Researchers proposed Parallel Power Tempering to improve small model reasoning. It uses multiple interacting replicas at different sharpening levels. Lower-power replicas explore diverse reasoning paths. Higher-power replicas exploit higher-likelihood responses. The method mitigates truncation bias found in prior power samplers. It investigates swap strategies under finite compute budgets. Experiments show it outperforms RL-post-trained models. It achieves performance comparable to frontier models. The paper was submitted to arXiv on 29 September 2026.