Perplexity and NVIDIA have announced the introduction of JOLT, a groundbreaking joint teacher-student reinforcement learning method. This innovative post-training approach allows a single model to function both as a teacher and a student, streamlining the training process.
In testing, JOLT achieved accuracy levels comparable to the GRPO method while processing 6.5 times fewer training tokens. This efficiency could potentially revolutionize how AI models are trained, making them more accessible and less resource-intensive.
The development of JOLT marks a significant advancement in reinforcement learning methodologies, highlighting the ongoing collaboration between Perplexity and NVIDIA in pushing the boundaries of AI research and application.
Compiled automatically by the Tech AI Newsdesk from public AI-news sources and summarised in our own words.
