Grading the Graders: Verification Autonomy Levels (L0-L5) for LLM Reasoning
The paper says most LLM “verifiers” can check a proposed answer, but cannot prove the model found every valid one.
Yajie Yin proposes Verification Autonomy Levels, or VAL, to classify verification systems by where their checking standard comes from and what their verdict actually guarantees. The scale runs from L0 self-declaration to L3/L4 decidable formal systems, while L5 is described as impossible for unrestricted reasoning. The paper argues open-world checks such as fact-checking, diagnosis, and many sampling-based methods top out at anchored correctness, not completeness. It says prior work has conflated VAL with granularity, abstraction, risk, and system-stack layers across 17 surveyed papers. ArXiv · AI/CL/LG's note
Yajie Yin proposes Verification Autonomy Levels, or VAL, to classify verification systems by where their checking standard comes from and what their verdict actually guarantees. The scale runs from L0 self-declaration to L3/L4 decidable formal systems, while L5 is described as impossible for unrestricted reasoning. The paper argues open-world checks such as fact-checking, diagnosis, and many sampling-based methods top out at anchored correctness, not completeness. It says prior work has conflated VAL with granularity, abstraction, risk, and system-stack layers across 17 surveyed papers. ArXiv · AI/CL/LG's note
score 4