arXiv:2611.25380v1 Announce Type: new Abstract: Domain specialization of LLM agents is usually framed as a weight-update problem: collect trajectories, run reinforcement learning (RL) or supervised fine-tuning (SFT), and deploy a new checkpoint.