[R] GRACE: Discriminator-Guided Chain-of-Thought Reasoning

**TLDR:** The paper proposes a decoding approach that improves multi-step (chain-of-thought) reasoning by using a discriminator to score and guide the generation of correct reasoning steps. It outperforms self-consistency and verifiers on various tasks and enhances both final answer accuracy and intermediate reasoning correctness.

​

**Paper:** [arxiv.org/abs/2305.14934](https://arxiv.org/abs/2305.14934)

**Code:** [https://github.com/mukhal/grace/](https://github.com/mukhal/grace/)

​

​

accuracy decoding grace guide intermediate machinelearning paper reasoning tasks thought