Google Research recently explored how large language models (LLMs) can enhance scientific research, particularly in cryptography. A significant case study revolved around how LLMs identified a bug in a paper on SNARGs, written by Ziyi Guan and Eylon Yogev. Their work, titled "SNARGs for NP from LWE," initially presented a breakthrough in constructing SNARGs based on the learning with errors (LWE) assumption. This was noteworthy because prior methods relied on non-falsifiable assumptions or idealized models. However, shortly after publication, the authors discovered a bug they couldn't fix, which led to a revised version of the paper that retracted the SNARG claim but retained some findings. The LLM's role in uncovering this error was facilitated by a structured prompting strategy labeled LLM-as-a-Judge. Instead of a simple verification request, the authors used a "rigorous iterative self-correction prompt," where the model reviewed the paper and then its own review. Details on the specific model and prompting text remain undisclosed. The Google Research report mentions that the LLM produced some irrelevant "noise" in its output, though it doesn't clarify if these included false positives. The implications of this technology extend beyond this case study. LLMs have been effectively used in academic contexts to formalize proofs and may soon flood conferences with AI-assisted submissions. There's potential for LLMs to streamline peer reviews, but challenges remain, such as the reliability of these methods across a broader sample of papers and the management of false-positive rates. At ZKSecurity, similar LLM methodologies are being applied in audits via a tool called zkao, which automates the review process and incorporates feedback loops for accuracy.