Enhancing LLM Reasoning with Multi-Path Collaborative Reactive and Reflection agents

Abstract: Agents have demonstrated their potential in scientific reasoning tasks through LLMs. However, they often face challenges such as insufficient accuracy and degeneration of thought when handling complex reasoning tasks, which impede their performance. To overcome these issues, we propose the Reactive and Reflection agents with Multi-Path Reasoning (RR-MP) Framework, aimed at enhancing the reasoning capabilities of LLMs. Our approach improves scientific reasoning accuracy by employing a multi-path reasoning mechanism where each path consists of a reactive agent and a reflection agent that collaborate to prevent degeneration of thought inherent in single-agent reliance. Additionally, the RR-MP framework does not require additional training; it utilizes multiple dialogue instances for each reasoning path and a separate summarizer to consolidate insights from all paths. This design integrates diverse perspectives and strengthens reasoning across each path. We conducted zero-shot and few-shot evaluations on tasks involving moral scenarios, college-level physics, and mathematics. Experimental results demonstrate that our method outperforms baseline approaches, highlighting the effectiveness and advantages of the RR-MP framework in managing complex scientific reasoning tasks.

• The paper introduces the RR-MP framework to enhance large language model reasoning by combining reactive and reflection agents.
• The multi-path reasoning approach enables diverse perspectives to collaborate, achieving 75.94% accuracy on scientific tasks.
• Ablation studies confirm that reflection agents are critical for preventing degeneration-of-thought and ensuring robust performance.

Enhancing LLM Reasoning with Multi-Path Collaborative Reactive and Reflection Agents

Introduction

The paper "Enhancing LLM Reasoning with Multi-Path Collaborative Reactive and Reflection Agents" introduces the RR-MP framework designed to bolster the reasoning capabilities of LLMs by addressing common issues such as insufficient accuracy and Degeneration-of-Thought (DoT). The RR-MP framework employs a dual-agent model consisting of reactive and reflection agents to facilitate multi-path reasoning, aiming to enhance scientific reasoning accuracy without additional training through strategic cooperation between agents.

Figure 1: Reactive and Reflection agents with Multi-Path Reasoning

Framework Overview

The RR-MP framework involves a multi-path reasoning mechanism where reactive agents generate preliminary answers and reflection agents critically analyze these answers to stimulate self-correction. This collaborative approach mimics human cognitive processes that rely on diverse reasoning paths to solve complex tasks effectively. The system operates in zero-shot and few-shot contexts to evaluate scenarios involving moral judgments, physics, and mathematics, demonstrating superior performance compared to baseline approaches.

Figure 2: The reasoning process of the reactive agent and reflection agent. The reactive agent receives information from the external environment, decomposes it into sub-tasks, and stores them in the database. The reflection agent performs each sub-task through a process of supplementation or critique and returns the results to the reactive agent.

Methodology

Multi-Path Reasoning

The multi-path reasoning approach enhances cognitive flexibility by creating diverse reasoning paths, akin to collaborative human problem-solving methods. This framework enables the consolidation of distinct agent-generated perspectives to derive higher quality solutions. Through theoretical analysis, the paper establishes that increasing the number of reasoning paths optimally enhances decision quality, theoretically grounding the framework’s advantages.

Multi-Agent Interaction

The RR-MP framework encompasses multiple interaction paradigms, including both collaborative and debate modes within same-domain and different-domain contexts. Reactive and reflection agents engage through shared memory modules to iteratively refine and optimize decision pathways, employing role-specific dialogue strategies to ensure adaptability and cross-agent communication without cognitive interference.

Figure 3: Accuracy (\%) with and without stimulation roles.

Experimental Results

Tests conducted on College Physics, College Mathematics, and Moral Scenarios datasets demonstrate that the RR-MP framework achieves substantial accuracy improvements in few-shot settings. The collaboration mode, particularly across different domains, proves most effective, achieving an average accuracy of 75.94\%. These findings underscore the benefits of multi-agent interaction designs and highlight the critical role of reflection in complex reasoning tasks.

Comparison and Ablation Studies

The paper includes detailed ablation studies that confirm the necessity of reflection agents and multiple dialogue instances for maintaining effective reasoning capabilities. Single-instance setups face performance degradation due to cognitive overlaps, affirming the significance of well-defined agent interactions and adaptive reasoning paths.

Conclusion

The RR-MP framework markedly improves the reasoning capabilities of LLMs in complex scientific tasks through innovative methods of multi-path reasoning and collaborative multi-agent interactions. Future work should focus on automating prompt engineering to further enhance framework scalability and adaptability. This research contributes significantly to the optimization of human-machine collaboration models, ensuring the accurate and efficient solution of intricate reasoning tasks.