Skip to main content
Back

Amazon's LitMT-8B Beats a 235B Model at Literary Translation

AWS and Peking University researchers train 8B and 14B models that match Claude Sonnet 4.5 on literary translation by generating better training data through iterative multi-agent refinement.

Paper, Full Access
  • LitMT-8B and LitMT-14B from AWS/Peking University match Claude Sonnet 4.5 (68.43) on English-to-Chinese literary translation, scoring 67.25 and 69.07 CEA100.
  • A multi-aspect iterative refinement pipeline uses two specialized agents — one for fluency, one for literary effect — to generate training data, beating human-written ground truth by 8.65 points.
  • DPO degrades performance in this setting (by 2–3 points vs. SFT); an explicit reward model with GRPO adds +1.51 points over SFT.
  • LitMT-8B outperforms its 235B teacher model (Qwen3-235B) using 30x fewer parameters, demonstrating genuine data quality improvement beyond distillation.
  • Data generation costs ~$168 (one-time); competing methods like ExTrans cost $1,038+ per training run due to online LLM-as-judge reward signals.
  • Current scope is English-to-Chinese only; no model weights released yet. Full paper and prompts at arxiv.org/abs/2606.05924.

Literary translation is one of the hardest problems in NLP. Metaphors, rhetorical devices, cultural tone , these are exactly the things that break standard machine translation pipelines. A team of researchers from Amazon Web Services and Peking University has published a new framework that attacks this problem from the data side, producing two fine-tuned models , LitMT-8B and LitMT-14B , that score competitively with Claude Sonnet 4.5 on an English-to-Chinese literary translation benchmark, using a fraction of the parameters.

The core tension in literary translation

Literary translation poses unique challenges due to the scarcity of high-quality annotated data and the need to balance expression fluency with literary effect. These two goals actively conflict: a translation that reads naturally in the target language often sacrifices the metaphors and stylistic devices of the original, while one that preserves every literary flourish can feel stilted and unnatural.

Recent approaches address this by training models with long chain-of-thought reasoning or using massive LLMs as judges during reinforcement learning. While effective, these methods face critical limitations: using LLM-as-a-judge as the reward signal requires expensive inference for every training sample, and this cost must be repeated for each new model, leading to prohibitive training costs. Long chain-of-thought reasoning also increases inference latency by up to 10x, making those approaches impractical for real-time use.

Splitting quality into two jobs

The key insight in this paper is that literary translation quality can be decomposed into two separable dimensions , expression fluency and literary effect , and that you can train specialized agents to optimize each one independently. The pipeline, called