A new paper on arXiv (2610.26932) tackles a growing problem: how to train large language model agents in non-verifiable domains, such as financial services. In these settings, there is no simple right-or-wrong answer, yet agents must still interpret user intent, rewrite queries, and retrieve relevant information. The authors argue that conventional reinforcement learning struggles here because reward signals are too sparse or ill-defined.
The paper's proposal is to use expert-guided rubrics as dense reward signals. Instead of relying on a single final outcome, a rubric breaks down what good behavior looks like into multiple criteria. This gives the model more frequent and more informative feedback during training.
The approach is notable because it moves beyond verifiable tasks like code execution or math problems, where rewards can be computed automatically. By leveraging human expertise in the form of rubrics, the method aims to make RL practical for domains where quality is subjective but still assessable. The abstract does not provide experimental results, so the paper's claims are presented as a proposed method rather than a demonstrated breakthrough.