PAMT: Process-Aligned Reinforcement Learning for Multi-Domain Translation

PAMT addresses credit-assignment bottlenecks in multi-domain machine translation where Long-CoT reasoning can cause terminology drift. The method combines domain-aware Long-CoT supervision with reinforcement learning to optimize specific intermediate translation steps rather than just the final output.