Training code generation models to debug their own outputs
39%. That's the success rate improvement when code generation models learn to debug themselves.

Why it matters
Amazon's research demonstrates a practical method to enhance AI model reliability through self-correction loops—a capability that could reshape how enterprises deploy code generation tools in production.
The key facts
5 to know39% improvement in code generation success rate
Method uses LLMs to generate training data
Combines fine-tuning and reinforcement learning approaches
Published by Amazon Science (Feb 20, 2025)
Addresses model self-correction and debugging capabilities
Go to the source
Amazon Scienceamazon.science
Publisher excerpt: Using large language models to generate training data and updating models through both fine tuning and reinforcement learning improves the success rate of code generation by 39%.