ACL 2026: Alibaba DAMO Academy's I2B-LPO Breaks RLVR Homogenization — From Repetitive Sampling to Effective Exploration

The I2B-LPO framework has been accepted at the ACL 2026 conference, aiming to enhance exploration strategies for reinforcement learning models post-training. By improving exploration behavior, the framework achieves an increase in model accuracy of up to 5.3% and a semantic diversity boost of 7.4% across various mathematical metrics. Reinforcement learning models with verifiable rewards (RLVR) are modern approaches that enhance model capabilities in mathematics and coding. These models utilize multiple thought pathways for the same problem, reinforcing correct paths while minimizing errors.