Swapping the Judge Changes the Ranking: A Preregistered Study of Cross-Provider Replication in LLM-Judged System Comparisons
Organizations: CEO Business School
Abstract
Practitioners who compare systems with an LLM judge and have no human labels often check a conclusion by repeating it with a judge from another provider. We test how far that check can be trusted. In a preregistered study (with a disclosed post-registration reduction of the judge set), five judge models from three providers re-judged 4,000 Chatbot Arena battles (nine systems, 33 system pairs) that carry human votes, each in both response orders. Agreement with the human vote was low at the verdict level (median Cohen's kappa = 0.22), and system rankings did not survive a change of judge: Kendall's tau with the human ranking ranged from 0.17 to 0.89 (median 0.67), so the preregistered hypothesis that rankings remain stable was not supported. Conclusions replicated by two low-cost judges from different providers matched the human conclusion more often than single-judge conclusions (+0.10 precision), but the gain came from filtering out one poorly aligned judge, and replication involving that judge left a conclusion for only 35% of comparisons. Judges that agreed with humans reached a stable conclusion with no more verdicts than human votes required; the others failed to converge on 26-70% of system pairs at the sample sizes tested (up to 100 verdicts per pair). Disagreements with the human vote co-occurred across judges far more often than independence predicts, although a single human vote per battle cannot separate shared judge bias from voter noise.
Figures & tables
| Judge | Provider | Order-consistent | Tie | Repeat | |||
|---|---|---|---|---|---|---|---|
| Opus 5.5 | Anthropic | 0.202 | 0.667 | 0.783 | 89.4% | 11.6% | 0.868 |
| Sonnet 5.5 | Anthropic | 0.217 | 0.667 | 0.783 | 87.6% | 13.3% | 0.845 |
| Haiku 4.5 | Anthropic | 0.266 | 0.889 | 0.967 | 83.0% | 17.5% | 0.942 |
| GPT-6 Luna | OpenAI | 0.052 | 0.167 | 0.217 | 78.5% | 22.1% | 0.807 |
| Gemini 3.1 Flash-Lite | 0.302 | 0.889 | 0.967 | 85.8% | 14.9% | 1.000 | |
| Majority of five | – | 0.267 | 0.778 | 0.900 | – | – | – |
| Source of conclusion | Precision | Coverage |
|---|---|---|
| Single judge | ||
| Haiku 4.5 | 22/23 (0.96) | 1.00 |
| GPT-6 Luna | 7/14 (0.50) | 0.61 |
| Gemini 3.1 Flash-Lite | 22/23 (0.96) | 1.00 |
| Replicated across providers | ||
| Haiku 4.5 + GPT-6 Luna | 7/8 (0.88) | 0.35 |
| Source | Censored | |||
|---|---|---|---|---|
| Human votes | 15 | 7 | 1 | 0 |
| Haiku 4.5 | 22 | 0 | 0 | 1 |
| Gemini 3.1 Flash-Lite | 22 | 0 | 0 | 1 |
| Sonnet 5.5 | 14 | 3 | 0 | 6 |
| Opus 5.5 | 14 | 1 | 0 | 8 |
| GPT-6 Luna | 3 | 3 | 1 | 16 |
| Judge | Anthropic | OpenAI | Mistral | Alibaba | Own minus others (95% CI) | |
|---|---|---|---|---|---|---|
| Opus 5.5 | +0.101 | +0.081 | -0.019 | -0.159 | -0.190 | +0.173 [0.141, 0.211] |
| Sonnet 5.5 | +0.072 | +0.091 | +0.002 | -0.154 | -0.169 | +0.129 [0.098, 0.163] |
| Haiku 4.5 | -0.104 | +0.126 | +0.099 | -0.066 | -0.032 | -0.136 [-0.167, -0.105] |
| GPT-6 Luna | +0.245 | -0.114 | -0.074 | -0.125 | -0.145 | -0.039 [-0.082, 0.003] |
| Gemini 3.1 Flash-Lite | -0.110 | +0.155 | +0.125 | -0.105 | -0.062 | +0.193 [0.161, 0.224] |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Stage | Battles remaining |
|---|---|
| Raw release | 135,634 |
| Duplicates removed | 135,619 |
| Two different systems | 135,609 |
| Non-empty responses | 135,472 |
| Single turn | 116,307 |
| First vote per conversation | 92,488 |
| Judge | Model ID | Temperature | Reasoning | Cap | Batch |
|---|---|---|---|---|---|
| Opus 5.5 | claude-opus-5-5 | Not sent (unsupported) | adaptive; effort=low | 5120 | Yes |
| Sonnet 5.5 | claude-sonnet-5-5 | Not sent (unsupported) | between_tools; effort=low; no tools | 1024 | Yes |
| Haiku 4.5 | claude-haiku-4-5-20251001 | 0 | disabled | 1024 | Yes |
| GPT-6 Luna | gpt-6-luna | Not sent (unsupported) | effort=none | 1024 | Yes |
| Gemini 3.1 Flash-Lite | gemini-3.1-flash-lite | 0 | thinkingLevel=minimal | 5120 | Yes |
LLM-as-a-Judge
Large language model (LLM)-as-a-judge research uses language models to score, compare, and verify AI outputs as scalable alternatives to human evaluation.
See all 604 papers