Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?

The paper's core contribution is treating dependency parses and grammatical rules as explicit intermediate supervision signals during training, not just features. This is distinct from prior chain-of-thought work because it operationalizes linguistic structure as a training objective across three different learning paradigms (in-context, fine-tuning, RL) rather than just a reasoning artifact.

This connects directly to the June 1st work on multilingual reasoning (Learning When to Translate for Multilingual Reasoning), which identified language comprehension gaps as a bottleneck in reasoning systems. Where that paper proposed selective translation as a workaround, this research takes the opposite approach: it assumes the model stays in the target language but gives it explicit grammatical scaffolding to compensate for low-resource data scarcity. Both papers treat language-specific failure modes as addressable through structured intervention rather than brute-force scaling. The HERO'S JOURNEY benchmark from the same day also tested whether models can extract and apply procedural rules; this paper asks whether providing those rules upfront (via dependency trees) helps models generalize better on translation tasks.

If the synthetic reasoning traces improve performance on held-out language pairs beyond Xibe and Chin (particularly on morphologically rich or syntactically distant languages), that confirms the approach generalizes. If performance gains collapse when the dependency treebank is noisy or incomplete, that signals the method is brittle and dependent on annotation quality rather than robust linguistic reasoning.