Tail-Influence Sampling for CVaR Policy Evaluation
The paper proposes a budget-allocation method for estimating a fixed policy’s lower-tail risk without relying on full rollouts.
Tail-Influence Sampling estimates which queryable parts of a stochastic workflow most affect CVaR, then spends fresh evaluation queries on those components. The authors report oracle-level asymptotic variance under stated conditions, with a visitation-anchored version kept within a factor of two of the oracle. In experiments, TIS cuts MSE versus learned occupancy and complete rollouts on CliffWalking, and the anchored variant improves error in language-model review workflows including MMLU-Pro and FinQA. Source: ArXiv · AI/CL/LG's note.
Tail-Influence Sampling estimates which queryable parts of a stochastic workflow most affect CVaR, then spends fresh evaluation queries on those components. The authors report oracle-level asymptotic variance under stated conditions, with a visitation-anchored version kept within a factor of two of the oracle. In experiments, TIS cuts MSE versus learned occupancy and complete rollouts on CliffWalking, and the anchored variant improves error in language-model review workflows including MMLU-Pro and FinQA. Source: ArXiv · AI/CL/LG's note.
score 4