Evaluator and supervisor agents are used to refine value extraction through debate.
- Experiments on two datasets and two recommendation models reported improved recommendation performance.
Disclosure
- Research title:
- ZOOM improves user value mining for recommendations
- Authors:
- Lijian Chen, Yuan Wei, Tong Chen, Xiangyu Zhao, Quoc Viet Hung Nguyen, Hongzhi Yin
- Institutions:
- City University of Hong Kong, Griffith University, The University of Queensland, The University of Queensland, The University of Queensland, The University of Queensland
- Publication date:
- 2026-06-30
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with scholarly research without reading academic papers — one filtered digest, every Friday.