Fine-tuning on 10% of successful agent trajectories beats fine-tuning on all of them, by up to 24.2%

arXiv 2608.27449·medium signalXBlueskyLinkedInCopy link

SWE-Prime's premise is that a successful trajectory still contains ineffective, redundant and risky steps, so SFT on all resolved runs teaches bad habits. It filters twice: trajectory-level on process quality, result quality and representativeness, then segment-level by grouping consecutive steps and scoring each on contribution, learnability and risk. All segments stay in the sequence for context but only selected ones contribute to the loss, and the 10% subset beats the full resolved dataset by up to 12.2% on SWE-Bench Pro and 24.2% on Verified.