ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models (2305.18323v1)

Abstract: Augmented LLMs (ALMs) blend the reasoning capabilities of LLMs with tools that allow for knowledge retrieval and action execution. Existing ALM systems trigger LLM thought processes while pulling observations from these tools in an interleaved fashion. Specifically, an LLM reasons to call an external tool, gets halted to fetch the tool's response, and then decides the next action based on all preceding response tokens. Such a paradigm, though straightforward and easy to implement, often leads to huge computation complexity from redundant prompts and repeated execution. This study addresses such challenges for the first time, proposing a modular paradigm ReWOO (Reasoning WithOut Observation) that detaches the reasoning process from external observations, thus significantly reducing token consumption. Comprehensive evaluations across six public NLP benchmarks and a curated dataset reveal consistent performance enhancements with our 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. Beyond prompt efficiency, decoupling parametric modules from non-parametric tool calls enables instruction fine-tuning to offload LLMs into smaller LLMs, thus substantially reducing model parameters. Our illustrative work offloads reasoning ability from 175B GPT3.5 into 7B LLaMA, demonstrating the significant potential for truly efficient and scalable ALM systems.

• The paper introduces a novel ReWOO framework that decouples reasoning from observations to reduce redundant token generation and lower computational cost.
• It employs a Plan-Work-Solve paradigm, splitting tasks into Planner, Worker, and Solver components to streamline multi-step reasoning.
• The approach achieves a fivefold token efficiency improvement and a 4% accuracy boost on benchmarks like HotpotQA, demonstrating scalability with smaller models.

An Analysis of ReWOO: Decoupling Reasoning from Observations for Efficient Augmented LLMs

In recent years, the integration of external tools with LLMs has resulted in the formation of Augmented LLMs (ALMs), which exhibit enhanced reasoning capabilities by retrieving knowledge and executing actions autonomously. Despite their potential, the current architectures of ALMs often incur substantial computational costs due to interleaved reasoning-observation sequences. They follow a paradigm where reasoning and retrieval processes are interdependent, causing computation complexity with redundant token generation and execution. The paper under review introduces AutoDistil, a novel prompting strategy in the form of ReWOO (Reasoning WithOut Observation), advocating a shift from traditional interleaved models towards a more modular, detached approach.

ReWOO strategically decouples reasoning from external tool observations, facilitating token-efficient ALMs by considerably reducing redundancy. With an innovative Plan-Work-Solve paradigm, reasoning and observation processes are compartmentalized into three separate components: the Planner, Worker, and Solver. The Planner establishes a chain of action plans while the Worker acquires external insights, subsequently synthesized by the Solver into a cohesive solution for the task. This separation allows for streamlined processing where token consumption is conserved drastically as compared to the conventional methods.

The paper's empirical evaluations span across six public datasets including HotpotQA, TriviaQA, GSM8K, and StrategyQA among others, alongside a curated dataset to test the real-life applicability of the model. ReWOO demonstrated a clear advantage, achieving a fivefold improvement in token efficiency coupled with a 4% boost in accuracy on the multistep reasoning benchmark, HotpotQA. Furthermore, its robustness to tool failures signifies its resilience and utility in diverse real-world scenarios beyond traditional NLP tasks.

An intriguing dimension to ReWOO is its potential in offloading reasoning tasks from large LLMs like GPT-3.5 to smaller models like LLaMA 7B, thus paving the way towards scalable ALMs. This was evidenced through fine-tuning experiments illustrating that specialized smaller models can emulate the reasoning abilities of larger, resource-intensive counterparts at a fraction of their computational expense. This alone signifies a frontier in creating modular systems with significant potential for both industry-wide applications and theoretical advancements in language processing.

However, the research acknowledges limitations where reasoning without available context becomes inherently challenging, citing tasks that require stage-wise environmental learning as problematic for a decoupled approach. Therefore, optimizing a Directed Acyclic Graph (DAG) representation for LLMs, tools, and sub-models may offer a promising solution, leveraging each component's strengths efficiently.

In summary, the paper presents a compelling argument for a paradigm shift towards modular, token-efficient ALMs. By presenting a robust framework through ReWOO, the authors challenge current conventions in ALM design, advocating for structures that not only conserve resources but also maintain or enhance the performance of complex reasoning tasks. Future directions include refining tool representation and optimizing graph-based execution, signifying a move towards more holistic, end-to-end trainable systems, facilitating the development of scalable AGI solutions. As the landscape of LLMs continues to evolve, AutoDistil signifies a critical step in optimizing models for efficient deployment and execution across various computational and environmental thresholds.