Source record

Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.34634.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

AgentProofBench evaluates reusable verifiers for web agents by pairing tests of frozen acceptance rules with execution from identical saved openings. Across 97 tasks, 79 families, and four web services, experiments with DeepSeek Flash and Pro compare decision replay, paired rebuilds, and deployment-state panel correctness, while releasing benchmark artifacts and trajectories.

Provides a reproducible test structure for separating verifier quality from downstream execution, with released tasks, rules, states, programs, and failure-linked trajectories.

Cited by

No citations on record.