AgentProofBench Benchmarks Reusable Verifiers for Web Agents
AgentProofBench links paired tests of frozen acceptance rules to web-agent executions from identical saved openings, evaluating reusable verification at both requirement and policy levels. The benchmark covers 97 tasks across 79 families and four web services, with instructions, states, programs, reference rules, and trajectories released. Experiments using DeepSeek Flash and DeepSeek Pro report that deployment-state panel correctness distinguishes program assignments with error-free panels from those containing panel errors.
AgentProofBench provides a concrete benchmark and released artifacts for testing whether reusable acceptance rules remain valid when applied to later web-agent executions.
Computer Science > Multiagent Systems
arXiv:2610.34634 (cs)
[Submitted on 25 Oct 2026]
Title:AgentProofBench: Benchmarking Reusable Verifiers for Web Agents
Authors:Jingjie Ning, Xueqi Li, Yi Lu, Yibo Kong
View a PDF of the paper titled AgentProofBench: Benchmarking Reusable Verifiers for Web Agents, by Jingjie Ning and 3 other authors