Improving AI Chatbots with Bayesian Inference

AI models known as Large Language Models (LLMs) have garnered considerable interest in recent times, especially in the context of chatbots such as ChatGPT and Gemini. While AI chatbots can converse with users in a human-like fashion and respond to a wide array of prompts and queries, they still face notable challenges, such as hallucinations (producing arbitrary or nonsensical statements as facts) and breaches of safety (producing dangerous/hateful/biased content).

It is difficult to pin down the root cause of these issues as typical training paradigms are opaque and offer little in terms of explainability.

Bayesian inference, or the process by which prior beliefs are updated after observations to become (more accurate) posterior beliefs, may offer a solution. The aim of this project is to use Bayesian techniques to enhance AI chatbots along two complementary lines of inquiry.

First is the idea of training a chatbot in a more Bayesian fashion. Rather than optimizing a set of weights, training can instead be optimizing a set of weight distributions, beginning with some prior distribution that can take into account varying levels of domain knowledge. Research has already been done in other contexts (such as classification) where prior knowledge about a problem is encoded into the initial weight distribution of an AI model, which is then trained in a way that approximates Bayesian inference.

The result is a model that has not learned from a blank slate, but rather one in which there was full control over its initial biases. By adapting these methods to the chatbot context, it would be possible to directly initiate models with prior knowledge specific to an application even before any training takes place. As a result, such models would be more likely to operate within intended constraints.

Second is the complementary idea of developing Bayesian models that approximate already trained chatbots. In doing so, the behavior of the chatbot could be gleaned from analyzing the more transparent behavior of a Bayesian model. It has been proven that certain types of neural networks can be approximated by models known as Gaussian processes to within an arbitrary margin of error.

Extending these results to LLMs remains to be explored, but there is evidence to suggest LLMs do exhibit properties like a Bayesian model, for instance, they tend to answer questions correctly more often when the answer appears more often in their training text. This project falls within the EPSRC Artificial Intelligence Technologies research area.