Trust Region Masking for Long-Horizon LLM Reinforcement Learning

Researchers propose Trust Region Masking (TRM) to address off-policy mismatch problems in Large Language Model reinforcement learning pipelines. The method provides the first non-vacuous monotonic improvement guarantees for long-horizon LLM-RL tasks by masking entire sequences that violate trust region constraints.