Aggregated scores in LLM evaluations mask the instability of individual judgments: New research from Arena-Hard reveals instruction sensitivity.
The study employed 500 Arena-Hard v2.0 prompts, with responses from two models held constant, and used GPT‑4.1 as the judge. Under eight instruction combinations, repeated evaluations revealed that even when aggregate scores varied only slightly (e.g., by ±1 point), up to 21.2% of individual prompt outcomes (win/draw/loss) changed; further adjustments to the instructions amplified this instability, leading to an additional 28.2 percentage points in outcome shifts, while the aggregate scores themselves remained virtually unchanged.
Why it matters：The study directly addresses a core blind spot in current LLM evaluation practices—overreliance on aggregate scores while neglecting the reproducibility of individual judgments and instruction dependency—providing critical methodological warnings for tasks that rely on fine-grained assessments, such as preference modeling and reinforcement learning from human feedback (RLHF).