SWE-Prime: Fewer Trajectories, Better Performance
SWE-Prime trains better by filtering supervision quality, not by adding more solved traces.
The paper argues that successful agent trajectories can still teach bad habits if they include redundant, ineffective, or risky steps. Its method first selects a smaller set of high-quality successful trajectories, then marks which semantic segments should actually contribute to fine-tuning loss while keeping full context in sequence. In experiments on SWE-Bench Pro and SWE-Bench Verified, using the selected 10% subset beat training on the full resolved dataset, with reported relative gains up to 12.2% and 24.2%. ArXiv · AI/CL/LG's note
The paper argues that successful agent trajectories can still teach bad habits if they include redundant, ineffective, or risky steps. Its method first selects a smaller set of high-quality successful trajectories, then marks which semantic segments should actually contribute to fine-tuning loss while keeping full context in sequence. In experiments on SWE-Bench Pro and SWE-Bench Verified, using the selected 10% subset beat training on the full resolved dataset, with reported relative gains up to 12.2% and 24.2%. ArXiv · AI/CL/LG's note
score 5