ReflectMT: An Efficient Machine Translation Method with Internalized Reflection Capability

ReflectMT internalizes the 'translate-reflect-optimize' capability into the model via two-stage reinforcement learning, enabling it to generate high-quality translations directly during inference without explicit reasoning. On the WMT24 benchmark, its translation quality surpasses DeepSeek-R1 (COMET score 88.7 vs. 86.5), while reducing token consumption by 94.33%, solving the dilemma of balancing quality and efficiency in existing Large Reasoning Model (LRM) translation methods.

Large Reasoning Models (LRMs) such as DeepSeek-R1 adopt the 'think-first-then-translate' paradigm: first generate a reasoning process (analyzing semantics, cultural differences, etc.), then generate the translation. Although this improves quality, it has three major problems:

1. Token explosion: Reasoning consumes several times more tokens than the translation
2. Latency surge: Additional reasoning steps increase end-to-end latency
3. Cost spike: API fees are proportional to the number of tokens These overheads are unacceptable in production environments.

Core insight of ReflectMT: Learn to think during training, translate directly during inference. It uses two-stage training:

First Stage: Cultivate Reflection and Optimization Capability

The model learns the 'translate → reflect (identify semantic deviations, style inappropriateness, etc.) → optimize' process, with reinforcement learning rewarding translation quality, reflection accuracy, and optimization effectiveness.

Second Stage: Internalize Reflection Knowledge

High-value reflection knowledge from the first stage is extracted via knowledge distillation, training the model to generate high-quality translations directly without explicit reflection steps.

Quality Comparison

• WMT24 en-de: ReflectMT COMET 88.7 vs. DeepSeek-R1 86.5 (+2.2)
• GPT-4 evaluation: ReflectMT average 9.96/10 vs. DeepSeek-R1 7.8/10 (+2.16)

Efficiency Improvement

• Token consumption: ReflectMT ~850 tokens vs. DeepSeek-R1 ~15000 tokens (reduced by 94.33%)
• Effects: Latency reduced to hundreds of milliseconds, cost cut by over 90%, throughput increased by more than 10x

Multilingual Validation

Effective across language pairs like English-German, English-French, English-Chinese, English-Japanese, demonstrating universality.

Quantification of Reflection Quality

80% of reflections (35% high-value +45% medium-value) directly or indirectly improve translation quality, and these knowledge points are extracted and internalized in the second stage.

Changes in Attention Patterns

ReflectMT's attention is more focused, enabling it to identify key semantic clues and reduce omissions.

Reduction in Error Types

• Semantic errors: -42%
• Style inconsistency: -38%
• Cultural misinterpretation: -51% These are all key issues focused on during the reflection stage.

1. Rethink LRM applications: The explicit reasoning capability of LRMs can be compiled into the model through training, balancing quality and efficiency, providing new ideas for other NLP tasks.
2. New paradigm of training-inference decoupling: Invest more computation during training to save computation during inference, optimizing the training-inference trade-off.
3. Simulate human learning: From explicit analysis (beginners) to intuitive judgment (skilled practitioners), metacognitive ability is key to efficient AI.

Limitations

1. Two-stage training requires large amounts of computing resources
2. Domain-specific reflection training is needed for specific fields (law, medicine)
3. No explicit reflection during inference reduces interpretability

Future Directions

• Incremental learning: Support online learning of new language pairs
• Hybrid mode: Explicit reflection for difficult sentences, direct translation for simple ones
• Multimodal extension: Scenarios like image description, speech translation

ReflectMT proves the effectiveness of the 'think during training, intuit during inference' paradigm, providing a general strategy for improving the practicality of AI systems.