Large language models are increasingly used as judges, producing numeric scores that feed into downstream decisions such as reviewing papers and making recommendations. According to the preprint abstract, these LLM-as-a-Judge pipelines are vulnerable to adversarial manipulation, which creates a risk for any system that relies on their outputs.
The preprint introduces StableJudge, a lightweight adversarial defense framework designed to make judge outputs more robust. The abstract frames it as a defense layer for LLM-based evaluation, though it does not include implementation details or experimental results.
This summary is based solely on the arXiv abstract, so the specific attack types and evaluation outcomes are not covered here. The significance of the work is in addressing a growing weak point: as LLM judges take on more consequential tasks, defending them against adversarial inputs becomes increasingly important.