Repro - On the Power of (Approximate) Reward Models for Inference-Time Scaling: Sequential Monte Carlo and Beyond