Test-Time Scaling of Reasoning Models for Machine Translation
Zihao Li, Shaoxiong Ji, Jörg Tiedemann
2025-10-21
Summary
This paper investigates whether making large language models 'think' for longer during translation improves the quality of the translations they produce, specifically focusing on machine translation tasks.
What's the problem?
While letting models spend more time on problems has helped with things like math and coding, it wasn't clear if the same trick would work for translating languages. The core issue is whether simply giving a model more computation time during translation actually leads to better results, or if there's a limit to how much benefit you can get from just letting it 'think' longer.
What's the solution?
Researchers tested 12 different language models on a variety of translation tasks. They looked at three different ways to use these models: directly translating, forcing the model to reason through the translation step-by-step, and having the model review and correct its own initial translation. They found that just letting general-purpose models think longer didn't help much, but if they first trained the model specifically for translation tasks, then letting it think longer *did* improve results, up to a certain point. Also, forcing the model to reason beyond its natural stopping point actually made things worse, but letting it self-correct was very effective.
Why it matters?
This research shows that simply increasing computation time isn't a magic bullet for better machine translation. Instead, the key is to either specialize the model for translation or use the extra computation time for specific tasks like self-correction. This helps us understand how to best utilize powerful language models for translation and where to focus our efforts to get the biggest improvements.
Abstract
Test-time scaling (TTS) has enhanced the performance of Reasoning Models (RMs) on various tasks such as math and coding, yet its efficacy in machine translation (MT) remains underexplored. This paper investigates whether increased inference-time computation improves translation quality. We evaluate 12 RMs across a diverse suite of MT benchmarks spanning multiple domains, examining three scenarios: direct translation, forced-reasoning extrapolation, and post-editing. Our findings show that for general-purpose RMs, TTS provides limited and inconsistent benefits for direct translation, with performance quickly plateauing. However, the effectiveness of TTS is unlocked by domain-specific fine-tuning, which aligns a model's reasoning process with task requirements, leading to consistent improvements up to an optimal, self-determined reasoning depth. We also find that forcing a model to reason beyond its natural stopping point consistently degrades translation quality. In contrast, TTS proves highly effective in a post-editing context, reliably turning self-correction into a beneficial process. These results indicate that the value of inference-time computation in MT lies not in enhancing single-pass translation with general models, but in targeted applications like multi-step, self-correction workflows and in conjunction with task-specialized models.