Researchers have proposed a new framework for evaluating the effectiveness of verifiers used with large language models (LLMs). The Verification Autonomy Levels (VAL) system categorizes verification schemes based on where the verification specification comes from and what guarantees the verdict provides. VAL ranges from L0, where the model self-declares its correctness, to L5, which is impossible in the unrestricted case. The authors argue that completeness can only be achieved for formally specifiable properties, while empirical open-world verification caps at anchored correctness (L2). They demonstrate the framework across four domains and a formal-verification baseline. --- Why it matters: This matters to researchers working with large language models because it provides a standardized way to evaluate the effectiveness of verifiers. By categorizing verification schemes based on their autonomy levels, researchers can better understand the strengths and limitations of different approaches and make more informed decisions about how to use them in practice. Source: https://arxiv.org/abs/2608.19009