SWE-Prime: Train software agents on fewer, higher-quality trajectory segments¶

Core mechanism plus oracle-free L2.1 evaluation: trajectory screening, segment screening and loss mask.

Paper information¶

Field	Details
Paper	arXiv 2608.27449
Institution	Sun Yat-sen University (first-listed affiliation of first author), with Huawei Cloud and Chongqing University.
First public release	2026-08-27 (arXiv v1)
Author code	No public author code found as of 2026-08-29.
Local adapter or method key	`swe-prime`
Local reproduction code	`src/auto_research/agent_research/latest_20260829.py`

Paper method and evidence¶

Even successful trajectories can be redundant, unsafe or unlearnable. Select traces for process, outcome and representativeness, then select semantic segments for contribution, learnability and risk. Keep all context but compute SFT loss only on retained segments.

flowchart LR
  T[Successful traces] --> R[Screen process, outcome, and coverage]
  R --> S[Screen semantic segments]
  S --> M[Retain context and mask SFT loss]

Original paper figure¶

The original figure shows the trajectory and segment selection stages. From the original paper; copyright remains with the authors.

Core equation¶

\[ \mathcal L=-\sum_t m_t\log p_\theta(y_t\mid y_{<t},x),\quad m_t\in\{0,1\}. \]

Reported results¶

With only 10% of the trajectories, the paper reports up to 12.2% / 24.2% improvements over all-successful-trajectory training on SWE-Bench Pro / Verified.

Local reproduction¶

Three-seed ToolRoute-L2.1 joint success 0.7708, plan F1 0.8004, mean cost 4.5497. Metrics and batch results. No code LLM training or official SWE-Bench; this compares filters under the same observation/tool interface.