FaithAlign
The first theory-level benchmark for the faithfulness of autoformalization. It employs a context-aware, phrase-level align-then-relabel mechanism to assess how well formal statements preserve the semantics of informal mathematical propositions.
Timeline (last 90 days)
The paper proposes FaithAlign, claimed to be the first theory-scale faithfulness benchmark, containing 460 entries and 4,982 phrase-level annotated links.
In the evaluation of 8 models, GPT-6 Astra achieved the highest defect F1, at 47.8%.
The value of FaithAlign lies in turning "context-aware fine-grained faithfulness evaluation" into a quantifiable, reproducible testbed.
All Facts (3)
In the evaluation of 8 models, GPT-6 Astra achieved the highest defect F1, at 47.8%.
50%UnverifiedThe value of FaithAlign lies in turning "context-aware fine-grained faithfulness evaluation" into a quantifiable, reproducible testbed.
50%UnverifiedThe paper proposes FaithAlign, claimed to be the first theory-scale faithfulness benchmark, containing 460 entries and 4,982 phrase-level annotated links.
50%