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

2026-08-03

AI Summary

This work addresses the performance degradation of large reasoning models in multidomain machine translation, particularly in terminology-intensive or style-constrained scenarios, where misalignment between the inference process and translation objectives limits effectiveness. To overcome this, the paper proposes the PAMT framework, which introduces a step-level process reward mechanism for the first time. PAMT integrates domain-aware Long Chain-of-Thought (Long-CoT) reasoning with reinforcement learning, jointly optimizing both the intermediate reasoning path and the final output through a combination of sequence-level outcome rewards and step-level process rewards. This approach effectively alleviates the credit assignment bottleneck. Experimental results demonstrate that PAMT consistently outperforms strong baselines across two backbone architectures, surpassing specialized machine translation systems in in-domain, out-of-domain, and multilingual settings, while achieving performance on par with state-of-the-art large language model (LLM) and large reasoning model (LRM) approaches.

Abstract

Multi-domain machine translation (MDMT) requires more than fluent generation: it demands domain-sensitive translation decisions such as domain disambiguation, terminology control, and stylistic adaptation. Large reasoning models (LRMs) make such decisions explicit through intermediate translation steps, but our analysis across 15 domains and four translation directions shows that this explicit reasoning is double-edged: it improves long-form and high-difficulty translation, yet often drifts in terminology-intensive and stylistically constrained settings. We trace this failure to a credit-assignment bottleneck: existing methods optimize final outputs or coarse trajectories, but cannot identify which translation steps actually help the final translation. To address this, we propose PAMT, a process-aligned training framework that combines cold-start domain-aware Long-CoT supervision with reinforcement learning. PAMT uses sequence-level format and outcome rewards for the final translation, together with a step-level process reward that measures how much each explicit translation step increases the likelihood of the reference translation. Across two backbones, PAMT improves over base models, outperforms MT-specialized baselines on average, and remains competitive with strong LLMs/LRMs across in-domain, OOD, and multilingual settings.

Problem

• multi-domain machine translation
• domain-sensitive translation
• credit assignment
• translation process alignment
• terminology control

Innovation

• process-aligned reinforcement learning
• multi-domain machine translation
• step-level process reward
• Long-CoT supervision
• credit assignment