Researchers have introduced SWE-at-Scale, a framework that generates realistic software engineering training tasks for coding agents by reverse-engineering historical bug fixes rather than relying on a language model’s own prompted guesses at plausible bugs. By inverting a known fix, the method turns examples of bug repair into examples of how bugs are actually introduced, then applies the learned bug patterns to new codebases to build executable, verifiable tasks. The authors report that the resulting tasks are more diverse and realistic than existing synthetic benchmarks and pose a greater challenge to state-of-the-art coding agents, with the code and models publicly released.
