Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Self-Correction in Large Language Models through Multi-Perspective Reflection

Mariana Costa, Alberlucia Rafael Soarez, Daniel Kim, Camila Ferreira

TL;DR

This work introduces PR-CoT, a poly-reflective, prompt-engineered extension to Chain-of-Thought reasoning designed to enhance self-correction in large language models. By organizing a preliminary CoT step around multiple explicit perspectives—logical consistency, information completeness, ethical considerations, and alternative solutions—the approach synthesizes a refined final answer without any model retraining. Across arithmetic, commonsense, ethical decision-making, and logical puzzles, PR-CoT yields higher logical consistency and greater error-correction rates than traditional CoT and single-reflection baselines, with pronounced gains in nuanced domains. While the method incurs higher computational overhead, the results demonstrate meaningful improvements in reliability and trustworthiness of LLM reasoning, suggesting strong practical value for complex, real-world tasks.

Abstract

While Chain-of-Thought (CoT) prompting advances LLM reasoning, challenges persist in consistency, accuracy, and self-correction, especially for complex or ethically sensitive tasks. Existing single-dimensional reflection methods offer insufficient improvements. We propose MyGO Poly-Reflective Chain-of-Thought (PR-CoT), a novel methodology employing structured multi-perspective reflection. After initial CoT, PR-CoT guides the LLM to self-assess its reasoning across multiple predefined angles: logical consistency, information completeness, biases/ethics, and alternative solutions. Implemented purely via prompt engineering, this process refines the initial CoT into a more robust and accurate final answer without model retraining. Experiments across arithmetic, commonsense, ethical decision-making, and logical puzzles, using GPT-three point five and GPT-four models, demonstrate PR-CoT's superior performance. It significantly outperforms traditional CoT and existing reflection methods in logical consistency and error correction, with notable gains in nuanced domains like ethical decision-making. Ablation studies, human evaluations, and qualitative analyses further validate the contribution of each reflection perspective and the overall efficacy of our poly-reflective paradigm in fostering more reliable LLM reasoning.

Enhancing Self-Correction in Large Language Models through Multi-Perspective Reflection

TL;DR

This work introduces PR-CoT, a poly-reflective, prompt-engineered extension to Chain-of-Thought reasoning designed to enhance self-correction in large language models. By organizing a preliminary CoT step around multiple explicit perspectives—logical consistency, information completeness, ethical considerations, and alternative solutions—the approach synthesizes a refined final answer without any model retraining. Across arithmetic, commonsense, ethical decision-making, and logical puzzles, PR-CoT yields higher logical consistency and greater error-correction rates than traditional CoT and single-reflection baselines, with pronounced gains in nuanced domains. While the method incurs higher computational overhead, the results demonstrate meaningful improvements in reliability and trustworthiness of LLM reasoning, suggesting strong practical value for complex, real-world tasks.

Abstract

While Chain-of-Thought (CoT) prompting advances LLM reasoning, challenges persist in consistency, accuracy, and self-correction, especially for complex or ethically sensitive tasks. Existing single-dimensional reflection methods offer insufficient improvements. We propose MyGO Poly-Reflective Chain-of-Thought (PR-CoT), a novel methodology employing structured multi-perspective reflection. After initial CoT, PR-CoT guides the LLM to self-assess its reasoning across multiple predefined angles: logical consistency, information completeness, biases/ethics, and alternative solutions. Implemented purely via prompt engineering, this process refines the initial CoT into a more robust and accurate final answer without model retraining. Experiments across arithmetic, commonsense, ethical decision-making, and logical puzzles, using GPT-three point five and GPT-four models, demonstrate PR-CoT's superior performance. It significantly outperforms traditional CoT and existing reflection methods in logical consistency and error correction, with notable gains in nuanced domains like ethical decision-making. Ablation studies, human evaluations, and qualitative analyses further validate the contribution of each reflection perspective and the overall efficacy of our poly-reflective paradigm in fostering more reliable LLM reasoning.
Paper Structure (29 sections, 3 equations, 4 figures, 4 tables)

This paper contains 29 sections, 3 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Chain-of-Thought Reasoning in Large Language Models
  4. Self-Correction and Reflection Mechanisms for LLMs
  5. Method
  6. Initial CoT Generation
  7. Multi-Perspective Reflection
  8. Logical Consistency Check ($v_1$)
  9. Information Completeness Check ($v_2$)
  10. Potential Bias and Ethical Consideration ($v_3$)
  11. Alternative Solution Exploration ($v_4$)
  12. Synthesis and Refinement
  13. Experiments
  14. Experimental Setup
  15. Baseline Methods
  16. ...and 14 more sections

Figures (4)

  • Figure 1: Overview of MyGO Poly-Reflective Chain-of-Thought (PR-CoT). The figure illustrates the motivations and challenges in LLM reasoning, conceptually distinguishes traditional Chain-of-Thought (CoT) from PR-CoT's multi-perspective reflection and iterative refinement, and highlights the resulting benefits for enhanced reasoning capabilities.
  • Figure 2: An overview of the methodological framework, illustrating the placement of our Proposed Method / Model (PR-CoT) within a broader data processing and evaluation pipeline. This figure outlines the overall flow from initial data input through to results analysis, encompassing preprocessing, feature extraction, and iterative refinement via a feedback loop.
  • Figure 3: Human Evaluation of Reasoning Quality (Average Score out of 5)
  • Figure 4: Efficiency Comparison: Average Tokens and Inference Time Per Task Instance