Curiosity Reward for Personalized Dialogue — NeurIPS 2025
Source: https://arxiv.org/pdf/2504.03206 (via tweet) Author: unknown Date: 2025
Summary
Paper accepted to NeurIPS 2025: “Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward.” Uses curiosity as an intrinsic reward signal to drive personalized multi-turn conversations, avoiding the sycophancy trap of pure preference optimization.
Key Claims
- Curiosity as intrinsic reward for multi-turn personalized dialogue
- Avoids sycophancy: curiosity pushes toward exploring user preferences, not just confirming them
- NeurIPS 2025 acceptance validates approach
- Multi-turn design: personalization requires sustained engagement
Concepts
- Sycophancy — curiosity reward as alternative to preference collapse
- RL Infrastructure — intrinsic reward for personalization