A study published on October 24, 2026, on the arXiv platform (ID 2610.30512v1) points out that when large language models (LLMs) are used as automated judges to compare the quality of two responses, their individual judgments are highly unstable and extremely sensitive to instruction details, a instability often masked by seemingly stable aggregated scores.
The study was based on 500 Arena-Hard v2.0 prompts, with the responses of two tested models held constant. Using GPT-4.1 at a temperature of 0 and with the model version locked, the system tested all eight instruction combinations—covering evaluation criteria, output formats, and task phrasing. The results showed that simply repeating the default settings caused the aggregate score to fluctuate by about one point, yet 21.2% of prompt-level judgments (win/draw/loss) changed.
When the instruction is changed, instability intensifies: a minimalist instruction setup (omitting reasoning requirements and other clarifications) leads to an additional 28.2 percentage points of variation in decision outcomes compared to the average re-run fluctuations, while the aggregate score drops by 4.09 points; another setup that prioritizes accuracy, though resulting in only a 6.7-percentage-point increase in decision variability, leaves the aggregate score virtually unchanged. This demonstrates that stable aggregate scores do not necessarily equate to reliable individual decisions.
The paper emphasizes that if such judgments are used as preference labels or for evaluating the quality of individual responses, it is essential to conduct repeated runs for each evaluation setting and report both aggregate scores and prompt-level judgment consistency; otherwise, the evaluation conclusions may be subject to systematic bias. All conclusions are based on the experimental setup and have not been generalized to other models or datasets.