Long Live The Balance: Information Bottleneck Driven Tree-based Policy Optimization

Researchers introduced IB-Score, an Information Bottleneck-based metric for measuring exploration-exploitation balance in LLM reinforcement learning.

They also proposed IB-TPO, a tree-based policy optimization framework with IB-guided sampling and Monte Carlo estimation for online RL training.

On standard benchmarks, the method improved sampling efficiency and delivered 2.9% to 3.6% gains over GRPO and other online RL approaches.

The work offers a new way to make LLM reinforcement learning more stable and effective.