Test-Time Scaling of Reasoning Models for Machine Translation

TL;DR

This paper investigates whether increasing inference-time computation via test-time scaling (TTS) improves machine translation when using reasoning models. It evaluates 12 RMs across eight MT benchmarks and three scenarios: direct translation, forced extrapolation, and post-editing. The main finding is that TTS yields limited gains for general-purpose RMs in direct translation, but domain-specific fine-tuning aligns reasoning with task demands to unlock benefits up to a natural planning depth, while forcing longer reasoning degrades quality; TTS is highly effective in post-editing workflows. These results suggest focusing resources on task-aligned training and multi-step correction pipelines rather than universal computation scaling, with potential for dynamic budgeting and retrieval-augmented setups in future work.

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.

Figures (6)

• Figure 1: Illustration of the effectiveness of test-time scaling in reasoning models for machine translation. (1) TTS for general-purpose RMs yields only a small initial performance gain, but quickly plateauing as increased inference cost. (2) Forcing RMs to reason beyond their natural stopping point degrades quality by introducing noise. (3) In contrast, TTS becomes effective when applied to RMs specifically developed for MT. (4) TTS shows improvements in post-editing workflows. All these highlight TTS's value in MT lies in task-specialized models and multi-step self-correction, rather than as a robust strategy for enhancing single-pass translation with general-purpose RMs.
• Figure 2: Average GRB scores of Qwen-3 and Cogito models across all datasets with varying thinking budgets.
• Figure 3: Performance of Grok-3-mini across tasks, showing the difference between high- and low-effort reasoning. Subfigure (a) reports results under the GRB metric, and (b) shows results under GRF.
• Figure 4: Performance (dashed lines, right axis) and actual generated thinking tokens (solid lines, left axis) of DRT models across 3 literary translation tasks.
• Figure 5: Effectiveness of test-time scaling in post-editing scenario.