Reproduce “Test-Time Scaling of Reasoning Models for Machine Translation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zihao Li, Shaoxiong Ji, Jörg Tiedemann
This paper systematically investigates the utility of test-time scaling (TTS) for Reasoning Models (RMs) in machine translation (MT), revealing limited benefits for general-purpose models in direct translation but consistent quality improvements in post-editing workflows and for domain-specific models. The research also demonstrates that forcing models to extend reasoning beyond their natural stopping point degrades translation quality.
Key insights
- The effectiveness of test-time scaling in MT is highly conditional, showing limited gains for general-purpose RMs in direct translation due to a 'reasoning ceiling,' but strong benefits for domain-specific fine-tuned models.
- Test-time scaling is particularly effective in multi-stage post-editing (self-correction) workflows, reliably improving translation quality, especially for mid-sized models.
- Forcibly extending a Reasoning Model's deliberation beyond its natural stopping point consistently degrades translation quality, indicating the importance of respecting intrinsic reasoning termination.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.