Tail-Influence Sampling for CVaR Policy Evaluation
The paper targets cheaper, sharper estimates of rare policy failures by spending evaluation budget where tail risk is most sensitive.
It introduces Tail-Influence Sampling, which estimates which queryable parts of a stochastic workflow most affect lower-tail CVaR and shifts fresh queries toward them. The authors say the method reaches oracle asymptotic variance under fixed dimension and a positive quantile margin, with an anchored variant staying within a factor of two. Reported tests show lower MSE than rollout-based baselines on CliffWalking and frozen language-model review workflows. Source: HF Daily Papers' note.
It introduces Tail-Influence Sampling, which estimates which queryable parts of a stochastic workflow most affect lower-tail CVaR and shifts fresh queries toward them. The authors say the method reaches oracle asymptotic variance under fixed dimension and a positive quantile margin, with an anchored variant staying within a factor of two. Reported tests show lower MSE than rollout-based baselines on CliffWalking and frozen language-model review workflows. Source: HF Daily Papers' note.
score 4