# Self-Contrast: Better Reflection Through Inconsistent Solving Perspectives

**Authors:** Wenqi Zhang, Yongliang Shen, Linjuan Wu, Qiuying Peng, Jun Wang, Yueting Zhuang, Weiming Lu  
**arXiv:** [2401.02009](https://arxiv.org/abs/2401.02009)  
**Date:** 2024-01-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2401.02009v3/

## Key points

- Authors Wenqi Zhang, Yongliang Shen, Linjuan Wu, Qiuying Peng, Jun Wang, Yueting Zhuang, and Weiming Lu propose a novel strategy to enhance the reflective abilities of Large Language Models (LLMs).
- Existing post-hoc prompting strategies like reflexion and self-refine rely on self-evaluated or external feedback to refine LLM responses.
- Recent research has shown the instability of LLM's intrinsic reflection without external feedback.
- The quality of self-evaluated feedback is identified as a key bottleneck in LLM's reflective process.
- LLMs often exhibit overconfidence or high randomness when evaluating themselves, leading to stubborn or inconsistent feedback that hampers effective reflection.
- The authors introduce a method involving adaptively exploring diverse solving perspectives tailored to specific requests, contrasting differences between them, and summarizing discrepancies into a checklist for re-examination and resolution.
- By providing LLMs with a range of perspectives through this method, the authors aim to mitigate biases and improve overall reflection accuracy and stability.
- Identified discrepancies in solving perspectives serve as indicators of potential errors or uncertainties that LLMs may overlook during reflection.
- Experiments conducted on various reasoning and translation tasks using different LLMs highlight the significant impact of incorporating diverse perspectives into the reflective process to enhance performance effectively.

## Summary

In their paper titled "Self-Contrast: Better Reflection Through Inconsistent Solving Perspectives," authors Wenqi Zhang, Yongliang Shen, Linjuan Wu, Qiuying Peng, Jun Wang, Yueting Zhuang, and Weiming Lu delve into the reflection capacity of Large Language Models (LLMs) and propose a novel strategy to enhance their reflective abilities. The existing post-hoc prompting strategies like reflexion and self-refine rely on self-evaluated or external feedback to refine LLM responses. However, recent research has highlighted the instability of LLM's intrinsic reflection in the absence of external feedback. The authors identify the quality of self-evaluated feedback as a key bottleneck in LLM's reflective process. They observe that LLMs often demonstrate overconfidence or high randomness when evaluating themselves, leading to stubborn or inconsistent feedback that hampers effective reflection. To address this issue, the authors introduce the  method. This approach involves adaptively exploring diverse solving perspectives tailored to specific requests, contrasting the differences between them, and summarizing these discrepancies into a checklist for re-examination and resolution. By providing LLMs with a range of perspectives through , the authors aim to mitigate stubborn biases and improve overall reflection accuracy and stability. The identified discrepancies in solving perspectives also serve as indicators of potential errors or uncertainties that LLMs may overlook during reflection. Through careful consideration of these discrepancies, more accurate and reliable reflections can be catalyzed. The effectiveness and generality of the  strategy are underscored through experiments conducted on various reasoning and translation tasks using different LLMs. The results highlight the significant impact of incorporating diverse perspectives into the reflective process to enhance performance and mitigate biases effectively. Overall, this innovative approach offers promising insights for improving LLM's reflective capabilities in various applications within natural language processing domains.