The DRT-o1 (Deep Reasoning Translation) project aims to introduce the successful experience of Long Chain-of-Thought reasoning into neural machine translation (MT). Related research papers include "DRT: Deep Reasoning Translation via Long Chain-of-Thought" (ACL 2025 Findings), "DeepTrans: Deep Reasoning Translation via Reinforcement Learning" (TACL 2025), and "ExTrans: Multilingual Deep Reasoning Translation via Exemplar-Enhanced Reinforcement Learning" (arXiv 2025).
Purpose: The project explores the technical route of using long thought reasoning for machine translation. The goal is not to achieve performance on MT that competes with OpenAI O1, but rather, as a byproduct of exploration, to advance research in this direction.
Main capabilities: DRT mines English sentences containing similes or metaphors from existing literary books. These sentences are suitable for translation through long thought. The project proposes a multi-agent framework composed of three agents: a translator, an advisor, and an evaluator, used to synthesize MT samples with long thought processes, synthesizing a total of 22,264 samples. Based on Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct, and Qwen2.5-14B-Instruct, DRT-8B, DRT-7B, and DRT-14B were trained respectively. During inference, the DRT models first generate the thought content and then provide the final translation. DeepTrans-7B enhances the free translation ability of a deep reasoning large model through reinforcement learning, using DeepSeek-v3 (671B) as the reward model, and designs scoring criteria for translation and the thought process. ExTrans-7B adopts exemplar-enhanced reinforcement learning, using DeepSeek-R1 (671B) to generate an exemplar translation for each training MT sample, and compares the policy model's translation result with the exemplar translation to provide rewards. mExTrans-7B extends ExTrans-7B from English-to-Chinese to 11 languages, including Chinese, English, Arabic, Czech, German, Spanish, French, Italian, Japanese, Russian, and Korean.
Usage: When performing DRT model inference, use the system prompt "You are a philosopher skilled in deep thinking, accustomed to exploring complex problems with profound insight.", and the user prompt is "Please translate the following text from English to Chinese:\n[An English text]". The project provides deployment and inference methods for Huggingface Transformers and VLLM; for specific code and commands, please refer to the repository. The synthetic data is named MetaphorTrans and is stored in data/MetaphorTrans_*.jsonl, where text and trans represent the source English sentence and the target Chinese translation, respectively, and thought represents the MT thought content.
Applicable scenarios: Suitable for studying the application of long chain-of-thought reasoning in machine translation, especially the translation of English sentences containing similes or metaphors, as well as multilingual deep reasoning translation research.
Limitations: The project explicitly states that the goal is not to compete with OpenAI O1 on MT. The evaluation is based on two literary books: "The Essential O. Henry Collection" and "Orbital". The content of "The Essential O. Henry Collection" has been released in data/ohenry.txt, while "Orbital" cannot be released due to copyright protection. If it is needed for research purposes only, you may send a request email to jawang.nlp[at]gmail.com. The project uses the cc-by-nc-sa-4.0 license. Regarding fees, pricing, and other information, no details are currently provided in the materials.
