MCJudgeBench - Prompt-Variant Inconsistency: leaderboard
Metric: Procedural inconsistency rate under evaluation-prompt variants with parse penalty (%, CIR dagger prompt): share of constraint predictions that differ from the reference prediction when the constraint list is reordered or reformatted or the prompt sections are reordered, unparsable outputs counted as inconsistent, temperature 0; the judge sees an instruction, a candidate response and its explicit constraint list and labels each constraint yes, partial or no, over the 653 constraints of 141 multi-constraint instructions of MCJudgeBench; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 2.91 |
| 2 | Gemini 2.5 Flash Lite | 6.13 |
| 3 | Claude Sonnet 4.6 (Thinking) | 6.18 |
| 4 | Qwen 3.5 4B (Non-reasoning) | 6.28 |
| 5 | Claude Haiku 4.5 | 6.33 |
| 6 | Claude Sonnet 4.6 | 6.74 |
| 7 | Claude Haiku 4.5 (Thinking) | 7.15 |
| 8 | Gemini 2.5 Flash Lite (Thinking) | 7.43 |
| 9 | GPT-5.2 | 8.63 |
| 10 | GPT-5.2 (Non-reasoning) | 9.04 |
| 11 | Llama 3.2 3B Instruct | 27.58 |
Interactive version: theaggregate.ai/benchmark?slug=mcjudgebench-prompt-variant-inconsistency · How It Works · Data refreshed daily, snapshot 2026-10-07.