⚖️ 3D LLM Judge Position Bias: Probit Choice Model, Swap Consistency and Order-Averaged Debiasing
A Blender-built balance scale tilts toward the slot a noisy, biased judge prefers while trays collect the verdicts of a Monte Carlo benchmark: accuracy of fixed, random and order-averaged judging against closed forms for the probit model, with five independent checks.
⚙ Under the hood
A Blender-built balance scale tilts toward the slot a noisy, biased judge prefers while trays collect the verdicts of a Monte Carlo benchmark: accuracy of fixed, random and order-averaged judging against closed forms for the probit model, with five independent checks.
LLM-as-a-JudgePosition BiasEvaluationProbit ModelSwap ConsistencyDebiasingMonte CarloBenchmarks
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install