What's this about?
Prior work on making models resolve real software issues has focused on building large agent-trajectory datasets and doing supervised fine-tuning on the successful ones. SWE-Prime argues for the opposite: fewer, better-selected trajectories produce better results.
Why would fewer examples be better?
Because "successful" is a weak filter. A trajectory that eventually fixed the bug may have wandered through six dead ends, and training on it teaches the wandering along with the fix. Curating for trajectories that were *efficiently* correct gives the model a cleaner signal about what good problem-solving looks like — the same reason a small set of well-worked examples teaches better than a large pile of correct-but-messy ones.
Why does this matter to you?
- It is a direct challenge to the scale-the-dataset default in agent training.
- Same theme as the SFT-versus-RL paper above: check whether the expensive fashionable thing beats a careful cheap thing.
- Practically relevant if you are curating any kind of example set for a model.