Transparent and Controllable Human-Ai Interaction Via Chaining of Machine-Learned Language Models

Smart overview of the Invention

The patent application addresses the challenges in human-AI interaction by leveraging the chaining of machine-learned language models. While existing large language models (LLMs) have shown potential in handling simple tasks, their lack of transparency and controllability limits their effectiveness in complex scenarios. This approach involves connecting multiple language model instances, where the output of one serves as the input for the next, enhancing performance through aggregated steps.

Technical Field

This disclosure pertains to machine learning and language modeling, focusing on creating transparent and controllable interactions between humans and AI. It describes a system where users can modularly build or edit a chain of language model instances using a graphical user interface. This setup aims to improve interpretability and control over the models' outputs.

Background

Language models, particularly LLMs, have made significant strides in natural language processing (NLP), demonstrating capabilities in tasks like translation and question answering. Despite this, they struggle with complex tasks requiring multi-step reasoning due to their limited ability to connect multiple functional blocks. Additionally, their opacity makes them challenging to interact with, as users often find it difficult to adjust prompts for better results.

Summary of Invention

The invention introduces a computing system that chains multiple machine-learned language models to enhance interpretability and performance. The system processes an initial language input through a sequence of model instances, each configured to handle specific subtasks. A user interface is provided for visualizing and editing these chains, allowing users to modify prompts and outputs at each step, thus tailoring the process to specific needs.

Implementation Details

The system supports both single and multiple language models within the chain, offering flexibility in task execution. Users can select predefined templates for prompts or customize them as needed. The interface allows viewing the chain structure or focusing on individual steps, enabling detailed edits. The system can generate outputs in natural or programming languages, with some models accessed via APIs if stored separately from the computing system.