# ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models

**Authors:** Binfeng Xu, Zhiyuan Peng, Bowen Lei, Subhabrata Mukherjee, Yuchen Liu, Dongkuan Xu  
**arXiv:** [2305.18323](https://arxiv.org/abs/2305.18323)  
**Date:** 2023-05-23  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2305.18323v1/

## Key points

- Augmented Language Models (ALMs) combine reasoning capabilities of Large Language Models (LLMs) with knowledge retrieval and action execution tools.
- Existing ALM systems have huge computation complexity due to redundant prompts and repeated execution.
- ReWOO (Reasoning WithOut Observation) is a modular paradigm proposed in this study that detaches the reasoning process from external observations, significantly reducing token consumption.
- Comprehensive evaluations across six public NLP benchmarks and a curated dataset reveal consistent performance enhancements with the proposed methodology.
- ReWOO achieves 5x token efficiency and 4% accuracy improvement on HotpotQA, a multi-step reasoning benchmark.
- ReWOO demonstrates robustness under tool-failure scenarios.
- Decoupling parametric modules from non-parametric tool calls enables instruction fine-tuning to offload LLMs into smaller language models, thus substantially reducing model parameters.
- An illustrative work offloads reasoning ability from 175B GPT3.5 into 7B LLaMA, demonstrating significant potential.
- The study provides detailed descriptions of provided tools appended into the context prompt to enable zero-shot evaluation for ReWOO Planner covering information retrieval, comparison, equation solving, and calculating for different benchmarks.
- The number of reasoning steps k in exemplars is typically 2 or 3.
- This study proposes a novel approach to tackle challenges associated with existing ALM systems by introducing ReWOO as a modular paradigm that detaches the reasoning process from external observations while achieving consistent performance enhancements across various benchmarks.
- This approach offers significant potential for truly efficient and scalable ALM systems by enabling instruction fine-tuning to offload LLMs into smaller language models while reducing model parameters.

## Summary

Augmented Language Models (ALMs) have been developed to combine the reasoning capabilities of Large Language Models (LLMs) with tools that enable knowledge retrieval and action execution. However, existing ALM systems suffer from huge computation complexity due to redundant prompts and repeated execution. To address this challenge, a modular paradigm called ReWOO (Reasoning WithOut Observation) has been proposed in this study, which detaches the reasoning process from external observations and significantly reduces token consumption. The comprehensive evaluations across six public NLP benchmarks and a curated dataset reveal consistent performance enhancements with the proposed methodology. Notably, ReWOO achieves 5x token efficiency and 4% accuracy improvement on HotpotQA, a multi-step reasoning benchmark. Furthermore, ReWOO demonstrates robustness under tool-failure scenarios. The study also highlights the potential for truly efficient and scalable ALM systems by decoupling parametric modules from non-parametric tool calls that enable instruction fine-tuning to offload LLMs into smaller language models, thus substantially reducing model parameters. An illustrative work offloads reasoning ability from 175B GPT3.5 into 7B LLaMA, demonstrating significant potential. The study provides a detailed description of the provided tools appended into the context prompt to enable zero-shot evaluation. Exemplars are manually crafted for ReWOO Planner covering information retrieval, comparison, equation solving, and calculating for different benchmarks. The number of reasoning steps k in exemplars is typically 2 or 3. In summary, this study proposes a novel approach to tackle the challenges associated with existing ALM systems by introducing ReWOO as a modular paradigm that detaches the reasoning process from external observations while achieving consistent performance enhancements across various benchmarks. This approach offers significant potential for truly efficient and scalable ALM systems by enabling instruction fine-tuning to offload LLMs into smaller language models while reducing model parameters.