When Requirements Interact: Scaling Long-Horizon Coding via Interdependent Specifications

When Requirements Interact. Scaling Long-Horizon Coding via Interdependent Specifications. By Yishuo Cai and seven others, from Xiaomi's LLM Core team, Peking University, and one other institution. Published on the twenty-ninth of October.

Real-world long-horizon software development asks a coding agent to satisfy many specifications at once, interface compatibility and error handling among them, and to do it without breaking dependencies that already exist. Current work falls short on specifications in both quantity and diversity. It tends to emphasize the primary goal of a task while ignoring the complex constraints that surround it.

Synthesizing tasks could close that gap, but two challenges stand in the way. Adding specifications naively risks reducing a task to a trivial checklist, where each item can be ticked off alone instead of compelling joint reasoning. And verifying how the specifications interact risks either overlooking hidden flaws or rejecting valid alternative solutions. CodeSpecWeave is a framework for synthesizing tasks with rich, interdependent specifications, designed to meet both challenges.

Tasks are grounded in real developer seeds. From each seed, a synthesis agent iteratively develops coupled task requirements together with executable tests. A review agent then audits solution trajectories from multiple models, to guarantee that the tests are reliable. The two agents map onto the two challenges. The synthesis agent couples the requirements so the task cannot be done as a checklist, and the review agent checks the tests against several models' attempts so that they neither miss flaws nor reject sound solutions. This pipeline yields four thousand six hundred and seventy training tasks, drawn from two thousand eight hundred and sixty-five repositories.

Reinforcement learning on these tasks significantly improves performance over the base model on DeepSWE-v1.1, largely outperforming existing baselines. The authors also curate CodeSpecWeave-Bench, a rigorous benchmark of one hundred tasks carrying one thousand and forty-four specifications. It clearly separates frontier models, and it reveals significant headroom in how they handle interdependent engineering constraints.

The weave in the name is the contribution. A list of requirements that can each be met in isolation tests very little, so these tasks are built with requirements that are coupled, and with tests that a second agent has checked against several models' attempts. Even frontier models, by the benchmark's account, still have a long way to go on that kind of work.