[Paper Review] AI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts

Tongshuang Wu, Michael Terry+1 more|arXiv (Cornell University)|Oct 4, 2021

Topic Modeling34 citations

TL;DR

The paper introduces LLM Chaining, a method to decompose complex tasks into sequential sub-tasks with intermediate outputs, enables a modular interactive interface, and shows that chaining improves task quality, transparency, and user sense of collaboration in a 20-person study.

ABSTRACT

Although large language models (LLMs) have demonstrated impressive potential on simple tasks, their breadth of scope, lack of transparency, and insufficient controllability can make them less effective when assisting humans on more complex tasks. In response, we introduce the concept of Chaining LLM steps together, where the output of one step becomes the input for the next, thus aggregating the gains per step. We first define a set of LLM primitive operations useful for Chain construction, then present an interactive system where users can modify these Chains, along with their intermediate results, in a modular way. In a 20-person user study, we found that Chaining not only improved the quality of task outcomes, but also significantly enhanced system transparency, controllability, and sense of collaboration. Additionally, we saw that users developed new ways of interacting with LLMs through Chains: they leveraged sub-tasks to calibrate model expectations, compared and contrasted alternative strategies by observing parallel downstream effects, and debugged unexpected model outputs by "unit-testing" sub-components of a Chain. In two case studies, we further explore how LLM Chains may be used in future applications

Motivation & Objective

• Address the lack of transparency and controllability in complex LLM-assisted tasks.
• Propose a modular set of LLM primitive operations to build Chains of subtasks.
• Provide an interactive interface to create, edit, and inspect LLM Chains and intermediate outputs.
• Evaluate how chaining affects task performance and user experience.
• Demonstrate through case studies potential future applications of LLM Chains.

Proposed method

• Define eight primitive LLM operations grouped by objective to enable task decomposition.
• Design prompt templates and default parameters for each operation to support Chain construction.
• Construct an interactive Chain interface that visualizes operations and data layers and supports editing at multiple granularities.
• Run a within-subject user study comparing Chaining with a non-Chaining baseline using the same underlying LLM.
• Analyze how Chains improve transparency, controllability, collaboration, and task outcomes.
• Present case studies illustrating future applications and discuss design challenges and opportunities.

Experimental results

Research questions

• RQ1 Does Chaining improve task outcomes compared to a non-Chaining approach with the same LLM?
• RQ2 Does Chaining enhance system transparency and controllability from the user’s perspective?
• RQ3 What new interaction patterns emerge when users work with Chains and intermediate data?
• RQ4 How can Chains help with debugging or calibrating expectations of LLM behavior?
• RQ5 What are the potential future applications and design considerations for LLM Chains?

Key findings

• Chaining significantly improved key dimensions of the human-AI experience, including transparency, controllability, collaboration, and mental support.
• Participants achieved higher-quality outcomes roughly 82% of the time when using Chaining.
• Users leveraged sub-tasks to calibrate model expectations and compare alternative strategies via parallel downstream effects.
• Users debugged unexpected model outputs by unit-testing sub-components of a Chain.
• The study reveals new interaction patterns for using LLMs through Chains, such as isolating sub-tasks and inspecting intermediate results, without altering the underlying model.
• Case studies illustrate potential future applications and highlight challenges and opportunities in deploying LLM Chains.