Hybrid Intrinsic Reward Models for Efficient and Diverse Skill Internalization

TL;DR: What if we combined multiple types of intrinsic rewards (e.g., curiosity, novelty, empowerment) during RL-based skill internalization, allowing the agent to balance exploration and exploitation more efficiently?

Research Question: Does the use of hybrid intrinsic rewards, as in the HIRE framework, improve efficiency, diversity, and robustness of skill internalization compared to extrinsic-reward-only approaches like SKILL0?

Hypothesis: Skill internalization guided by hybrid intrinsic rewards will yield agents with richer exploratory behaviors and more generalized skill representations, especially in sparse-reward or open-ended environments.

Experiment Plan: Develop a hybrid intrinsic reward scheme that fuses curiosity, novelty, and goal-reaching signals during RL skill internalization (cf. Yuan et al., 2025). Compare efficiency (sample complexity), skill diversity, and zero-shot performance to SKILL0 on challenging multi-task benchmarks. Also analyze transfer to open-ended or creative problem-solving settings. Expected: Hybrid rewards accelerate learning and foster more versatile internal skills.

References:

• Yuan, M., Li, B., Jin, X., & Zeng, W. (2025). Deep Reinforcement Learning with Hybrid Intrinsic Reward Model. arXiv.org.
• Lu, Z., Yao, Z., Wu, J., Han, C., Gu, Q., Cai, X., Lu, W., Xiao, J., Zhuang, Y., & Shen, Y. (2026). SKILL0: In-Context Agentic Reinforcement Learning for Skill Internalization.