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

TL;DR

This work tackles reliability and coherence challenges in LLM-driven scientific reasoning by introducing RR-MP, a framework that couples reactive and reflection agents across multiple reasoning paths. By enabling diverse, independently running paths and external reflection to correct and improve outputs, the approach reduces degeneration of thought without additional training. The authors provide theoretical justification via implicit welfare optimization across paths and validate the method on moral, physics, and mathematics tasks, achieving superior zero-shot and few-shot performance, particularly in cross-domain collaboration. The findings underscore the value of multi-path, multi-agent collaboration and reflective stabilization for robust complex reasoning in LLMs, with future work aimed at automated prompt design to enhance adaptability.

Abstract

Agents have demonstrated their potential in scientific reasoning tasks through large language models. 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.

Figures (5)

• Figure 1: Reactive and Reflection agents with Multi-Path Reasoning
• Figure 2: A comparison of the self-consistency method and our approach. Using the College Physics datasets as an example, our multi-path, multi-role interactive framework effectively mitigates errors caused by the majority of incorrect judgments in majority voting and leverages accurate reflection stimulated by the reactive agents' input on the reflection agents' reasoning. Even if the first path yields an incorrect result, the final answer is achieved through reflection analysis of the second path. Refer to B.1 in the appendix for details.
• Figure 3: 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. Based on the feedback, the reactive agent refines its reasoning and completes the scientific reasoning process.
• Figure 4: Accuracy (%) with and without stimulation roles.
• Figure 5: Three typical interaction paradigms in human-agent collaboration frameworks.