VeriBound: PAC-Bayesian Generalization Bounds for Process Reward Models Trained with Formal Verification Tools

Short summary

VeriBound provides the first PAC-Bayesian generalization bounds for Process Reward Models trained with formal verification tools, closing a theoretical gap in LLM step-level reasoning verification. The framework establishes sample complexity bounds of O(d log(d/δ) / ε²), linear convergence guarantees, and error propagation analysis for downstream Best-of-K performance.

• •PAC-Bayesian framework for PRMs trained with formal verification tools (Z3, Isabelle)
• •Sample complexity bound: O(d log(d/δ) / ε²) examples needed for ε-generalization error
• •Linear convergence proof and error propagation bounds for Best-of-K performance

Generated with AI, which can make mistakes.