Unlocking Reasoning Capability on Machine Translation in Large Language Models

2026-02-16

AI Summary

This work addresses the limitations of current reasoning-oriented large language models in machine translation, which suffer from suboptimal performance due to the use of generic reasoning strategies that lack task-specific structural guidance. To overcome this, the authors propose a structured reasoning framework tailored for machine translation, comprising multi-stage drafting, accuracy refinement, fluency enhancement, and selective iterative revision. This approach dynamically synthesizes reasoning trajectory data and leverages it for post-training. The study is the first to reveal the adverse impact of generic reasoning on translation quality and transcends the constraints of conventional linear reasoning paradigms. Evaluated on the WMT24++ benchmark, the proposed method significantly outperforms both standard fine-tuning and generic reasoning injection techniques, demonstrating the efficacy of task-customized structured reasoning in machine translation.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Knowledge Representation and Reasoning: Argumentation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Reasoning-oriented large language models (RLMs) achieve strong gains on tasks such as mathematics and coding by generating explicit intermediate reasoning. However, their impact on machine translation (MT) remains underexplored. We systematically evaluate several open- and closed-weights RLMs on the WMT24++ benchmark and find that enabling explicit reasoning consistently degrades translation quality across languages and models. Analysis reveals that MT reasoning traces are highly linear, lacking revision, self-correction and exploration of alternative translations, which limits their usefulness. Furthermore, injecting higher-quality reasoning traces from stronger models does not reliably improve weaker models'performance. To address this mismatch, we propose a structured reasoning framework tailored to translation, based on multi-step drafting, adequacy refinement, fluency improvement, and selective iterative revision. We curate a synthetic dataset of dynamic structured reasoning traces and post-train a large reasoning model on this data. Experiments show significant improvements over standard translation fine-tuning and injected generic reasoning baselines. Our findings demonstrate that reasoning must be task-structured to benefit MT.

Problem

• machine translation
• reasoning capability
• large language models
• reasoning traces
• translation quality

Innovation

• structured reasoning
• machine translation
• reasoning traces
• large language models
• post-training