[Paper Review] Self-Contrast: Better Reflection Through Inconsistent Solving Perspectives

Wenqi Zhang, Yongliang Shen+5 more|arXiv (Cornell University)|Jan 4, 2024

Topic Modeling4 citations

TL;DR

This paper proposes Self-Contrast, a novel self-reflection method that enhances LLM reasoning by generating diverse solving perspectives, contrasting their discrepancies, and using these contrasts to build a corrective checklist. By mitigating overconfidence and inconsistency in self-evaluation, Self-Contrast improves reflection reliability and achieves significant performance gains across reasoning and translation tasks without external feedback.

ABSTRACT

The reflection capacity of Large Language Model (LLM) has garnered extensive attention. A post-hoc prompting strategy, e.g., reflexion and self-refine, refines LLM's response based on self-evaluated or external feedback. However, recent research indicates without external feedback, LLM's intrinsic reflection is unstable. Our investigation unveils that the key bottleneck is the quality of the self-evaluated feedback. We find LLMs often exhibit overconfidence or high randomness when self-evaluate, offering stubborn or inconsistent feedback, which causes poor reflection. To remedy this, we advocate Self-Contrast: It adaptively explores diverse solving perspectives tailored to the request, contrasts the differences, and summarizes these discrepancies into a checklist which could be used to re-examine and eliminate discrepancies. Our method endows LLM with diverse perspectives to alleviate stubborn biases. Moreover, their discrepancies indicate potential errors or inherent uncertainties that LLM often overlooks. Reflecting upon these can catalyze more accurate and stable reflection. Experiments conducted on a series of reasoning and translation tasks with different LLMs serve to underscore the effectiveness and generality of our strategy.

Motivation & Objective

• To address the instability of LLM self-reflection due to overconfident or inconsistent self-evaluation.
• To improve the reliability of post-hoc reflection in LLMs without relying on external feedback.
• To develop a method that leverages internal diversity of thought to uncover overlooked errors.
• To create a flexible, adaptive reflection mechanism that enhances both breadth and depth of reasoning.
• To demonstrate the generality and effectiveness of the approach across multiple LLMs and task types.

Proposed method

• The method generates multiple solving perspectives through self-curated prompts tailored to the input request.
• It selects pairs of responses with significant discrepancies for contrastive analysis.
• Discrepancies between perspectives are analyzed to identify potential errors, overlooked details, or reasoning flaws.
• The system synthesizes these discrepancies into a structured checklist for self-correction.
• The checklist guides revision by prompting the model to re-examine key assumptions and reasoning steps.
• The entire process is end-to-end LLM-driven, requiring no pre-defined agent roles or manual prompt engineering.

Experimental results

Research questions

• RQ1 Why does self-reflection in LLMs often fail despite post-hoc prompting strategies?
• RQ2 What causes the instability in LLM self-evaluation, particularly overconfidence and inconsistency?
• RQ3 Can contrasting multiple solving perspectives improve the quality of self-reflection?
• RQ4 How does internal diversity of reasoning lead to more reliable error detection and correction?
• RQ5 Can a self-driven, contrastive reflection mechanism outperform standard self-evaluation without external feedback?

Key findings

• Self-Contrast significantly improves reflection performance across multiple reasoning and translation tasks.
• The method reduces the failure rate of self-correction by mitigating overconfidence and inconsistency in self-evaluation.
• On average, Self-Contrast increases the correction rate of incorrect initial responses from 15.1% to over 40% in tested benchmarks.
• The approach achieves consistent gains across different LLMs, demonstrating strong generalization.
• Contrastive analysis of diverse perspectives uncovers errors that single-perspective evaluation misses.
• The generated checklists are highly effective in guiding accurate revision, outperforming standard self-reflection baselines.