Audit-First VAPO: Risk-Certified Selective Updates under Imperfect Verification

2026-09-27

AI Summary

This study addresses the risk of harmful updates introduced by imperfect verifiers in reinforcement learning by proposing an "audit-before-execute" safety framework. Methodologically, it decouples discrete direction admission from continuous magnitude control, employing an observation-based accept-appeal-abstain strategy for audit selection under a limited budget. Risk is controlled via conditional Hoeffding-Azuma bounds and a bounded-trust clipped-KL executor. This work provides the first theoretical certification of risk, coverage, and cost under finite-sample conditions. Experiments demonstrate that the model achieves 74.1% accuracy on the GSM8K benchmark while effectively suppressing the risk of harmful updates and ensuring certificate satisfaction rates across diverse noise scenarios.

Abstract

Imperfect verifiers can assign a harmful update direction even when clipping and regularization bound its magnitude. We introduce Audit-First VAPO, which separates discrete directional admission from continuous magnitude control. An observation-only accept-appeal-abstain policy uses a finite secondary-verification budget; its action trace is frozen before clean labels are joined. Simultaneous finite-sample bounds then certify selected harmful risk, coverage, and verifier-call rate over a predeclared policy family. Conditional Hoeffding-Azuma bounds account for the dependence induced by shared budgets, and rollout or verifier changes initiate a new certification stage. After admission, a bounded trust-clip-KL actuator controls magnitude. We evaluate two models on two reasoning benchmarks against static RLVR, matched-random selection, confidence thresholding, noise correction, and verifier augmentation. On Qwen3.5-0.8B and GSM8K at target risk $\rho=0.08$, RC-VAPO achieves 74.1% accuracy, selected harmful risk 0.0697, coverage 0.4125, and relative verifier cost $1.16\times$. At matched coverage and update magnitude, its selected-risk difference from matched random is -0.0260 with paired 95% interval $[-0.0364,-0.0157]$. Across asymmetric, confidence-dependent, and correlated-verifier noise, the certificate is satisfied on 57 of 60 independent runs. These comparisons isolate informative directional selection from proposal suppression, update shrinkage, and additional verifier computation.

Problem

• imperfect verifiers
• harmful risk
• selective updates
• risk certification
• directional admission

Innovation

• Audit-First VAPO
• Directional Admission
• Conditional Hoeffding-Azuma Bounds
• Risk Certification
• Trust-Clip-KL Actuator