LLM Position Bias First-Option Lift
Chat & Writing · 2026-04-21
First-Option Lift measures first-shown pick rate minus the 50% symmetric baseline. Values closer to zero indicate less position bias; positive values indicate first-position bias and negative values indicate second-position bias.
Top models (closer to zero is better)
| Model | Score |
|---|---|
| Trinity Large Thinking | -1.1 |
| Doubao Seed 2.0 Pro | -1.9 |
| Gemini 3.1 Flash-Lite Preview | 2.5 |
| DeepSeek-V3.2 | 3.7 |
| MiMo-V2-Pro | 4.5 |
| Qwen3.5 122B A10B | 6.7 |
| MiniMax M2.7 | 8.2 |
| Claude Fable 5 | 8.9 |
| GLM-5.1 | 9.8 |
| Opus 4.8 | 10.7 |
| Gemini 3.5 Flash | 12.1 |
| Qwen3.6 Plus (2026-04-02) | 14.3 |
| Opus 4.6 | 14.4 |
| Sonnet 4.6 | 15.1 |
| Qwen3.7-Max | 15.3 |
| MiniMax M3 | 15.3 |
| Opus 4.7 | 15.6 |
| Grok 4.20 | 15.8 |
| Qwen3.5 397B A17B | 15.8 |
| Gemini 3.1 Pro Preview | 16.0 |
| ERNIE 5.0 | 16.1 |
| DeepSeek-V4-Pro | 16.7 |
| Mistral Medium 3.1 | 17.2 |
| GPT-5.5 | 19.9 |
| Kimi K2.6 | 22.5 |
| Mistral Large 3 675B Instruct 2512 | -22.6 |
| GPT-5.4 | 22.7 |
| Gemma 4 31B IT | 24.6 |
| GPT-5.4 Mini | 25.7 |
| Kimi K2.5 | 25.8 |
| Mistral Medium 3.5 | 32.8 |