What's this about?
Language models are increasingly paired with verifiers — step checkers, self-consistency filters, tool-based fact checkers, formal proof assistants — that claim to catch the model's errors. This paper proposes verification autonomy levels, L0 to L5, for how much verification a system genuinely provides.
What exactly is the levels framing borrowed from?
It's the structure used for self-driving cars, where Level 0 is no automation and Level 5 is full autonomy in all conditions. Its value there was ending a marketing free-for-all: "self-driving" meant everything and nothing, and levels forced specific claims about what the system does and when a human must intervene. Verification has the same problem now. "The output is verified" can mean a second model was asked whether it looks right, or that a formal proof assistant machine-checked every step — and those are not remotely the same guarantee.
Why is this worth noticing?
- The variance in what "verified" means is enormous, and it's the gap between an agent you can leave alone and one you must watch.
- Verification is the load-bearing assumption in the current agent stack. StateM's fitness came from benchmark verifiers, DarwinX's from the same, RLVR depends on it — so how good the graders are determines how much of that work holds.
- It converges with this week's Authority Resolution Framework and PACE — all three are attempts to make agent claims checkable rather than asserted.