Chain-of-Thought

Hard trust regions for long-horizon LLM-RL: sequence-level masked importance sampling (Seq-MIS) rejects out-of-distribution samples instead of clipping them, and geometric masking makes the trust region length-invariant.