VStress: Correlation-Aware Auditing and Adaptive Budget Allocation for Repeated Verifiers
VStress-CA spends repeated verifier calls only when they add conditional information, improving accuracy under a fixed budget.
VStress defines an auditable replay contract for repeated verifier calls, and VStress-CA allocates those calls by estimated conditional marginal information, uncertainty, and cost. The controller freezes its decision before consulting a clean oracle and falls back when a dependence-shift alarm fires. At 35% symmetric corruption, majority-5 improved balanced accuracy from 0.6578 to 0.7739, while at 65% it lost 0.1226 points. In a fixed-budget comparison, adaptive allocation reached 0.6538 balanced accuracy with 3.4216 calls per item and an RLVR score of 0.6417.
- VStress-CA allocates verifier calls by estimated conditional information per cost.
- At 35% corruption, majority-5 raised balanced accuracy from 0.6578 to 0.7739.
- Adaptive allocation reached 0.6538 accuracy with about 3.42 calls per item.
- A dependence-shift alarm disables channel preference and falls back to exact-stop.
Full article168 words · extracted from arxiv.org · click to collapse
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.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.36958