Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning

A new reinforcement-learning policy called Translate-R1 learns when to invoke translation for large language models, cutting cost while preserving accuracy across 22 languages, according to research posted to arXiv on June 5, 2026 [1]. The performance gap across languages in large language models is well documented, and closing it natively requires pretraining or fine-tuning on corpora that, for most languages, do not exist [1]. 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 [1]. Prior work resolves this with language-specific rules, domain heuristics, language identifiers, or external routers, each requiring manual engineering [1]. The Translate-R1 framework instead learns 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 [1]. The researchers continued reinforcement learning on the post-trained Qwen3-4B model across 22 languages in three resource tiers—High, Low, and XLow—and five domains, introducing a confidence-gated GSPO mechanism for cost-sensitive tool use [1]. The gated policy lifts reward over the baseline by +4.6 on High, +23.5 on Low, and +17.5 on XLow [1]. 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 [1]. Existing approaches to cost-sensitive tool use either apply a fixed penalty that cannot adapt to language difficulty, or use group-relative signals that are corrupted by lucky guesses on low-resource languages [4]. Both over-suppress tool use for exactly the languages that need it most. The confidence gate introduced in Translate-R1 applies cost pressure only when the model demonstrates strong native competence, adapting automatically to language difficulty without tier labels [4]. To simulate behavior on a completely unseen language, the team created two synthetic languages, where the gated policy improved +18.7 over the overconfident baseline policy that underutilizes the tool even on these incomprehensible inputs [1]. The policy also transferred zero-shot to nine held-out languages absent from training [1]. Large language models are typically based on transformer architecture and are pre-trained to predict the next word, then fine-tuned to follow instructions [6]. The Translate-R1 work builds on a broader push to equip LLMs with tool-use capabilities through reinforcement learning, as seen in frameworks such as Tool-R1, which generates executable Python code for multi-step tool workflows [5]. The Translate-R1 paper includes analysis of how tool use emerges over training, per language and per domain [1].