LLM Judges Reverse 85% of Verdicts When Given Reference Answers

LLM judges reverse up to 85% of their verdicts on open-ended responses when a reference answer is added to the prompt, as per research from the University of Potsdam and Canada's National Research Council. This indicates that reference-free evaluation pipelines are systematically inflating correctness scores.

In a two-stage calibration and sensitivity study across three languages, Kranti and Vajjala defined three evaluation conditions: NR (no reference), RV (reference visible), and RC (reference comparison). In the no-reference condition, the judge model—relying solely on its parametric knowledge—often marks both correct and incorrect candidate answers as correct. Introducing the reference answer into the prompt, either for visibility or explicit comparison, prompts the judge to flip the incorrect response to its proper failing grade. The added reference makes the judge stricter, and a comparison against human annotations shows that stricter direction generally aligns with human judgment. Without the anchor, the judge over-credits wrong answers.

The finding adds to a growing inventory of judge-specific failure modes. A separate study from Ant Group on scoring bias in LLM-as-a-judge tests multiple state-of-the-art judge models, identifying rubric order bias, score ID bias, and reference answer score bias as prompt-level distortions that shift verdicts even when the candidate response is unchanged. Dongryeol Lee et al. (arXiv:2601.07506) compound the concern by showing that judges may ignore a provided reference entirely when it conflicts with their parametric knowledge. These papers indicate that the judge is not a neutral arbiter but a model with its own priors, prompt sensitivities, and knowledge gaps.

Neither study provides production deployment evidence—there are no reported inference costs, latencies, or throughput benchmarks for judge-based evaluation pipelines at scale. The operational numbers that matter here are reliability metrics: the 85% flip rate in the primary study, and comparable ~35% pairwise preference flips noted in related work by Tripathi et al. 2025—generator-embedded distractor features such as assertiveness or sycophancy cause pairwise judges to flip preferences while pointwise judges flip only ~9%. For an architect running nightly evals on a product's open-ended outputs, this means a no-reference judge could be reporting correctness rates that collapse once a reference is introduced, and the bias is consistent enough to look like a trend rather than noise. A practitioner summary on Tel Channel makes the point directly: teams optimize against these biased scores, improving metrics that do not reflect actual quality, while treating the drift as real performance gain.

What makes this particularly costly is that reference-free judging is often adopted precisely to avoid the latency and expense of human annotation or the engineering overhead of code-based evals. The paper's proposed mitigation is a pre-deployment calibration experiment: test the judge on a held-out sample that includes reference answers to discover which tasks, languages, or domains the judge can assess without an anchor. This requires maintaining a reference bank, reintroducing the annotation bottleneck the judge was meant to replace. Architects should also audit prompt artifacts—rubric ordering, score labels, and reference placement—because the Ant Group work shows these routinely move scores across all tested models.

There is no production evidence yet for automated mitigations, and the open question is whether any single judge model can be trusted as a reference-free oracle across multilingual, multi-domain workloads without continuous human calibration.

An architect should treat every no-reference LLM judge score as an uncalibrated upper bound and force it to prove its strictness against a held-out reference set before it is allowed to gate a release.