StateFlow: Enhancing LLM Task-Solving through State-Driven Workflows

Abstract: It is a notable trend to use LLMs to tackle complex tasks, e.g., tasks that require a sequence of actions and dynamic interaction with tools and external environments. In this paper, we propose StateFlow, a novel LLM-based task-solving paradigm that conceptualizes complex task-solving processes as state machines. In StateFlow, we distinguish between "process grounding" (via state and state transitions) and "sub-task solving" (through actions within a state), enhancing control and interpretability of the task-solving procedure. A state represents the status of a running process. The transitions between states are controlled by heuristic rules or decisions made by the LLM, allowing for a dynamic and adaptive progression. Upon entering a state, a series of actions is executed, involving not only calling LLMs guided by different prompts, but also the utilization of external tools as needed. Our results show that StateFlow significantly enhances LLMs' efficiency. For instance, StateFlow achieves 13% and 28% higher success rates compared to ReAct in InterCode SQL and ALFWorld benchmark, with 5x and 3x less cost respectively. We also show that StateFlow can be combined with iterative refining methods like Reflexion to further improve performance.

• The paper introduces StateFlow, a finite state machine framework that enhances control and accuracy in multi-step LLM task-solving.
• It details a methodology that combines internal LLM responses with external tool interactions across states like Init, Observe, Solve, Verify, and Error.
• Experimental results show significant improvements, with StateFlow achieving higher success rates and lower interaction costs compared to traditional methods like ReAct.

StateFlow: Enhancing LLM Task-Solving through State-Driven Workflows

The paper "StateFlow: Enhancing LLM Task-Solving through State-Driven Workflows" introduces a paradigm shift in leveraging LLMs to effectively handle complex, multi-step tasks. This research outlines a framework, termed StateFlow, which conceptualizes the LLM task-solving process as a finite state machine (FSM), thereby enhancing control, accuracy, and efficiency in task completion.

Overview of StateFlow

Motivation and Problem Statement

The existing methodologies for complex task-solving using LLMs, such as Chain of Thought (CoT) and ReAct prompting, rely heavily on the LLM's implicit judgment to determine the progress status and take subsequent actions. However, these models often fall short in consistently making correct state inferences and tracking their actions, leading to inefficiencies and errors. The paper addresses this gap by posing a key research question: How can we exert more precise control and guidance over LLMs?

Conceptual Framework

StateFlow models the LLM task-solving process as a state machine, a mathematically rigorous control system commonly used in practical applications. This model comprises defined states, state transitions, and execution of specific actions within each state. The FSM approach in StateFlow allows for a state to represent the LLM’s task-solving phase, and transitions between states are governed by rules based on the current context and outputs, which can dynamically adapt using specific prompts or tools. This modeling ensures that each phase of the process is tracked, controlled, and managed precisely.

Practical Implementation and Evaluation

StateFlow employs a combination of internal LLM responses and external tool usage to navigate through the states. The framework initiates at an Init state and navigates through various states such as Observe, Solve, Verify, and Error, each designed to perform specific actions and transitions. The transitions are guided by context history and string matching or explicit conditional checks using the LLM.

The research demonstrates the application of StateFlow using GPT-3.5-Turbo and GPT-4-Turbo across complex tasks like SQL and Bash scripting from the InterCode benchmark. The results show significant improvements in success rates: StateFlow achieves 60.83% in SQL tasks with GPT-3.5-Turbo, a substantial increase from 50.68% using ReAct. Similarly, in Bash tasks, StateFlow attains a success rate of 37% compared to 32.5% with ReAct. Additionally, the StateFlow's efficiency metrics indicate a notable reduction in interaction costs, and execution errors, with a cost reduction up to five times compared to the ReAct prompting method.

Implications and Future Work

The introduction of StateFlow has several important implications:

1. Enhanced Control: By using state machines, StateFlow allows developers and researchers to have fine-grained control over the task-solving process.
2. Efficiency: The framework reduces unnecessary computations and interactions, leading to cost-effective solutions.
3. Robustness: StateFlow's structured approach increases the robustness and reliability of LLMs in handling complex tasks.

Future research avenues include automating the construction of StateFlow models using LLMs, enabling them to dynamically generate and refine workflows and employing active learning strategies to iteratively adjust the state machine based on performance feedback. There is also potential to expand the framework to handle even more intricate and heterogeneous tasks by incorporating parallel actions and asynchronous processing.

Conclusion

StateFlow presents a significant advancement in the landscape of LLM-based task-solving frameworks. Its underlying FSM-based methodology aligns complex task-solving with enhanced control, efficiency, and consistency. The experimental results substantiate its effectiveness over existing prompting methods, marking a promising direction for future AI research in automating and optimizing LLM-driven workflows.