Position Bias (Lechmazur) — leaderboard
Tests whether LLM judges preserve the same underlying preference when two lightly edited stories are shown in opposite orders. Lower order-flip rates indicate more stable pairwise judgment.
Metric: Order Flip % (lower is better). Source: github.com. Status: saturation imminent. 36 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | MiMo-V2-Pro | 19.8 |
| 2 | Seed 2.0 Pro | 28 |
| 3 | Gemini 3.5 Flash | 29.8 |
| 4 | Claude Opus 4.6 (Thinking, High) | 30.2 |
| 5 | DeepSeek V3.2 | 30.3 |
| 6 | GLM-5.1 | 31.5 |
| 7 | Qwen 3.6 Plus | 33.8 |
| 8 | Qwen 3.7 Max | 34.8 |
| 9 | MiniMax-M3 | 34.9 |
| 10 | Gemini 3.1 Pro (Preview) | 35.4 |
| 11 | MiniMax-M2.7 | 36.5 |
| 12 | Arcee Trinity Large (Thinking) | 36.6 |
| 13 | Claude Sonnet 4.6 (Thinking, High) | 37.4 |
| 14 | Claude Opus 4.6 (Non-reasoning) | 38.1 |
| 15 | Grok 4.20 0309 (Reasoning) | 39 |
Interactive version: theaggregate.ai/benchmark?slug=position-bias-lechmazur · How the rankings work · Data refreshed daily, snapshot 2026-07-22.