Google unveils AI framework for faster, more diverse search results

Image credit: Abondance

Google introduced its new artificial intelligence framework, Retrieve-for-Train, on Tuesday, designed to improve the diversity and speed of AI-powered search results by separating training from execution.

The innovation addresses long-standing issues of redundancy and high latency prevalent in traditional language models used for generating search queries, according to the company.

Traditional language models often suffer from what Google described as ‘paraphrastic collapse,’ a phenomenon where the models generate numerous redundant sub-queries, leading to less diverse results.

Furthermore, the autoregressive nature of classic language models contributes to high latency because they generate tokens sequentially, slowing down the overall process.

Retrieve-for-Train tackles these challenges by employing a one-time, offline reinforcement learning process to create training data for a more efficient, lightweight diffusion model.

This approach allows the system to generate diverse sub-queries without the real-time computational burden of autoregressive models, Google stated.

The final diffusion model developed through this framework contains 53.9 million parameters, a relatively compact size for such a system.

Google reported that this model offers a significant speed increase, performing 12 to 20 times faster than conventional autoregressive methods.

To ensure the quality and relevance of the generated results, Retrieve-for-Train incorporates a three-criterion reward system during its training phase.

This system evaluates groundedness, diversity via a Vendi Score, and alignment, which collectively prevent the model from producing absurd or repetitive outcomes.

The company said that by focusing on these criteria, the framework ensures that the AI generates useful and varied search suggestions.

The global technology firm emphasized that the separation of training and execution phases is key to achieving both efficiency and improved output quality in AI search.

This development aims to enhance user experience by delivering more comprehensive and quicker search results across Google’s platforms.

Source: Abondance