DiscoverLLM: Training LLMs to Help Users Form and Discover Intents
DiscoverLLM is an innovative framework designed to train large language models to help users identify their own intentions, rather than just following direct commands. This method tackles vague and open-ended inquiries where users may not have established specific objectives. At the heart of DiscoverLLM lies a user simulator that represents cognitive states through a hierarchy of intents, which become clearer as the model offers pertinent choices. The level of clarity acts as a reward signal for optimization. Consequently, the models are trained to work together by flexibly exploring different options. The research paper can be found on arXiv with the ID 2602.03429.
Key facts
- DiscoverLLM is a framework for training LLMs to help users discover intents.
- It addresses ambiguous and open-ended requests where users have not formed intents.
- The approach uses a user simulator with a hierarchy of intents.
- Intents progressively concretize as the model surfaces relevant options.
- The degree of concretization serves as a reward signal for training.
- Models learn to collaborate by adaptively diverging to explore options.
- The paper is on arXiv with ID 2602.03429.
Entities
Institutions
- arXiv