Alibaba’s Qwen team has introduced QwenGyre, an elastic reinforcement learning (RL) framework specifically designed for extreme-long-horizon agent training. This framework represents a significant advancement in online reinforcement learning, enabling the training of coding agents through unmodified, black-box harnesses.
QwenGyre is tailored for rollouts that can last for hours and process nearly 1 million tokens, allowing for more complex and extended training scenarios. The introduction of this framework is expected to enhance the capabilities of agents in handling long-term tasks, marking a notable development in the field of AI and reinforcement learning.
Compiled automatically by the Tech AI Newsdesk from public AI-news sources and summarised in our own words.
