New Framework Launched to Enable Smarter Management of External AI Skills

Large Model Agents, also known as LLM Agents, are rapidly progressing beyond basic chatting functions toward the ability to make continuous decisions and complete real-world tasks. Yet, figuring out how to effectively manage an agent’s external capabilities has emerged as a new challenge for the entire industry. Recently, a research team from The Chinese University of Hong Kong introduced a dynamic skill lifecycle management framework named “SLIM” in their paper titled “Dynamic Skill Lifecycle Management for Agentic Reinforcement Learning.” This innovative approach moves away from the previous industry practice of simply “stacking skills” and offers a new way to tackle complex tasks in both physical and virtual environments.

In complex, long-tail scenarios such as web searching, automated office tasks, and operations involving embodied robots, agents often need to leverage external skills to handle steps that are prone to errors or require specialized handling. Traditional methods either keep adding skills indefinitely, which increases retrieval noise and context interference, or attempt “zero-skill reasoning” by trying to embed all necessary capabilities directly into the model parameters, resulting in the loss of locally important abilities. To address these issues, the SLIM framework treats external skills as a dynamic capability system with a defined lifecycle, allowing the model to autonomously decide whether to retain, remove, or expand these external skills during reinforcement learning training.

SLIM operates through a sophisticated closed-loop mechanism. During training, the system retrieves appropriate skills—whether general or task-specific—based on the current situation and updates the agent’s decision-making strategy using the GRPO algorithm. Afterward, it conducts a unique “leave-one-skill-out” audit: it temporarily disables a specific skill to assess its contribution to the agent’s performance. If performance drops significantly after disabling the skill, it is classified as “Retain”; if its contribution remains consistently low, it indicates that the model has already absorbed the associated ability or that the skill causes interference, in which case it is classified as “Retire.” For new scenarios where the agent keeps failing, the system uses the “Expand” mechanism to develop and incorporate new skills based on those failure cases.

Experimental results show that this framework outperforms existing best comparison methods by an average of 7.1 percentage points. In more action-oriented and complex ALFWorld home environment tasks, SLIM achieved an 87.5% success rate thanks to its efficient external skill management approach, far surpassing the 75.0% success rate of the baseline method SkillRL. It also demonstrated strong performance in information retrieval and reasoning-focused SearchQA tasks, further validating the feasibility of integrating certain search strategies directly into the model.

Industry analysts note that SLIM’s greatest value lies in transforming external skill libraries from static auxiliary tools into training elements that can be optimized in coordination with overall strategies. It not only helps determine technically which capabilities should be incorporated into the model and which should remain external but also enables large model agents to learn when to seek external assistance in dynamic and changing environments. This dynamic capability management approach undoubtedly provides a solid theoretical and practical foundation for the next phase of development in embodied intelligence and large model agents as they move toward broader industrial applications.