Tag
This paper introduces a signal-coverage matrix that decomposes type-correctness gains in autoformalization into four strata, revealing the mechanisms behind LLM refinements and showing that headline metrics can obscure which errors are actually resolved.