Foundation Models for Conversation

Reducing hallucination in conversation systems and developing models that can perform a variety of specialized tasks


The project explores strategies to anchor dialogue response generation on dependable enterprise content, utilizing structures like decision trees, flow charts, and structured data for grounding responses. It also investigates using large language models (LLMs) to harness approved unstructured documents for generating responses. A core aspect of this exploration is to ensure that generated responses remain faithful to the content of documents fed into the model.

By building foundation models for digital interaction data, the project team aims to overcome research challenges around data modeling, identifying the right pre-training objectives, and ensuring cost-effectiveness and efficiency.

The ambitious goal is to support a myriad of downstream tasks, often new unseen tasks, with the same foundational model requiring minimal tuning. Examples of such tasks include predicting a user's following action, analyzing a user group's activity footprint to match a product with its target user group, or initiating a personalized dialog flow with a customer.