English
Share
Sign In
Automated Reasoning and Tooling (ART): Simplifying complex problem solving
ART (Automatic Reasoning and Tool-use) is a new method to improve the capabilities of language models (LMs) for complex tasks, and was proposed in 2023 by Bhargavi Paranjape in his study <ART: Automatic multi-step reasoning and tool-use for large language models>.
ART has the following progression:
Select a task: ART selects a relevant task from a library of tasks that include demos for a variety of tasks.
Tool integration: Call external tools whenever needed, integrate the tool's output, and resume the inference process.
Zero-shot generalization: ART learns to decompose new tasks and use tools appropriately through demonstrations.
Human feedback: Users can improve ART's performance by modifying the task library or adding new tools.
Example prompts
"Hector가 수평선 위로 35도 각도로 72.0 N 힘으로 체인을 당깁니다. 이 힘의 수평 구성 요소를 결정하십시오."Prompt Processing Process:
1.
Search: Do a search for "What is the formula for the horizontal component?"
2.
Code Generation: Generate Python code to solve the problem using the formula "Fx = Ftens * cosine(θ)"
3.
Run the code: Run the generated code to derive the "Fx" value.
4.
Result: "Fx" value is calculated as 58.9789
How to apply and use ART
Task Library: ART uses a task library that includes demo programs for various tasks.
Tool library: Calls external tools as needed (e.g., search, code generation, and execution) and integrates their results into your program.
Constructing Prompts: ART constructs prompts for new tasks using demonstrations of related tasks.
Automated Reasoning: ART breaks down a task into steps and uses tools to gather information and make inferences at each step.
Human feedback: Users can improve ART's performance by modifying the task library or adding new tools.
Comparison with CoT (Chain of Thought):
CoT: CoT is a method for generating step-by-step reasoning processes in natural language for complex inference.
ART: ART extends the CoT approach to include the use of external tools (e.g., search, code execution) when necessary, providing additional information and computational power to the inference process.
ⓒ 2023. Haebom, all rights reserved.
It may be used for commercial purposes with permission from the copyright holder, provided the source is cited.