Companies like Google, Anthropic, and OpenAI are continuously improving their AI models, which affects how users interact with them. Apple’s research team has developed a new method to streamline the transition between model upgrades, reducing inconsistencies by up to 40%.
In a study titled “MUSCLE: A Model Update Strategy for Compatible LLM Evolution,” researchers emphasize the importance of ensuring a seamless transition between models to maintain user experience quality. They aim to reduce negative flips, where the new model predicts an incorrect output that was previously predicted correctly by the older model.
The study acknowledges that users have personalized ways of interacting with chatbots and constantly adapting to new models can be challenging and counterproductive to Apple’s user experience goals.
Apple presents MUSCLE
A Model Update Strategy for Compatible LLM Evolution
Large Language Models (LLMs) are frequently updated due to data or architecture changes to improve their performance. When updating models, developers often focus on increasing overall performance… pic.twitter.com/ATm2zM4Poc
— AK (@_akhaliq) July 15, 2024
The researchers developed MUSCLE, a strategy that minimizes regression between models without requiring developers to retrain the entire base model. It involves using training adapters, which are small AI modules that can integrate at various points within the LLM.
By fine-tuning these specific modules instead of the entire model, developers can enhance the performance of specific tasks at a lower training cost and with minimal parameter increase. These adapters act as plug-ins for large language models, enabling targeted adjustments rather than comprehensive retraining.
In testing with LLMs like Meta’s Llama and Microsoft’s Phi, the study found that negative flips occurred in 60% of cases. Implementing the MUSCLE strategy reduced this occurrence by up to 40% compared to the control group.
