← 研踪 PaperTrailarXiv:2611.23876 · v1
SWE-growth: Teacher-Guided Adaptive Self-Play for Software Engineering Agents
首次提交 2026-11-18 13:29
版本更新 2026-11-18 13:29
雷达观测 2026-11-19 19:08
摘要
Self-play offers a scalable way to train software engineering agents by generating repository-level repair tasks. However, with binary repair rewards and finite rollout budgets, all-fail groups provide no reward-based group-relative advantage. Such tasks may still offer learning opportunities when a Teacher can provide verified repairs absent from the Student's own attempts. We introduce SWE-Growth, a teacher-guided adaptive self-play framework that uses verified Teacher outcomes to identify learning opportunities for the current Student. Tasks with all students failure but reliable Teacher success provide demonstrations for supervised fine-tuning (SFT), and post-SFT Student rollouts identify mixed-outcome tasks with a positive Teacher--Student gap for reinforcement learning. Task eligibility also guides Injector updates, connecting Student training with subsequent task generation. Starting from 101 seed issues, \textsc{SWE-Growth} improves Qwen3-Coder-30B-A3B-Instruct from 27.7% to 31.7% on SWE-bench Pro and from 11.0% to 17.2% on held-out TerminalBench 2.1 by Round~2. It exceeds Student-only self-play by 2.6 and 4.1 percentage points, respectively, and Teacher-augmented self-play by 1.8 and 2.5 points under approximately matched budgets. These results support adapting both supervision and task selection to the Student's evolving repair capability.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-18 13:29 · 最新观测版本
SWE-growth: Teacher-Guided Adaptive Self-Play for Software Engineering Agents
Self-play offers a scalable way to train software engineering agents by generating repository-level repair tasks. However, with binary repair rewards and finite rollout budgets, all-fail groups provide no reward-based group-relative advantage. Such tasks may still offer learning opportunities when a Teacher can provide verified repairs absent from the Student's own attempts. We introduce SWE-Growth, a teacher-guided adaptive self-play framework that uses verified Teacher outcomes to identify learning opportunities for the current Student. Tasks with all students failure but reliable Teacher success provide demonstrations for supervised fine-tuning (SFT), and post-SFT Student rollouts identify mixed-outcome tasks with a positive Teacher--Student gap for reinforcement learning. Task eligibility also guides Injector updates, connecting Student training with subsequent task generation. Starting from 101 seed issues, \textsc{SWE-Growth} improves Qwen3-Coder-30B-A3B-Instruct from 27.7% to 31.7% on SWE-bench Pro and from 11.0% to 17.2% on held-out TerminalBench 2.1 by Round~2. It exceeds Student-only self-play by 2.6 and 4.1 percentage points, respectively, and Teacher-augmented self-play by 1.8 and 2.5 points under approximately matched budgets. These results support adapting both supervision and task selection to the Student's evolving repair capability.
查看此版本解读 →查看此版本原文 ↗