“Experiments by Xie et al. (2023) using the Codex model with per-step self-evaluation in beam search achieved a 5-6% improvement on few-shot GSM8k, AQuA, and StrategyQA benchmarks.”