Artificial intelligence systems that perform complex tasks over long periods often need to juggle multiple skills or capabilities. However, training these systems to learn new skills without forgetting old ones is a major challenge, known as “catastrophic forgetting.” A newly published research paper introduces an innovative approach called WeAct that helps AI agents dynamically switch between different internal skill sets during task execution, improving their accuracy and adaptability.
Key Takeaways
- WeAct trains separate, skill-specific modules (called LoRA adapters) on a fixed core AI model, avoiding interference between skills.
- The AI agent’s action space is expanded to include selecting which internal skill module to activate at each step, allowing dynamic switching.
- On eight benchmark tests, WeAct improved average accuracy by 6.2% compared to full fine-tuning and by 27.5% compared to static skill-merging methods.
- WeAct supports modular continual learning, enabling new skills to be added or existing ones improved without retraining the whole model.
Traditional AI training often involves fine-tuning all model parameters together, which can cause the system to overwrite prior knowledge when learning new tasks. This is especially problematic for long-horizon tasks where different stages require different capabilities. For example, an AI writing assistant might need creative language generation at one point and precise code understanding at another. Existing approaches typically use static combinations of skill modules, which cannot adapt dynamically as the task unfolds.
WeAct addresses this by freezing the core AI model (often called the “backbone”) and training multiple small, specialized modules known as LoRA adapters, each focusing on a specific capability. LoRA stands for Low-Rank Adaptation, a technique that efficiently fine-tunes models by adjusting a small number of parameters. Instead of merging these adapters into a fixed configuration, WeAct treats the selection of which adapter to activate as part of the agent’s own set of actions—internal “skill switches” it can choose at each step.
By integrating adapter selection into the agent’s native tool-calling interface, WeAct enables on-demand composition of capabilities within a shared context. This means the AI can dynamically switch between skills as needed, rather than relying on a one-size-fits-all parameter setup. The researchers tested this approach using Qwen3-32B, a large language model, across eight deep-research benchmarks. WeAct not only improved accuracy significantly but also required less computational power compared to fully fine-tuning the entire model.
Another important feature of WeAct is its support for modular continual adaptation. The researchers demonstrated that new skill adapters—such as one for code computation—can be added to the agent without changing the frozen backbone or existing adapters. Similarly, existing adapters can be updated independently to improve performance on specific tasks, like report writing. This modularity allows AI systems to evolve over time without the risk of erasing previously learned abilities.
While WeAct is a promising step toward more flexible and resilient AI agents, the approach is currently demonstrated on research benchmarks and large language models. Future work will explore how this dynamic skill-switching can be applied to a broader range of real-world applications, such as multi-task assistants, robotics, or adaptive tutoring systems. By enabling AI to manage its internal capabilities more intelligently, WeAct may help create systems that learn and adapt continuously without forgetting, a crucial milestone for long-term autonomous AI.
Based on research published on arXiv by Rui Shi, Xunjian Yin, Jiacheng Sang et al..