arXiv:2511.12784v3 Announce Type: replace Abstract: Large Language Models (LLMs) have recently emerged as powerful tools for autoformalization. Despite their impressive performance, these models can still struggle to produce grounded and verifiable formalizations. Recent work in text-to-SQL, has revealed that LLMs can be sensitive to paraphrased natural language (NL) inputs, even when high degrees of semantic fidelity are preserved. In this paper, we investigate this claim in the autoformalization domain. Specifically, we evaluate the robustness of LLMs generating formal proofs with semantical
The proliferation of LLMs in formal reasoning necessitates robust evaluation methods as autoformalization applications become more common.
This research provides a critical lens on the reliability and consistency of LLMs for generating formal proofs, which is crucial for high-stakes applications.
The understanding of LLM limitations in autoformalization is deepened, guiding future development towards more semantically stable and robust AI systems.
- · AI researchers focusing on formal verification
- · Developers of more robust LLM architectures
- · Industries requiring high-assurance AI outputs
- · Developers of brittle LLM autoformalization tools
- · Users relying on unverified LLM output for formal proofs
Evaluation frameworks for LLM robustness will incorporate semantic faithfulness metrics, leading to improved testing methodologies.
This could drive demand for LLMs specifically designed with enhanced semantic stability and explainability to mitigate paraphrasing sensitivity.
Increased skepticism about the 'general intelligence' of LLMs might emerge when faced with seemingly trivial semantic perturbations impacting core functionalities like autoformalization.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL