VStress: Correlation-Aware Auditing and Adaptive Budget Allocation for Repeated Verifiers

2026-09-29

AI Summary

This study addresses the inefficiencies in redundant information utilization and budget allocation caused by correlations among repeated verifier calls. To this end, we propose a correlation-aware adaptive budget allocation framework. Methodologically, we construct an auditable replay contract that estimates conditional marginal information via a sealed calibration set, discounts uncertainty, and normalizes costs. By introducing dependency drift alerts and exact stopping fallback mechanisms, the framework shifts from post-hoc warnings to ex-ante auditable decisions, enabling adaptive stopping or abstention. Experimental results demonstrate that VStress-CA achieves a balanced accuracy of 0.6538 under a fixed budget with an average of 3.42 calls and an RLVR score of 0.6417, significantly outperforming existing baselines.

Abstract

Repeated verifier calls are useful only when they contribute conditional information. We introduce VStress, an auditable replay contract, and VStress-CA, a correlation-aware allocation policy that estimates the conditional marginal information of an unqueried verifier on a sealed calibration split, discounts uncertainty, normalizes by call cost, and stops or abstains when the next call is not informative. The controller freezes its decision and cost ledger before joining the clean oracle; a dependence-shift alarm disables channel preference and falls back to exact-stop. The controlled audit gives the mechanism boundary: at 35% symmetric corruption, majority-5 improves balanced accuracy from 0.6578 to 0.7739, whereas at 65% it loses 0.1226 points. In the matched fixed-budget comparison, breadth, redundancy, and adaptive allocation obtain balanced accuracies 0.6048, 0.6375, and 0.6538, with 3.4216 calls per item and an RLVR score of 0.6417 for VStress-CA. Dependence diagnostics also increase from same-model repeats to cross-family channels, with conditional marginal gains of 0.0126, 0.0462, and 0.0913. These measurements turn correlation from a post-hoc warning into an auditable allocation decision.

Problem

• verifier correlation
• budget allocation
• auditable replay
• conditional information
• repeated verifiers

Innovation

• Correlation-Aware Allocation
• Auditable Replay Contract
• Conditional Marginal Information
• Dependence-Shift Alarm
• Adaptive Budget Allocation