Adapting a legal agent without fine-tuning

The paper studies how agents adapt to long legal proceedings at test time, without updating the model’s parameters. Two mechanisms are proposed: memory evolution, which extracts experience from previous cases, adjusts it to the factual and procedural context of the current case and accumulates it for subsequent decisions, and rubric alignment, which checks and corrects the actions of each role against behavioral and procedural requirements, linking roles and stages of the proceedings. Experiments on the J1-EVAL and LegalWorld datasets with five base models showed gains over reasoning approaches and agentic baselines at moderate interaction and computation costs. Ablations show that the two mechanisms complement each other: one is responsible for transferring experience, the other for coordinating roles.