Self-play curriculum learning improves language model code repair

The genuinely novel structural choice here is that a single model plays both roles simultaneously: it generates bugs calibrated to its own current repair ability, which means the difficulty ceiling rises automatically as the model improves, without any external curriculum designer. Most prior self-play work in code separates the generator and fixer into distinct models, so the anchoring to one shared set of weights is the architectural bet worth scrutinizing.

BugSourceBench's three-category taxonomy, covering human-written code, LM-generated code, and human-edited LM output, directly addresses the evaluation gap that GameEngineBench (covered July 3rd) exposed from a different angle: real-world code tasks resist synthetic benchmarks because the bug distributions differ sharply across authorship contexts. The self-play training loop also rhymes with the SEA architecture covered July 1st, where self-modification is gated by formal certificates to prevent drift. This paper's acknowledged distribution drift problem during training is precisely the failure mode SEA was designed to contain, though the two papers do not appear to be in direct dialogue.

Watch whether BugSourceBench gets adopted as an evaluation layer in coding agent papers over the next two conference cycles. If it does, the three-category authorship split will pressure other repair benchmarks to disaggregate their numbers the same way, which would expose how much existing repair performance is concentrated on the easiest human-written category.