Technical insights on causal evaluation, LLM judges, and off-policy learning.
9 plots that tell you exactly when to ship your LLM changes and when to collect more data. From weight explosions to confidence intervals—see what reliable evaluation actually looks like.
Read moreStandard offline evaluation fails catastrophically for LLMs—0.6% effective sample size on Arena data. Here's how importance weighting with calibration and stabilization enables reliable counterfactual evaluation.
Read moreStandard importance sampling breaks catastrophically when evaluating LLMs offline. We explain why effective sample size drops to near zero and how Causal Judge Evaluation's projections restore reliable evaluation.
Read more