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. Saturation forecast: Not forecast. 36 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiMo-V2-Pro | 19.8 | #168 |
| 2 | Seed 2.0 Pro | 28 | #118 |
| 3 | Gemini 3.5 Flash | 29.8 | #60 |
| 4 | Claude Opus 4.6 (Thinking, High) | 30.2 | #63 (Claude Opus 4.6) |
| 5 | DeepSeek V3.2 | 30.3 | #205 |
| 6 | GLM-5.1 | 31.5 | #97 |
| 7 | Qwen 3.6 Plus | 33.8 | #120 |
| 8 | Qwen 3.7 Max | 34.8 | #75 |
| 9 | MiniMax-M3 | 34.9 | #136 |
| 10 | Gemini 3.1 Pro (Preview) | 35.4 | #58 |
| 11 | MiniMax-M2.7 | 36.5 | #258 |
| 12 | Trinity-Large-Thinking | 36.6 | #356 |
| 13 | Claude Opus 4.6 (Non-reasoning) | 38.1 | #63 (Claude Opus 4.6) |
| 14 | Grok 4.20 0309 (Reasoning) | 39 | #112 (Grok 4.20 0309) |
| 15 | DeepSeek V4 Pro | 43.6 | #102 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=position-bias-lechmazur · How It Works · Data refreshed daily, snapshot 2026-10-20.