# FReM: A Flexible Reasoning Mechanism for Balancing Quick and Slow Thinking in Long-Context Question Answering

**Authors:** Zhengyi Zhao, Shubo Zhang, Zezhong Wang, Bin Liang, Binyang Li, Kam-Fai Wong  
**arXiv:** [2503.22985](https://arxiv.org/abs/2503.22985)  
**Date:** 2025-03-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2503.22985v1/

## Key points

- Existing systems in the field of question answering benefit from large language models (LLMs) but face limitations in slow and quick reasoning modes.
- Slow thinking involves exploring every possible reasoning path, leading to overthinking, while quick thinking relies on pattern matching and lacks true understanding.
- A novel framework called FReM is proposed to address these challenges by adjusting reasoning depth based on question complexity.
- FReM leverages synthetic reference QA examples to guide the model's own reasoning process and balances quick-thinking and slow-thinking approaches by offering flexible reasoning steps.
- Experimental results demonstrate that FReM outperforms quick-thinking models by capturing deeper logic beyond pattern matching and surpasses slow-thinking models by reducing unnecessary exploration.
- Different thinking strategies for QA systems include quick-thinking approaches that match patterns in input data and slow-thinking methods that focus on extended reasoning processes.
- FReM introduces a dynamic mechanism to adapt reasoning complexity to the demands of each question, offering self-directed and flexible reasoning steps for efficient handling of both simple and complex queries in LCQA tasks.
- The approach helps models move beyond superficial pattern matching and prevents unnecessary exploration during lengthy reasoning paths, demonstrating improved accuracy and scalability across seven QA datasets.

## Summary

In the field of , existing systems have benefited from large language models (LLMs) but face limitations in slow and quick reasoning modes. Slow thinking involves exploring every possible reasoning path, leading to overthinking, while quick thinking relies on pattern matching and lacks true understanding. To address these challenges, a novel framework called  is proposed. This method adjusts reasoning depth based on question complexity, leveraging synthetic reference QA examples to guide the model's own reasoning process. FReM balances quick-thinking and slow-thinking approaches by offering flexible reasoning steps. It dynamically decides whether to use simple or detailed pathways for handling single-step and multi-hop questions efficiently. Experimental results demonstrate that FReM outperforms quick-thinking models by capturing deeper logic beyond pattern matching and surpasses slow-thinking models by reducing unnecessary exploration. In the context of LCQA, previous works have utilized strategies such as RAG-based retrieval, iterative planning-retrieval-generation, and coreference resolution. However, these methods may become inefficient or overly complex for low-complexity questions. Additionally, different thinking strategies for QA systems include quick-thinking approaches that match patterns in input data and slow-thinking methods that focus on extended reasoning processes. The proposed  introduces a dynamic mechanism to adapt reasoning complexity to the demands of each question. By synthesizing different reference demos to guide the model's own reasoning process, FReM offers self-directed and flexible reasoning steps for efficient handling of both simple and complex queries in LCQA tasks. This approach helps models move beyond superficial pattern matching and prevents unnecessary exploration during lengthy reasoning paths. Overall, FReM demonstrates improved reasoning accuracy and scalability across seven QA datasets, particularly enhancing performance on complex multihop questions in Long-Context Question Answering tasks. This innovative framework shows promise in advancing LCQA methodologies by balancing quick and slow thinking strategies effectively.