Why test relations?
Imagine a log recording both an alarm and a production stop. That alone does not establish that the alarm caused the stop. This illustrative example shows why checking events individually can miss an unsupported connection.
What the benchmark tests
BeyondAtoms is organized around structural errors: substituting a relation, injecting an unsupported link, attaching a relation to the wrong events, or omitting a required premise.
How CARVE checks
CARVE preserves relational structure, extracts claim-anchored evidence and aligns the two in both directions. It asks whether asserted links exist and whether required premises were dropped. The training-free method is evaluated for both accuracy and computational cost.
Implications for evaluation
A useful evaluation should distinguish correct events from supported connections. The benchmark supplies analyzable error categories; applying it to a new domain still requires checking evidence coverage and task distribution.
Paper and resources
BeyondAtoms: Benchmarking and Verifying Relational Factuality Beyond Atomic Facts
This overview introduces the work; consult the paper for complete methods and experimental settings.