CLIPR Framework Learns Transferable User Preferences for LLMs
A new framework called CLIPR (Conversational Learning for Inferring Preferences and Reasoning) enables large language models to infer latent user preferences from limited interactions and transfer them across tasks. The approach addresses a key limitation of LLMs in human-aligned decision making, where ambiguous situations require understanding unstated preferences. CLIPR learns actionable natural language rules that represent these preferences, allowing downstream reasoning to align with user intent without extensive repeated interactions. The framework is designed for high-level reasoning modules in LLMs, improving their ability to resolve ambiguities in a human-aligned manner. The paper is available on arXiv under identifier 2605.12682.
Key facts
- CLIPR stands for Conversational Learning for Inferring Preferences and Reasoning
- The framework learns transferable natural language rules representing latent user preferences
- It addresses LLM limitations in human-aligned decision making
- Requires only limited user interactions to infer preferences
- Preferences are actionable and transferable across tasks and contexts
- Designed for high-level reasoning modules in LLMs
- Paper published on arXiv with ID 2605.12682
- Focuses on resolving ambiguous situations in decision making
Entities
Institutions
- arXiv