MAR: Multi-Agent Reflexion Improves Reasoning Abilities in LLMs 📝

Onat Ozer¹, Grace Wu¹, Yuchen Wang¹, Daniel Dosti¹, Honghao Zhang¹, Vivi De La Rue¹

Machine Learning Research FA 25
¹University of Michigan

LLMs have shown the capacity to improve their performance on reasoning tasks through reflecting on their mistakes, and acting with these reflections in mind. However, continual reflections of the same LLM onto itself exhibit degeneration of thought, where the LLM continues to repeat the same errors again and again even with the knowledge that its wrong. To address this problem, we instead use multi-agent multi-persona debate between LLMs to generate the reflections.

Under the traditional Reflexion framework, the agent uses a single LLM instance to reflect on its mistakes, and then uses these reflections as it attempts the task again. Under our proposed MAR framework, a multi-agent debate is initiated to generate the reflection, with the judge LLM giving the verdict.

Here we can see an example trial under the MAR framework, where each of the different personas generate a possible reason for failure, which the judge LLM then converts into a verdict that the LLM uses for self-reflecion.

As we can see, the MAR framework both outperforms the reflexion framework as well as the baseline gpt3.5 on question answering (HotPotQA) and programming (HumanEval).

MAR reflexion leads to consistent improvements in HotPotQA question-aswering, eventually outperforming the original Reflexion framework.

MLRE Conference Poster

Lightning Talk Presentation

BibTeX

If you found this work helpful and want to cite us, use the following BibTex given below:


      @misc{MAR-2025,
          title={MAR:Multi-Agent Reflexion Improves Reasoning Abilities in LLMs}, 
          author={Onat Ozer and Grace Wu and Yuchen Wang and Daniel Dosti and Honghao Zhang and Vivi De La Rue},
          year={2025},
          eprint={2512.20845},
          archivePrefix={arXiv},
          primaryClass={cs.AI},
          url={https://arxiv.org/abs/2512.20845}
        }