This research demonstrates how to adapt decoder-only LLMs to encoder-decoder architecture for improved inference efficiency while maintaining quality.

• Achieves 32x faster inference with comparable performance metrics
• Introduces techniques for effective parameter initialization and optimization
• Develops novel pretraining objectives specifically for adaptation
• Creates a path to leverage strengths from both model architectures

For engineering teams, this approach offers a practical solution to deploy high-quality LLMs in resource-constrained environments where inference speed matters, without sacrificing core capabilities.

Original Paper: Encoder-Decoder Gemma: Improving the Quality-Efficiency Trade-Off via Adaptation