From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents

Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• The SWE-ZERO to SWE-HERO two-stage SFT recipe achieves state-of-the-art results on SWE-bench by distilling open-weight frontier LLMs.result 90%
• SWE-HERO-32B achieves a 62.2% resolution rate on SWE-bench Verified.result 95%
• The pipeline replaces resource-heavy dependencies with an evolutionary refinement strategy.method 85%
• SWE-ZERO utilizes large-scale, execution-free trajectories to master code semantics and repository-level reasoning.method 90%
• SWE-HERO applies targeted, execution-backed refinement to transition semantic intuitions into rigorous engineering workflows.method 90%
• The agents demonstrate robust zero-shot transferability on SWE-bench Multilingual, reaching 44.1% despite being trained exclusively on Python.result 90%
• The paper releases a dataset of 300k SWE-ZERO and 13k SWE-HERO trajectories distilled from Qwen3-Coder-480B.technical 95%
• The empirical results set a new benchmark for open-source models of comparable size.result 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card