AI

Tail-Influence Sampling for CVaR Policy Evaluation

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 18:18:24 UTC.

Briefings

A 2026 study proposes Tail-Influence Sampling (TIS) to allocate evaluation budgets for CVaR policy evaluation by estimating tail influence from pilot models. TIS reallocates queries to high-impact kernels, achieving oracle asymptotic variance and first-order MSE including pilot cost. In frozen language-model review workflows, anchored TIS outperformed mean-influence blends in 23 of 24 MMLU-Pro settings and reduced MSE by 2.4-3.4 times compared to rollouts on six-call FinQA reviews. The method also identifies an exact-grid regime where tail- and mean-optimal allocations coincide.

Key points

• Tail-Influence Sampling (TIS) estimates tail influence scales from a pilot model to allocate fresh queries toward high-impact kernels for CVaR estimation.
• In frozen language-model review workflows, anchored TIS achieved 2.4-3.4 times lower MSE than rollouts on six-call FinQA reviews.

Context and limitations

The study covers only frozen language-model review workflows and CliffWalking environments; results are limited to positive quantile margins and fixed dimension settings.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 10/10/2026, 01:09:19 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiaw67wkn4z2um5ftkaaxon5jzuozqafd35jjz3vdmajrzc5l5c6de