MDSKILL internalizes skill memories via multi-view distillation for RL agents
Original title: MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution
MDSKILL improves ALFWorld and WebShop performance by up to 8.1% without external teachers.
Skill memories in RL often contain noise and redundancy. MDSKILL organizes them into state grounding, action planning, and execution control views. It selects high-quality skills dynamically and uses OPSD to consolidate knowledge across branches. This enables efficient agent self-evolution on benchmarks.
- Improves SOTA by up to 7.1% on ALFWorld and 8.1% on WebShop
- Organizes skill memories into three complementary functional views
- Uses dynamic metrics for high-quality skill selection
- No additional computational overhead
- Code available at anonymous repository
From the original article
1 min read at arxiv.orgSkill memories encode past experiences as textual prompts to improve reinforcement learning (RL) for self-evolving agents. Complementing RL, On-Policy Self-Distillation (OPSD) provides dense, fine-grained supervision to internalize knowledge from skill-guided trajectories and accelerate policy learning.
By Dian Jin · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.