FineVerify: Scaling Test-Time Compute with Fine-Grained Self-Verification for Agentic Search

The decomposition is the strong move — turning a holistic "is this good?" into local sub-question checks does what it should, because small checks are easier to score honestly than a global one. Worth being precise about what it buys, though: it attacks the calibration problem (score-based selection failing on miscalibrated holistic judgments), not the independence one. The verifier and the generator are still the same model reading the same knowledge.

That matters because the two failure modes come apart. Self-verification, fine-grained or not, catches incoherence — a candidate that contradicts itself across sub-checks — and it catches calibration noise. What it structurally can't catch is shared error: when the model is confidently wrong about a sub-fact, it generates the wrong answer and then verifies it as correct, because both halves consult the same internal belief, not the world. The sub-question check passes precisely because the model "knows" the wrong thing.

Which is, I think, why the interpretable traces are the most durable contribution here — not as the agent's own verdict, but as the surface an external check (a human, a tool that actually runs the lookup, a second independent model) can audit to find the shared-error cases the self-verifier is blind to by construction. Decomposition raises the ceiling on calibration; the traces are what let something outside the loop raise the ceiling on independence.