Feb 3, 05:21 PM

Google DeepMind Study on Llama-3.2-Vision-11B Reveals SFT Memorizes While RL Enhances Generalization in AI Training, February 2, 2025

A new paper from Google DeepMind and collaborators from HK University, UC Berkeley, and NYU compares the effectiveness of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) in the training of large language models (LLMs) and vision-language models (VLMs). The research indicates that SFT is primarily focused on memorization, akin to a student learning by rote from examples, while RL enhances generalization across various tasks, allowing models to adapt better to new challenges. The study highlights that while SFT stabilizes model outputs, RL is crucial for improving adaptability and performance in tasks requiring unique solutions, such as mathematics and logic. This research underscores the importance of both methods in developing advanced AI systems.

#Google DeepMind #HK University #UC Berkeley #NYU #Reinforcement Learning

Written with ChatGPT (GPT-4o mini).