Per-category training lifts software-engineering agent 5.4 points over base model
single source· 1 articles · confidence: medium · first seen 2026-09-19 20:00 UTC
What this means for you
Nothing to act on yet. This is a preprint — the paper names no released model and no evaluation date, so the 5.39-point gain cannot be checked against a harness you already run. If you train coding agents on mixed tasks, the transferable idea is the category split, not the number.
A preprint splits software-engineering agent training into separate experts per task category, then distils them into one deployable model by training it on the others' outputs. Pooled training on mixed tasks lifts some categories while hurting others, so the authors train category-specific policies, alternating reinforcement learning with a refresh-repair-expand cycle that reuses each agent's own verified successful runs. The combined policy reports 58.04% mean resolution (issues fixed) on Pro-618 and 59.00% on SWE-bench Multilingual, 5.39 and 2.78 percentage points above its base model. The paper gives no evaluation date, and the scores are from a preprint rather than a third-party run.
Key facts
- ·The final distilled policy reports 58.04% mean resolution on Pro-618 and 59.00% on SWE-bench Multilingual. source
- ·Those figures are 5.39 and 2.78 percentage points above the base model, according to the paper. source
- ·The paper gives no evaluation date or harness configuration for either score. source
- ·Category-specific reinforcement learning alternates with a Refresh-Repair-Expand cycle that reselects tasks and reuses the agent's own verified successful trajectories. source
- ·The experts are consolidated into one model by label-routed multi-teacher on-policy distillation, which requires no external model to supply solution trajectories or action targets. source
- ·The work is posted as arXiv 2609.23377, dated 19 September 2026. source
What the sources say
- Hugging Face Daily Papers (research) — Hugging Face listing for an arXiv preprint on per-category expert training, distillation and reported SWE-bench scores.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersOne to More, More to One: Category-Aware Iterative Expert Training for Software Engineering Agents2026-09-19