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

Alibaba DAMO Academy introduced I2B-LPO, an exploration-enhancement framework for RLVR post-training.

Instead of relying on repetitive sampling, it encourages more diverse and informative reasoning trajectories.

The approach reportedly improves math benchmark accuracy by up to 5.3% and semantic diversity by up to 7.4%.

It highlights a key RLVR limitation: more samples do not automatically lead to better learning.

The work, from Alibaba’s Intelligent Decision team, was accepted to ACL 2026 Main.