A Tsinghua SIGS research team open sourced VeriLoop Coder-E1, an evidence-governed 27B model achieving state-of-the-art results across major repository-level code repair benchmarks.
A research team led by Professor Liu Houde at the Intelligent Robotics Laboratory of Tsinghua Shenzhen International Graduate School (Tsinghua SIGS) has open-sourced VeriLoop Coder-E1, a specialised vertical coding model designed for repository-level code repair and agentic software engineering.
The model is built on Qwen3.6-27B with the base parameters remaining frozen. VeriLoop Coder-E1 employs narrow-domain Parameter-Efficient Fine-Tuning (PEFT) delivered via a detachable Surface Host Adapter to reinforce tool-contract following, evidence-conclusion binding, uncertainty identification, failure explanation, local repair, and rollback control.
The system utilises a proprietary Self-Harness mechanism that structures code generation as a closed, evidence-governed multi-round loop, recompiling generated code, test receipts, tool feedback, and counter-evidence into structured Markdown. Inspired by Karl Popper’s principles of falsifiability, the model implements an “Evidence-Governed Spiral” where iterative corrections are validated by new evidence to dynamically adapt how the system detects, judges, and fixes software defects.
In benchmark evaluations, VeriLoop Coder-E1 demonstrated state-of-the-art capabilities for open-source vertical coding models, achieving impressive scores of 85.20 on SWE-bench Verified, 62.38 on SWE-bench Pro, 76.40 on Terminal-Bench 2.0, and 33.63 on DeepSWE.
Among open-source models at or under 32B parameters, VeriLoop Coder-E1 ranks first on SWE-bench Verified, SWE-bench Pro, and Terminal-Bench 2.0, and second on DeepSWE. Across all open-source models regardless of size, it ranks second, first, first, and fifth respectively on those benchmarks. The core model weights and adapter assets are fully open-sourced, whilst the upper-level Self-Harness control stack remains retained by the team.
