Search papers, labs, and topics across Lattice.
The paper introduces a two-stage Supervised Fine-Tuning (SFT) recipe, SWE-ZERO to SWE-HERO, for training software engineering agents. The first stage, SWE-ZERO, uses large-scale, execution-free data to learn code semantics, while the second stage, SWE-HERO, refines these intuitions with execution-backed data. SWE-HERO-32B achieves a 62.2% resolution rate on SWE-bench Verified, setting a new state-of-the-art for open-source models of comparable size and demonstrating strong zero-shot transferability to other languages.
Open-source models can now solve 62% of SWE-bench problems, rivaling proprietary models, thanks to a novel two-stage training approach that first learns code semantics without execution, then refines with execution feedback.
We introduce SWE-ZERO to SWE-HERO, a two-stage SFT recipe that achieves state-of-the-art results on SWE-bench by distilling open-weight frontier LLMs. Our pipeline replaces resource-heavy dependencies with an evolutionary refinement strategy: (1) SWE-ZERO utilizes large-scale, execution-free trajectories to master code semantics and repository-level reasoning, and (2) SWE-HERO applies targeted, execution-backed refinement to transition these semantic intuitions into rigorous engineering workflows. Our empirical results set a new benchmark for open-source models of comparable size. We release a dataset of 300k SWE-ZERO and 13k SWE-HERO trajectories distilled from Qwen3-Coder-480B, alongside a suite of agents based on the Qwen2.5-Coder series. Notably, SWE-HERO-32B achieves a 62.2% resolution rate on SWE-bench Verified. Furthermore, despite being trained exclusively on Python, our agents demonstrate robust zero-shot transferability on SWE-bench Multilingual, reaching 44.1% and confirming the paradigm's generalizability across diverse languages.
