WeAct: Dynamic Weight Switching as Agent Actions
arXiv cs.AI · 20 Oct 2026, 03:59 AEST
WeAct introduces a method for dynamic weight switching in agents to enhance their capabilities without degrading previously learned skills. This approach allows for modular continual adaptation while maintaining a frozen backbone and improving accuracy across various benchmarks.
Key facts
- Long-horizon agents must learn heterogeneous capabilities from interleaved trajectories.
- We introduce WeAct (Weight-level Internal Actions), which trains independent capability-specific LoRA adapters over a frozen backbone and adds adapter selection to the agent's action space.
- Across eight deep-research benchmarks, Qwen3-32B with WeAct improves average accuracy by 6.2% relative to full fine-tuning counterparts despite using less compute.
- WeAct also supports modular continual adaptation, by adding a new code computation adapter to expand the agent's capabilities and updating the existing writing adapter to improve report quality.
Read aloud
WeAct introduces a method for dynamic weight switching in agents to enhance their capabilities without degrading previously learned skills. This approach allows for modular continual adaptation while maintaining a frozen backbone and improving accuracy across various benchmarks. Long-horizon agents must learn heterogeneous capabilities from interleaved trajectories. We introduce WeAct Weight-level Internal Actions, which trains independent capability-specific LoRA adapters over a frozen backbone and adds adapter selection to the agent's action space. Across eight deep-research benchmarks, Qwen3-32B with WeAct improves average accuracy by six point two percent relative to full fine-tuning counterparts despite using less compute. WeAct also supports modular continual adaptation, by adding a new code computation adapter to expand the agent's capabilities and updating the existing writing adapter to improve report quality.
Sources in this cluster
- WeAct: Dynamic Weight Switching as Agent Actions
arXiv cs.AI · 20 Oct 2026, 03:59 AEST