When Requirements Interact: Scaling Long-Horizon Coding via Interdependent Specifications
Coding tasks rarely test constraints that depend on each other. CodeSpecWeave synthesizes interdependent specs and audits the tests: 4,670 tasks across 2,865 repositories.
The claim
Coding tasks need constraints that interact, not one primary goal.
The problem
Real development has interface, error, and dependency constraints at once.
Synthesized tasks often ignore those interacting constraints.
Weak tests either miss hidden flaws or reject valid solutions.
How it works
A synthesis agent starts from real developer seeds.
It writes coupled task requirements together with executable tests.
A review agent audits solutions from several models.
That audit keeps the tests reliable.
Reinforcement learning on the tasks trains stronger coding agents.
Figures
What changes
The pipeline builds 4,670 training tasks across 2,865 repositories.
Training beats the base model on DeepSWE-v1.1.
CodeSpecWeave-Bench holds 100 tasks with 1,044 specifications.
Limits
The benchmark covers 100 tasks, a small sample.
Test quality still depends on the review agent.
Training gains are reported against one base model setting.
Source · arXiv