MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution
arXiv cs.CL · 4 Nov 2026, 04:19 AEST
MDSKILL introduces a multi-view distillation approach for skill memory internalization in reinforcement learning agents. This method enhances policy learning and agent self-evolution by organizing skill memories into complementary functional views.
Key facts
- Skill memories encode past experiences as textual prompts to improve reinforcement learning (RL) for self-evolving agents.
- MDSKILL organizes skill memories into three complementary functional views (state grounding, action planning, and execution control) and selects high-quality skills using dynamic metrics.
- Without additional computational overhead, MDSKILL improves upon SOTA performance by up to 7.1% and 8.1% on ALFWorld and WebShop.
Read aloud
MDSKILL introduces a multi-view distillation approach for skill memory internalization in reinforcement learning agents. This method enhances policy learning and agent self-evolution by organizing skill memories into complementary functional views. Skill memories encode past experiences as textual prompts to improve reinforcement learning RL for self-evolving agents. MDSKILL organizes skill memories into three complementary functional views state grounding, action planning, and execution control and selects high-quality skills using dynamic metrics. Without additional computational overhead, MDSKILL improves upon SOTA performance by up to seven point one percent and eight point one percent on ALFWorld and WebShop.
Sources in this cluster
- MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution
arXiv cs.CL · 4 Nov 2026, 04:19 AEST