Mending the Holes: Mitigating Reward Hacking in Reinforcement Learning for Multilingual Translation

Rewriting source text with large language models (LLMs) before translation has been shown to improve machine translation (MT) quality. However, we find that prompt-based rewriting can degrade translation quality rather than improve it, particularly when smaller LLMs, such as 4B-parameter models, are used. We argue that this limitation stems from the difficulty of controlling rewriting behavior through natural-language prompts alone: a rewrite is useful only if it improves downstream translation, yet existing prompt-based methods do not explicitly optimize for this signal. To address this issue, we propose RLSR (Reinforcement Learning for Source Rewriting), a reinforcement learning framework that trains the rewriting model with a reward based on the downstream translation-quality improvement produced by each rewrite. Experiments across six MT systems and 16 language pairs show that our 4B RLSR-trained rewriting models significantly outperform both the no-rewriting baseline and prompt-based rewriting baselines at the same model scale, while remaining competitive with baselines that use a 235B LLM.

The performance gap across languages in LLMs is well documented, and closing it natively requires pretraining or fine-tuning on corpora that, for most languages, do not exist. Translation offers an alternative: converting an input into the model's dominant language unlocks its full capabilities at once. Applying translation to every input, however, is wasteful for languages the model already handles, while leaving the choice to the model fails in the opposite way, as LLMs are overconfident and skip the tool even when they cannot understand the input. Prior work resolves this with language-specific rules, domain heuristics, language identifiers, or external routers, each requiring manual engineering. We instead learn a single policy that decides when to translate from reward alone, developing language- and domain-adaptive introspection that assesses its own comprehension and invokes translation only when it cannot solve a task natively. Using data built by our answer-preserving translation pipeline, we continue RL on the post-trained Qwen3-4B across 22 languages in 3 resource tiers (High, Low, XLow) and 5 domains, and introduce confidence-gated GSPO for cost-sensitive tool use. The gated policy lifts reward over the baseline by +4.6 on High, +23.5 on Low, and +17.5 on XLow. Against an unconstrained policy that almost always translates, it preserves full reward at 63% of the cost and is Pareto-optimal across 87% of the cost-sensitivity range. Additionally, to simulate behavior on a completely unseen language, we create 2 synthetic languages, where our gated policy improves +18.7 over the overconfident baseline that underutilizes the tool even on these incomprehensible inputs. The policy transfers zero-shot to 9 held-out languages, and we analyze how tool use emerges over training, per language and per domain.

Extending large language models (LLMs) to low-resource languages often incurs an "alignment tax": improvements in the target language come at the cost of catastrophic forgetting in general capabilities. We argue that this trade-off arises from the rigidity of supervised fine-tuning (SFT), which enforces token-level surface imitation on narrow and biased data distributions. To address this limitation, we propose a semantic-space alignment paradigm powered by Group Relative Policy Optimization (GRPO), where the model is optimized using embedding-level semantic rewards rather than likelihood maximization. This objective encourages meaning preservation through flexible realizations, enabling controlled updates that reduce destructive interference with pretrained knowledge. We evaluate our approach on Tibetan-Chinese machine translation and Tibetan headline generation. Experiments show that our method acquires low-resource capabilities while markedly mitigating alignment tax, preserving general competence more effectively than SFT. Despite producing less rigid surface overlap, semantic RL yields higher semantic quality and preference in open-ended generation, and few-shot transfer results indicate that it learns more transferable and robust representations under limited supervision. Overall, our study demonstrates that reinforcement learning with semantic rewards provides a safer and more reliable pathway for inclusive low-resource language expansion.