Syntactically robust NLI for semantics of imperfectly generated text? [R]

Syntactically robust NLI for semantics of imperfectly generated text? [R]

Hi all, I'm looking for literature on relatively specific tooling.

In autoregressive LLMs, there is substantial published work that used NLI on sub-claims produced by LLMs to gauge correctness of LLM answers.

In diffusion (or D-) LLMs, the SoTA model generations that I see (outside of perhaps LLaDA) seem to struggle to be as correct syntactically as the generations from premier AR LLMs, in addition to the issue of semantic correctness.

My intuition is that this complicates the usage of NLI (the syntactic noise).

What is the SoTA on syntax-robust NLI?