If the goal is to review every citation fully with 100% accuracy, then, sure, exhaustive human review is needed. But I suspect human review of a random sample would add value, catching some fraud, missing others, but having zero false positives (or as close to zero as human review can get).

An LLM could replace the random sampling. It doesn't need to be particularly good for the approach to provide value. I would worry about LLM bias though.

Another thing to consider is that readers can detect fake citations after publication, report to arXiv, and the author gets banned.