The SWE-Nexus paper, posted on arXiv, argues that historical code changes are a convenient way to define software engineering tasks, but they have two key shortcomings: they do not identify which production functions are exercised by selected tests, and they do not support composing validated behaviors that span more than one change.
To address this, the paper introduces SWE-Nexus, a system for automated composition of software engineering tasks from test execution. The abstract does not provide details on the method or evaluation, but it positions the system as a way to move beyond single-change boundaries.
The significance is that task composition based on test execution could better align with actual behavior validation, rather than relying on the arbitrary boundaries of historical commits. However, without further details, the claims are limited to the problem statement and the system's intended purpose.