Self-Contrast: Better Reflection Through Inconsistent Solving Perspectives This paper investigates the limitations of self-evaluation in large language models (LLMs) for reflection and proposes a contrastive strategy called Self-Contrast to improve reflection accuracy and stability. The key challenge is that LLMs often provide overconfident or inconsistent feedback when self-evaluating, which hinders effective reflection. Self-Contrast addresses this by exploring diverse solving perspectives, contrasting differences between responses, and generating a checklist for re-examination and correction. The method involves three steps: creating diverse perspectives, contrasting inter-perspective discrepancies, and eliminating discrepancies through a checklist. By generating multiple perspectives, the LLM can identify inconsistencies and generate more accurate feedback. The experiments show that Self-Contrast significantly improves performance on mathematical reasoning and translation tasks compared to standard reflection and other baselines. It also demonstrates better generalization across different LLMs and tasks. The paper highlights that the self-evaluation process is often unreliable due to overconfidence or inconsistency in feedback. Self-Contrast mitigates this by contrasting multiple perspectives, leading to more accurate and stable reflection. The results show that Self-Contrast reduces the number of invalid and toxic reflections, and improves the accuracy of responses by identifying and correcting errors. The study also shows that contrasting incorrect solutions can be instructive, as it helps identify errors and biases. The method is flexible and can be applied to various tasks, making it a promising approach for improving LLM reflection capabilities. The paper concludes that Self-Contrast is an effective strategy for enhancing the self-correction ability of LLMs by leveraging contrastive evaluation.Self-Contrast: Better Reflection Through Inconsistent Solving Perspectives This paper investigates the limitations of self-evaluation in large language models (LLMs) for reflection and proposes a contrastive strategy called Self-Contrast to improve reflection accuracy and stability. The key challenge is that LLMs often provide overconfident or inconsistent feedback when self-evaluating, which hinders effective reflection. Self-Contrast addresses this by exploring diverse solving perspectives, contrasting differences between responses, and generating a checklist for re-examination and correction. The method involves three steps: creating diverse perspectives, contrasting inter-perspective discrepancies, and eliminating discrepancies through a checklist. By generating multiple perspectives, the LLM can identify inconsistencies and generate more accurate feedback. The experiments show that Self-Contrast significantly improves performance on mathematical reasoning and translation tasks compared to standard reflection and other baselines. It also demonstrates better generalization across different LLMs and tasks. The paper highlights that the self-evaluation process is often unreliable due to overconfidence or inconsistency in feedback. Self-Contrast mitigates this by contrasting multiple perspectives, leading to more accurate and stable reflection. The results show that Self-Contrast reduces the number of invalid and toxic reflections, and improves the accuracy of responses by identifying and correcting errors. The study also shows that contrasting incorrect solutions can be instructive, as it helps identify errors and biases. The method is flexible and can be applied to various tasks, making it a promising approach for improving LLM reflection capabilities. The paper concludes that Self-Contrast is an effective strategy for enhancing the self-correction ability of LLMs by leveraging contrastive evaluation.