Tsinghua Team Open-Sources VeriLoop Coder-E1: Evidence-Governed Spiral for Verifiable Recursive Self-Improvement

On July 29, 2026, the team from Tsinghua University's Shenzhen International Graduate School released VeriLoop Coder-E1, built on Qwen3.6-27B, targeting repository-level code repair and agentic software engineering tasks. According to Machine Intelligence, the model scored 85.20 on SWE-bench Verified, 62.38 on SWE-bench Pro, 76.40 on Terminal-Bench 2.0, and 33.63 on DeepSWE.

As of July 27, 2026, among open-source models with 32B parameters or fewer, VeriLoop Coder-E1 ranks first on SWE-bench Verified, SWE-bench Pro, and Terminal-Bench 2.0, and second on DeepSWE; among all open-source models, it ranks second, first, first, and fifth respectively. Data is sourced from the Hugging Face leaderboard, but the article emphasizes that benchmarks only reflect relative performance on specific task distributions and cannot infer overall capability in real-world environments.

The core of VeriLoop Coder-E1 is the synergy between narrow-domain PEFT fine-tuning and Self-Harness. The base model is frozen, and only a small number of parameters are trained to strengthen capabilities such as tool contract adherence and evidence-conclusion binding, with adapters that can be loaded and unloaded. Self-Harness organizes tasks into multi-round execution chains, where each round generates candidates and then goes through rebuttal, targeted exploration, revision, and verification; only corrections that pass verification proceed to the next round.

The team defines this mechanism as the evidence spiral, following the logic of 'evidence, falsification, exploration, repair, verification, consolidation': collect evidence, test hypotheses, explore gaps, repair according to mismatch levels, verify effectiveness, and compress experience into future rules. This differs from simply increasing reasoning rounds; instead, each round's invocation is changed by new evidence in how judgments are made.

The team argues that merely modifying one's own components cannot prove true improvement; if goals and evaluation assumptions are not tested, deeper modifications are more likely to solidify biases. They redefine recursive self-improvement: methods corrected by evidence change how the system discovers and corrects errors in the future, and such methods remain falsifiable by new evidence.

During operation, verified revisions do not directly become permanent rules; they are registered as method candidates. Only when they are actually invoked in subsequent independent tasks, changing evidence obligations, tool selection, etc., and again receive evidence support, are they written into new error-correction methods. If later evidence conflicts, they can be narrowed, downgraded, or rolled back.

The team open-sources model weights, tokenizer, configuration, partial PEFT adapters, and evaluation materials under Apache 2.0, but the full implementation of Self-Harness is not yet open because it is a complex runtime control plane that determines whether the system can become a general Code Agent. The team's next phase shifts commercial focus to a general Code Agent.

Within 48 hours of release, third-party developers, including from Brazil, created and released GGUF quantized versions supporting local inference via llama.cpp, Ollama, and others. As of July 30, 2026, the original repository had 413 downloads in a single day, and the GGUF repository had 955. Third parties also report strong anti-abliteration capabilities, with refusal rates dropping from about 95% to about 82% after 200 attempts, but this data is not officially confirmed.

The team's vision differs from the monolithic model-centric AGI path, emphasizing that systems must form corrigible state representations, predict action consequences, and change future questioning and action when reality overturns predictions. Software engineering is a rigorous starting point for testing this view because code repositories provide environmental states, and tests and execution traces provide external evidence.

A general Code Agent will predict code, dependencies, and interface changes, and after execution, compare predictions against Git diffs and test logs; if mismatches occur, it will re-explore or roll back. The team states they will continue to compete on public benchmarks and real tasks, but will not just keep digging in holes others have dug.