Developing AI that can actually perform software engineering—navigating files, running tests, and fixing bugs—requires more than just reading code. SII-GAIR has introduced daVinci-Dev, a project focused on "agentic mid-training." This technique involves training models on large-scale data that mimics real-world developer workflows, moving beyond simple code generation to autonomous problem-solving. The team addressed the distribution mismatch between static text and interactive coding environments by using agent-native data. This includes contextually-native trajectories, which capture the full information flow a developer sees, and environmentally-native trajectories, which record actual tool uses and test results. By training on these dynamic feedback loops, the model learns the foundational behaviors needed to act as an independent Coding Agent. The results are impressive: their 72B model reached a 58.5% resolution rate on the SWE-Bench Verified benchmark. Notably, it outperformed previous methods like Kimi-Dev while requiring less than half the training tokens. This efficiency suggests that the quality of interactive data is far more important than sheer volume when teaching models to navigate complex software repositories. By starting with a general-purpose base model rather than a specialized coder, daVinci-Dev proves that mid-training is a powerful, scalable alternative to expensive reinforcement learning. The researchers plan to release their datasets and model checkpoints, offering the community a blueprint for building high-performance software agents without massive computational overhead.

Developing AI that can actually perform software engineering—navigating files, running tests, and fixing bugs—requires more than just reading code. SII-GAIR has introduced daVinci-Dev, a project focused on "agentic mid-training." This technique involves training models on large-scale data that mimics real-world developer workflows, moving beyond simple code generation to autonomous problem-solving. The team addressed the distribution mismatch between static text and interactive coding environments by using agent-native data. This includes contextually-native trajectories, which capture the full information flow a developer sees, and environmentally-native trajectories, which record actual tool uses and test results. By training on these dynamic feedback loops, the model learns the foundational behaviors needed to act as an independent Coding Agent. The results are impressive: their 72B model reached a 58.5% resolution rate on the SWE-Bench Verified benchmark. Notably, it outperformed previous methods like Kimi-Dev while requiring less than half the training tokens. This efficiency suggests that the quality of interactive data is far more important than sheer volume when teaching models to navigate complex software repositories. By starting with a general-purpose base model rather than a specialized coder, daVinci-Dev proves that mid-training is a powerful, scalable alternative to expensive reinforcement learning. The researchers plan to release their datasets and model checkpoints, offering the community a blueprint for building high-performance software agents without massive computational overhead.