Paper 2601.18418
daVinci-Dev: Agent-native Mid-training for Software Engineering
- Published
- Jan 2026
- Research lab
- Independent
- Citations
- 10
- GitHub
- 73 stars
01 In brief
Summary
The paper introduces daVinci-Dev, a training recipe for agentic software engineering that uses agent-native mid-training data to bridge the gap between static training corpora and dynamic, feedback-rich development environments.
The authors construct two complementary trajectory types: contextually-native trajectories (68.6B tokens) derived from GitHub pull requests, preserving the full localization-to-edit workflow, and environmentally-native trajectories (3.1B tokens) collected from agent rollouts in executable Docker environments with real test feedback.
Starting from Qwen2.5-Base models, they achieve state-of-the-art SWE-Bench Verified resolution rates of 56.1% (32B) and 58.5% (72B) using the SWE-AGENT scaffold, surpassing the prior open recipe KIMI-DEV while using less than half the mid-training tokens (73.1B vs ~150B).
The approach also improves performance on general code generation and scientific benchmarks.
Key findings include the synergy between the two data types, the importance of bundling context with edits, and predictable scaling with training steps.
The authors plan to open-source datasets, recipes, and checkpoints to facilitate further research in this underexplored paradigm.
Limitations include potential privacy issues from developer identifiers, evaluation sensitivity, and focus on a single model family and benchmark.
The work demonstrates that agentic mid-training is a scalable and token-efficient path to instilling foundational agentic behaviors in LLMs, complementing post-training methods like SFT and RL.
02 From the paper
Abstract
Recently, the frontier of Large Language Model (LLM) capabilities has shifted from single-turn code generation to agentic software engineering-a paradigm where models autonomously navigate, edit, and test complex repositories. While post-training methods have become the de facto approach for code agents, **agentic mid-training**-mid-training (MT) on large-scale data that mirrors authentic agentic workflows-remains critically underexplored due to substantial resource requirements, despite offering a more scalable path to instilling foundational agentic behaviors than relying solely on expensive reinforcement learning. A central challenge in realizing effective agentic mid-training is the distribution mismatch between static training data and the dynamic, feedback-rich environment of real development. To address this, we present a systematic study of agentic mid-training, establishing both the data synthesis principles and training methodology for effective agent development at scale. Central to our approach is **agent-native data**-supervision comprising two complementary types of trajectories: **contextually-native trajectories** that preserve the complete information flow an agent experiences, offering broad coverage and diversity; and **environmentally-native trajectories** collected from executable repositories where observations stem from actual tool invocations and test executions, providing depth and interaction authenticity. We verify the model's agentic capabilities on `SWE-Bench Verified`. We demonstrate our superiority over the previous open software engineering mid-training recipe `Kimi-Dev` under two post-training settings with an aligned base model and agentic scaffold, while using less than half mid-training tokens (73.1B). Besides relative advantage, our best performing 32B and 72B models achieve **56.1%** and **58.5%** resolution rates, respectively, which are ...