Agent-as-a-Verifier: From Scores to Evidence for General Verification

Verification is increasingly critical across various inference and learning paradigms, such as test-time scaling, recursive intelligence and automated research, where feedback can improve efficiency and optimization. Common score-based verifiers, or discriminative reward models, predict a scalar reward for an output. While useful for filtering, rewarding, and ranking, such scores provide no explicit, auditable evidentiary basis for the judgment. In parallel, although existing generative process reward models offer text feedback, most of design do not fully exploit agentic capabilities, such as tool use and iterative reasoning, which are essential for fact-checking and fine-grained claim verification. In this paper, we address these limitations by introducing Agent-as-a-Verifier, an agentic, certificate-producing verification system that places role-separated LLM agents inside a formal-methods-inspired claim-obligation-certificate protocol. Given a generation candidate, Agent-as-a-Verifier produces a certificate by extracting a verdict-blind dependency graph of claims, independently compiling the task's acceptance contract, and deriving checkable obligation. On Terminal-Bench 2.1, Agent-as-a-Verifier achieves higher accuracy than the state-of-the-art approach. More importantly, Agent-as-a-Verifier is able to identify reward-hacking solutions that deceive the benchmark evaluator, highlighting its broader potential for iterative and recursive agent optimization.