Shanghai Jiao Tong University researchers introduced SWE-at-Scale, a framework that creates realistic software engineering tasks by reversing historical bug fixes. The paper, titled “SWE-at-Scale: Scalable Issue Synthesis from Historical Bug Fixes,” was published on October 15 and submitted to Hugging Face Papers on October 16. It aims to supply broader, more varied tasks for evaluating coding agents.

The researchers say real-world software issues are limited by how often developers report them, while manually curating tasks is costly. Existing synthetic methods mainly use large models’ pretrained knowledge and heuristic prompts, which the authors say often miss the complex mechanics of actual software failures. SWE-at-Scale instead treats a real fix as evidence of how a bug formed: reversing the fix turns repair examples into examples for inducing bugs grounded in development scenarios.

The framework transfers learned bug patterns to unfamiliar code contexts and samples different bug types to create executable tasks. In experiments, it generated tasks the authors describe as more diverse and realistic than current synthesis baselines, while also making them harder for state-of-the-art coding agents. A single code context could yield distinct bug types, and the authors report that bug diversity scaled robustly through 200 generations per file. They present the approach as a way to reuse a finite set of real-world fixes for scalable task generation.

Shanghai Jiao Tong University researchers introduced SWE-at-Scale, a framework that creates realistic software engineering tasks by reversing historical bug fixes. The paper, titled “SWE-at-Scale: Scalable Issue Synthesis from Historical Bug Fixes,” was published on October 15 and submitted to Hugging Face Papers on October 16. It aims to supply broader, more varied tasks for evaluating coding agents.

The researchers say real-world software issues are limited by how often developers report them, while manually curating tasks is costly. Existing synthetic methods mainly use large models’ pretrained knowledge and heuristic prompts, which the authors say often miss the complex mechanics of actual software failures. SWE-at-Scale instead treats a real fix as evidence of how a bug formed: reversing the fix turns repair examples into examples for inducing bugs grounded in development scenarios.

The framework transfers learned bug patterns to unfamiliar code contexts and samples different bug types to create executable tasks. In experiments, it generated tasks the authors describe as more diverse and realistic than current synthesis baselines, while also making them harder for state-of-the-art coding agents. A single code context could yield distinct bug types, and the authors report that bug diversity scaled robustly through 200 generations per file. They present the approach as a way to reuse a finite set of real-world fixes for scalable task generation.