MCJudgeBench - Response-Variant Inconsistency: leaderboard

Judge robustness to response variants on MCJudgeBench (2026), where an LLM judge sees an instruction, a candidate response and its explicit constraint list and labels each of 653 constraints from 141 multi-constraint instructions as satisfied, partly satisfied or not satisfied: the share of constraint verdicts that change on human-validated label-preserving rewrites of the candidate response (local paraphrase, structural reordering), unparseable outputs counted as changed. Percent; lower is better.

Metric: Procedural inconsistency rate under candidate-response variants with parse penalty (%, CIR dagger resp): share of constraint predictions that differ from the reference prediction on 201 human-validated label-preserving response perturbations (local paraphrase, structural reordering), 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. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3.5 4BNon-reasoning3.59#1241
2Gemini 2.5 Flash Lite3.81#964
3Gemini 3.1 Pro (Preview)4.02#99
4Claude Sonnet 4.6Thinking4.02#173
5Claude Haiku 4.55.29#583
6Llama 3.2 3B Instruct6.17#7241
7GPT-5.27.29#185

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=mcjudgebench-response-variant-inconsistency · How It Works · Data refreshed daily, snapshot 2026-11-21.