Researchers from Xiaomi MiMo present CodeSpecWeave, a framework that generates coding tasks with interdependent specifications and executable tests to train coding agents.
Pipeline: A synthesis agent builds coupled task requirements and tests from real developer seeds, and a review agent audits solution trajectories from multiple models to check test reliability.
Training data: The framework produces 4,670 tasks across 2,865 repositories, used for reinforcement learning. The abstract says this significantly improves over the base model on DeepSWE-v1.1.
Benchmark: CodeSpecWeave-Bench has 100 tasks and 1,044 specifications. The authors say it clearly separates frontier models and leaves significant headroom on interdependent engineering constraints.
Sign in to comment.
TL;DR of the paper, generated automatically.
Researchers from Xiaomi MiMo present CodeSpecWeave, a framework that generates coding tasks with interdependent specifications and executable tests to train coding agents.