Atlas

Benchmarks

← All benchmarks

LLM Position Bias Rating Bonus

Chat & Writing · 2026-04-21

Rating Bonus is the average score advantage the same story receives when shown first rather than second. Values closer to zero indicate more order-invariant ratings; positive and negative values reflect opposite position effects.

Top models (closer to zero is better)

ModelScore
Trinity Large Thinking-0.0
Doubao Seed 2.0 Pro0.0
Gemini 3.1 Flash-Lite Preview0.0
MiMo-V2-Pro0.1
DeepSeek-V3.20.1
Opus 4.60.1
Qwen3.5 122B A10B0.1
MiniMax M2.70.1
Claude Fable 50.2
GPT-5.50.2
GLM-5.10.2
Opus 4.80.2
MiniMax M30.2
Gemini 3.5 Flash0.2
Qwen3.6 Plus (2026-04-02)0.2
Gemini 3.1 Pro Preview0.3
ERNIE 5.00.3
Opus 4.70.3
Sonnet 4.60.3
Mistral Medium 3.10.3
Qwen3.5 397B A17B0.3
Qwen3.7-Max0.4
DeepSeek-V4-Pro0.4
Gemma 4 31B IT0.4
Mistral Large 3 675B Instruct 2512-0.4
Grok 4.200.4
GPT-5.4 Mini0.4
GPT-5.40.5
Kimi K2.60.5
Mistral Medium 3.50.5
Kimi K2.50.6
Loading Atlas data…