MCJudgeBench - Intrinsic Inconsistency: leaderboard

Judge self-consistency 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 constraints whose verdict changes across five repeated judgments of the same instance at temperature 1.0. Percent; lower is better.

Metric: Intrinsic constraint-level inconsistency rate (%, CIR intr): share of constraints whose predicted label is not identical across 5 repeated judgments of the same instance at temperature 1.0; 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
1Claude Haiku 4.52.94#583
2Claude Sonnet 4.6Thinking5.21#173
3Gemini 3.1 Pro (Preview)6.58#99
4Gemini 2.5 Flash Lite7.35#964
5GPT-5.2Non-reasoning9.49#185
6Qwen 3.5 4BNon-reasoning24.2#1241
7Llama 3.2 3B Instruct35.38#7241

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

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