Reproduce “Calibrate, Don't Curate: Label-Efficient Estimation from Noisy LLM Judges” on OpenResearch

About this paper

Yanran Li

This research from Columbia University challenges the conventional practice of curating top-performing judges in multi-judge evaluation for large language models, demonstrating that calibrating a full panel of diverse judges yields more reliable probabilistic estimates. Including all judges, even weaker ones, consistently improves calibrated metrics like Negative Log-Likelihood and Brier score compared to selecting only the most accurate, often leading to substantial performance gains.

Key insights

• Judges perceived as 'weak' in terms of point accuracy can still contribute valuable, learnable conditional information that improves overall probability estimation after proper calibration.
• The benefit of including all judges for calibrated risk is theoretically supported by principles like the monotonicity of oracle-calibrated risk and an extended Calibrated Jury Theorem.
• The choice and capacity of the post-hoc calibrator (e.g., beta calibration) are more critical for achieving high-quality calibrated predictions than the specific heterogeneity-aware aggregation method.

Summary by alphaXiv.