LLM judges change their verdicts when prompts are reworded, new benchmark reveals

When large language models act as judges to evaluate the outputs of other AI systems, do they render consistent decisions? A new benchmark called JudgeSense suggests the answer is troubling: reworded prompts that ask the same underlying question often yield different verdicts.

The Sensitivity Problem#

Researchers tested 25 different judge models from six major providers across four evaluation tasks, using a dataset of 880 items drawn from human-labeled corpora. Each item was evaluated under two differently worded instructions designed to ask the same question while varying the language. The results showed that rewording consistently degraded agreement—meaning the judges disagreed with themselves simply because the phrasing changed, not because the core request differed.

On two of the four tasks tested, this wording-driven inconsistency crossed what the researchers defined as a practically meaningful threshold for unreliability. The ordinal ranking task proved especially vulnerable, exhibiting both the lowest stability and the lowest accuracy across the tested judge models.

Key Findings#

The benchmark revealed several important patterns:

• Wording matters more than expected: Changing prompt language while keeping the meaning identical caused evaluators to reverse course.
• Scale doesn't guarantee consistency: Within a single model family, parameter count did not predict how stable a judge would be under rewording.
• Harness integration introduces additional noise: When judges ran inside an agent framework rather than through direct API calls, their agreement with themselves degraded faster than their cross-wording agreement, producing underestimated performance metrics.

Why This Matters#

As organizations increasingly rely on LLMs to evaluate AI system outputs—whether for benchmarking, competition judging, or quality assurance—this sensitivity to prompt wording undermines confidence in those evaluations. The benchmark enables researchers to examine how any judge model performs on the same test, providing a standard way to measure vulnerability to this form of instability.

The work was submitted to arXiv in April 2026 and revised through September of that year, with complete decision logs and the JudgeSense benchmark made available for external validation.