Adaptive LLM Judge Tuning Shows Universal Gains Over Fixed Parameters in Early Sims

IanArawjo · x · 2026-08-13

The developer reconsiders the previous rule of thumb of 0.4 IRR (Information Retrieval Rate), which was developed under a PPI implementation with a fixed shrinkage parameter for λ.

He proposes an adaptive solution called "power tuning tuning," where the target shrinkage for λ dynamically depends on how aligned the judge appears to be. Early simulations indicate that this approach is almost universally better than either fixed alternative.

Original post →

More from Research

Research channel →